Thesis Fulltext

Digital Image Elasto-Tomography:
Mechanical Property
Reconstruction from Surface
Measured Displacement Data
Ashton Peters
A thesis presented for the degree of

Doctor of Philosophy
in
Mechanical Engineering
at the
University of Canterbury,
Christchurch, New Zealand.
July 2007
This thesis is dedicated to the memory of Madge Peters (1918–2003), and all other New
Zealanders whose lives have been affected by breast cancer. May the perseverance of
researchers throughout the world yield a cure that will render our work unnecessary.
Acknowledgments
The research presented in this thesis would not have been possible, or as enjoyable as
I have found it, without the assistance of a number of individuals, who I will attempt
to acknowledge in this short note of appreciation.
It was my supervisor, Dr. Eli Van Houten, who first proposed that I extend my
university studies to work on the DIET project. Eli, I thoroughly appreciate the
many hours you have spent with me brainstorming ideas, debugging computer code,
or discussing strategies when I was stuck at one of the countless obstacles encountered
during the past three years. I have enjoyed your friendly, easy-going, and patient
nature, and I consider myself very fortunate to have a supervisor who I also consider
to be a friend. I also owe a significant thanks to my associate supervisor, Prof. Geoff
Chase, whose experience and helpful assistance when writing and reviewing material
for publication, and thoughts on my research from an independent perspective, have
been most appreciated. I look forward to working with both of you in the future.
The DIET project has been, and continues to be, a team effort that involves a
number of students and staff. I would like to acknowledge the assistance of all those
who have contributed over the past few years, including (but not limited to): Jérôme
Rouzé, Arnaud Milsant, Stefan Wortmann, Fabrice Jandet, Edouard Ravni, Crispin
Berg, Anthony Hii, Shig Kinoshita, Richard Brown, Hans Uwe-Berger, Rodney Elliott,
Mark Staiger, Kevin Stobbs, and Chris Hann. Thanks also to my fellow research
students in the Center for Bioengineering for contributing to the sense of community
we share, and to Prof. Tim David for creating a work environment where the overall
‘feel’ is determined by research students, and not by administrators.
To my family, including those in New Plymouth, Hamilton, Christchurch and
Timaru, thank you for you words of encouragement, and for the various forms of
support you have given me, particularly during the writing of this thesis, which at
one point seemed an insurmountable task. Finally, to Alison, you are both family and
my best friend, and I thank you for all the love and encouragement you have given
me. After more than three years of additional study, I am finally a salaried employee
(hooray!), and I look forward to spending the rest of my life with you.
Contents
Abstract xxiii
1 Introduction and Background 1

1.1 Basic Cancer Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Breast Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Breast Cancer Screening . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Mammography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Alternative Breast Imaging Modalities . . . . . . . . . . . . . . . 11
1.4 Elastography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4.1 Mechanical Properties of Soft Tissue . . . . . . . . . . . . . . . . 15
1.4.2 Ultrasound Elastography . . . . . . . . . . . . . . . . . . . . . . 17
1.4.3 Magnetic Resonance Elastography . . . . . . . . . . . . . . . . . 18
1.4.4 Novel Elastographic Techniques . . . . . . . . . . . . . . . . . . . 19
1.5 Stochastic Optimization in Elastography . . . . . . . . . . . . . . . . . . 20
1.6 Digital Image Elasto-Tomography . . . . . . . . . . . . . . . . . . . . . . 22
1.7 From Proof of Concept to Phantom Elastography . . . . . . . . . . . . . 23
2 Proof of Concept Studies 25

2.1 Data Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Three Region Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Dual-Mesh Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Experimental Methods 37
3.1 Phantom Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.1 Gelatine Phantoms . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.2 Silicone Phantoms . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Material Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Actuation and Motion Capture . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.1 Motion Capture from Gelatine Phantoms . . . . . . . . . . . . . 52
3.3.2 Motion Capture from Silicone Phantoms . . . . . . . . . . . . . . 54
3.4 Photogrammetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.5 Displacement Data Correlation . . . . . . . . . . . . . . . . . . . . . . . 58
viii CONTENTS
3.5.1 Mesh Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.5.2 Surface Projection . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.2.1 Reference Point Centroid . . . . . . . . . . . . . . . . . 59
3.5.2.2 Surface Element Normal Vectors . . . . . . . . . . . . . 61
3.5.2.3 Surface Node Normal Vectors . . . . . . . . . . . . . . . 62
3.5.2.4 Linear Basis Function Values . . . . . . . . . . . . . . . 62
3.6 Measured Data Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.6.1 Steady State Measured Motion . . . . . . . . . . . . . . . . . . . 63
3.6.2 Undamped Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.6.3 Damped Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7 Surface Error Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.7.1 Soft Gelatine Phantom Surface Error . . . . . . . . . . . . . . . . 66
3.7.2 Hard Gelatine Phantom Surface Error . . . . . . . . . . . . . . . 66
3.7.3 Silicone Phantom Surface Motion Error . . . . . . . . . . . . . . 66
3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4 Phantom Surface Motion Error Studies 69

4.1 Homogeneous Gelatine Phantoms . . . . . . . . . . . . . . . . . . . . . . 70
4.1.1 Measured Phantom Displacement . . . . . . . . . . . . . . . . . . 70
4.1.2 Finite Element Simulation . . . . . . . . . . . . . . . . . . . . . . 70
4.1.3 Stiffness Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Homogeneous Silicone Phantoms . . . . . . . . . . . . . . . . . . . . . . 79
4.2.3 Mechanical Property Estimation . . . . . . . . . . . . . . . . . . 83
4.2.3.1 Storage Modulus Frequency Dependence . . . . . . . . 83
4.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 Heterogeneous Silicone Phantoms . . . . . . . . . . . . . . . . . . . . . . 87
4.3.3 Model Parameter Estimation . . . . . . . . . . . . . . . . . . . . 90
4.3.3.1 Corroborative Method . . . . . . . . . . . . . . . . . . . 91
4.3.3.2 Contradictive Method . . . . . . . . . . . . . . . . . . . 93
4.3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5 Non-Linear Reconstruction Algorithms 101

5.1 Gradient Descent Optimization . . . . . . . . . . . . . . . . . . . . . . . 102
5.1.1 Reconstruction Algorithm Structure . . . . . . . . . . . . . . . . 104
5.1.2 Stiffness Reconstruction Results . . . . . . . . . . . . . . . . . . 105
5.1.2.1 Simulated Data . . . . . . . . . . . . . . . . . . . . . . 106
5.1.2.2 Experimental Data . . . . . . . . . . . . . . . . . . . . 106
5.1.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.2 Hybrid Reconstruction Simulation Study . . . . . . . . . . . . . . . . . . 110
CONTENTS ix
5.2.2 Reconstruction Algorithms . . . . . . . . . . . . . . . . . . . . . 113

5.2.2.1 Gradient Descent Reconstruction . . . . . . . . . . . . . 113
5.2.2.2 Combinatorial Optimization . . . . . . . . . . . . . . . 115
5.2.2.3 Hybrid Algorithm . . . . . . . . . . . . . . . . . . . . . 117
5.2.3 Parameter Reconstruction Results . . . . . . . . . . . . . . . . . 117
5.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.3 Hybrid Reconstruction Optimization . . . . . . . . . . . . . . . . . . . . 122
5.3.1 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.3.1.1 Combinatorial Optimization . . . . . . . . . . . . . . . 123
5.3.1.2 Gradient Descent Optimization . . . . . . . . . . . . . . 124
5.3.2 Heterogeneous Phantom Reconstruction Results . . . . . . . . . 126
5.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.4 Spherical Phantom Reconstructions . . . . . . . . . . . . . . . . . . . . . 132
5.4.1 Spherical Silicone Phantom . . . . . . . . . . . . . . . . . . . . . 133
5.4.2 Parameter Reconstruction . . . . . . . . . . . . . . . . . . . . . . 134
5.4.2.1 Spherical Inclusion Phantom . . . . . . . . . . . . . . . 134
5.4.2.2 Homogeneous Phantom Reconstruction . . . . . . . . . 135
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6 Conclusions 143
6.1 Technical Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.2 Clinical Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7 Future Work 147

7.1 Short-Term Development Goals . . . . . . . . . . . . . . . . . . . . . . . 147
7.2 Long-Term DIET System Development . . . . . . . . . . . . . . . . . . . 149
Appendix A: The Finite Element Method 151
Appendix B: Silicone Material Data 159

List of Figures
1.1 The acquired capabilities that differentiate cancerous and healthy cells. 2
1.2 Anatomy of the female breast. . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 A stained pathological specimen from ductal carcinoma in situ. . . . . . 5
1.4 The equipment used and resulting images from a screening mammogram. 9
1.5 MRI patient positioning, and an MRI image produced with a contrast-
enhancing agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Diagnostic ultrasound of the breast. . . . . . . . . . . . . . . . . . . . . 12
1.7 Healthy and abnormal images from breast thermography. . . . . . . . . 13
1.8 A prototype EIT system, and the resulting image of electrical properties. 14
1.9 A tissue sample and elastogram from early research using ultrasound for
elastography. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.10 Modern diagnostic ultrasound elastography of a biopsy-proven breast
cancer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.11 Displacement and stiffness images from a MRE system. . . . . . . . . . 19
1.12 Key components in the DIET system. . . . . . . . . . . . . . . . . . . . 22
2.1 The FE mesh used for computer model simulation in the three-region
proof of concept study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 The geometry used to generate simulated data sets in the three-region
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Surface displacement amplitudes for four of the inclusion scenarios sim-
ulated in the three-region study. . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Convergence of the three-region reconstruction algorithm for a selection
of inclusion scenarios. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
xii LIST OF FIGURES
2.5 A summary of the reconstruction results from the three-region study. . . 31

2.6 The fine mesh used for FE displacement simulation during dual-mesh
reconstructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.7 The stiffness distribution and surface motion from the target model in
the dual-mesh study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.8 A two-dimensional representation of a dual mesh scheme. . . . . . . . . 34
2.9 The best reconstructed stiffness distributions achieved during the dual-
mesh reconstruction study. . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1 Granules of processed gelatine, a raw ingredient in the first experimental

phantoms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 The cylindrical mold developed for gelatine phantom preparation. . . . . 40
3.3 Two gelatine phantoms created during the mold development process. . 41
3.4 The venturi vacuum setup used to remove air from the silicone mixture. 43
3.5 The mold used to cast silicone phantoms. . . . . . . . . . . . . . . . . . 44
3.6 The four silicone phantoms used for experimental motion capture. . . . 46
3.7 The Pyris Diamond DMA machine used for mechanical testing of phan-
tom materials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.8 Static testing results from the gelatine and silicone materials used for
phantom manufacture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.9 The experimental setup for phantom motion capture. . . . . . . . . . . . 50
3.10 A photo of the typical setup for phantom displacement imaging. . . . . 50
3.11 The dSPACE Control Desk software interface. . . . . . . . . . . . . . . . 51
3.12 A schematic detailing the timing of the image capture process. . . . . . 53
3.13 Motion frames from the right camera during actuation of the soft gelatine
phantom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.14 Left and right camera views of the object used for calibration of the
imaging system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.15 An example of reference point position conversion from two dimensional
image co-ordinates to the three dimensional world space coordinate system. 57
3.16 An example of the mesh alignment technique. . . . . . . . . . . . . . . . 58
3.17 An example of the error introduced when a curved surface is discretized
with linear elements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.18 A graphical representation of the steps in the surface projection process. 60
3.19 A reference point and a triangular surface element in a three dimensional
coordinate system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
LIST OF FIGURES xiii
3.20 The evolution of measured motion fitting techniques. . . . . . . . . . . . 63
4.1 The experimentally observed motion of the homogeneous gelatine phan-

toms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 A selection of measured and fitted reference point motion from the soft
and hard homogenous gelatine phantoms. . . . . . . . . . . . . . . . . . 72
4.3 The FE meshes used to simulate soft and hard gelatine phantom motion. 73
4.4 FE-simulated displacement for the soft and hard gelatine phantom mod-
els across a range of Young’s Modulus values. . . . . . . . . . . . . . . . 75
4.5 Motion error metric across a range of homogenous stiffness values for
the soft gelatine phantom. . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.6 A comparison of reference point displacement amplitudes for the exper-
imental soft gelatine phantom. . . . . . . . . . . . . . . . . . . . . . . . 76
4.7 Motion error metric across a range of homogenous stiffness values for
the hard gelatine phantom. . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.8 The match between FE-simulated and experimental displacements for
the hard gelatine phantom.. . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.9 The experimentally observed motion of the homogeneous silicone phan-
toms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.10 The FE mesh used for all silicone phantom simulations. . . . . . . . . . 82
4.11 Displacement sweep surface error results for the soft homogeneous sili-
cone phantom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.12 Displacement sweep surface error results for the hard homogeneous sili-
cone phantom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.13 The best fitting FE-simulated cases compared with measured motions
for both soft and hard silicone phantoms. . . . . . . . . . . . . . . . . . 85
4.14 Motion error plotted against stiffness for the soft silicone phantom ac-
tuated at 50 Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.15 The experimentally observed motion of the stacked and concentric het-
erogeneous silicone phantoms. . . . . . . . . . . . . . . . . . . . . . . . . 89
4.16 Parameter sweep results for the heterogeneous silicone phantoms, using
the corroborative method with Es0 = 28 kPa. . . . . . . . . . . . . . . . . 91
4.17 Corroborative method sensitivity analysis results for the heterogeneous
phantom data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.18 Parameter sweep results for the heterogeneous silicone phantoms, using
the contradictive method with Eh0 = 103 kPa. . . . . . . . . . . . . . . . 93
4.19 Contradictive method sensitivity analysis results for the heterogeneous
silicone phantom data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
xiv LIST OF FIGURES
4.20 The best fitting FE-simulated cases compared with experimental mo-
tions for both heterogeneous silicone phantoms. . . . . . . . . . . . . . . 95
4.21 Parameter sweep results from the homogenous soft silicone phantom,
using the contradictive method with Eh0 = 103 kPa. . . . . . . . . . . . . 96
4.22 The range of modulus values estimated and measured for both soft and
hard silicones. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.1 An example of a gradient descent algorithm. . . . . . . . . . . . . . . . . 103

5.2 A schematic of the gradient descent algorithm used for elastic recon-
struction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3 Gradient descent reconstruction results when assuming that simulated
motion at E = 27.1 kPa represented an experimental data set. . . . . . . 107
5.4 Gradient descent reconstruction results from the real experimental data
captured from the hard gelatine phantom. . . . . . . . . . . . . . . . . . 108
5.5 An enlarged view of the error domain from Figure 5.3(a), detailing the
local minimum caused by a resonant stiffness. . . . . . . . . . . . . . . . 109
5.6 The geometry of both phantom models used in the hybrid algorithm
simulation studies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.7 The displacement error metric from Equation 5.7 evaluated across the
parameter domain for both heterogeneous silicone geometries. . . . . . . 114
5.8 A schematic of the CO reconstruction algorithm. . . . . . . . . . . . . . 116
5.9 The distribution of results from 1,521 GD reconstructions performed on
both heterogeneous geometries. . . . . . . . . . . . . . . . . . . . . . . . 118
5.10 The distribution of results from 1,521 CO reconstructions performed on
both heterogeneous geometries. . . . . . . . . . . . . . . . . . . . . . . . 119
5.11 The distribution of results from 1,521 instances of the hybrid algorithm
performed on both heterogeneous geometries. . . . . . . . . . . . . . . . 119
5.12 A schematic of the hybrid reconstruction algorithm. . . . . . . . . . . . 123
5.13 Results from the sensitivity analysis performed on the CO reconstruction
algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.14 The performance of the GD reconstruction algorithm applied to the con-
centric phantom experimental data. . . . . . . . . . . . . . . . . . . . . 126
5.15 The reconstructed stiffness profile distribution from 100 GD reconstruc-
tions on concentric phantom data. . . . . . . . . . . . . . . . . . . . . . 127
5.16 The stiffness profile distribution from 100 CO reconstructions on con-
centric phantom data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.17 The stiffness profile distribution from 100 hybrid reconstructions on con-
centric phantom data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
LIST OF FIGURES xv
5.18 The reconstructed stiffness profile distribution from 100 hybrid recon-
structions performed on the soft homogenous silicone phantom data. . . 131
5.19 The experimentally observed motion of the spherical silicone phantom. . 133
5.20 A spatial representation of the spherical phantom modulus reconstructed
using the hybrid algorithm with varying numbers of genes. . . . . . . . . 135
5.21 The spherical silicone phantom modulus distribution reconstructed using
a contradictive hybrid algorithm with 100 genes. . . . . . . . . . . . . . 136
5.22 The soft homogenous silicone phantom modulus distribution reconstructed
using a contradictive hybrid algorithm with 100 genes. . . . . . . . . . . 137
5.23 A plane at Es0 = 27 kPa through the spherical phantom error domain. . 139
5.24 A plane at Es0 = 26 kPa through the soft homogeneous phantom error
domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
A.1 A three-dimensional linear tetrahedral element with nodes numbered

according to a right-handed convention. . . . . . . . . . . . . . . . . . . 153
A.2 A Voigt element, where a spring and dashpot are connected in parallel. 156
A.3 Incompressible FE model displacement fields at a range of Poisson’s ratio
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
A.4 Compressible FE model displacement fields at a range of Poisson’s ratio
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
List of Tables
1.1 The 2002 American Joint Committee on Cancer TNM system for breast
cancer classification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1 The ingredients used to make to make the two types of gelatine used for
phantom experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 The weight percentage composition of the two types of silicone used for
phantom preparation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Young’s Modulus values obtained from static testing of phantom materials. 49
4.1 Specifications of the hard and soft gelatine models used for FE simulation. 74
4.2 A comparison of surface error estimated and measured Young’s Modulus
values for the soft and hard gelatine. . . . . . . . . . . . . . . . . . . . . 78
4.3 Elastic modulus values for the soft and hard silicone phantoms. . . . . . 87
4.4 Silicone material properties estimated from the homogenous silicone
phantoms using surface error. . . . . . . . . . . . . . . . . . . . . . . . . 88
4.5 The measured interface position parameter, P , for both heterogeneous
silicone phantom geometries. . . . . . . . . . . . . . . . . . . . . . . . . 90
5.1 The reconstruction algorithms considered during development of the hy-

brid reconstruction method. . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2 The parameter values representing real data for the reconstructions per-
formed using simulated data. . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 The reconstruction algorithm success metric, Ψ1 , across both stacked
and concentric geometries. . . . . . . . . . . . . . . . . . . . . . . . . . . 120
xviii LIST OF TABLES
5.4 The total number of FE simulations required for each reconstruction

algorithm in the two-parameter simulated case. . . . . . . . . . . . . . . 121
5.5 Optimal settings for the CO algorithm determined via sensitivity analysis.124
5.6 Results from the hybrid algorithm optimization. . . . . . . . . . . . . . 130
5.7 Computational cost considerations for the hybrid algorithm applied to
the spherical silicone phantom. . . . . . . . . . . . . . . . . . . . . . . . 141
Nomenclature
Acronyms
BEM Boundary element methods

BLAS Basic linear algebra sub-programs
CO Combinatorial optimization
CCD Charge-coupled device
DCIS Ductal carcinoma in situ
DIET Digital image elasto-tomography
EIT Electrical impedance tomography
DMA Dynamic mechanical analysis
FE Finite element
GA Genetic algorithm
LAPACK Linear algebra package
LVDT Linear velocity displacement transducer
MRE Magnetic resonance elastography
MRI Magnetic resonance imaging
PVC Polyvinyl chloride
SA Simulated annealing
TDLU Terminal duct lobular unit
TNM Tumor, lymph node, metastasis breast cancer classification system
USB Universal serial bus
Roman Symbols
C Stiffness contrast parameter

Ĉ Stiffness contrast corresponding to global minimum error
C1 Stiffness contrast using a corroborative approach
C2 Stiffness contrast using a contradictive approach
C+ Stiffness contrast parameter providing the closest match to experimental data
xx NOMENCLATURE
E Young’s modulus
Ê Young’s modulus corresponding to global motion error
E+ Young’s modulus providing the closest match to experimental data
E∗ Complex elastic modulus
E0 Storage modulus
Es0 Soft material storage modulus
Eh0 Hard material storage modulus
E 00 Loss modulus
Ē 0 Mean reconstructed storage modulus
G Combinatorial optimization gene
G̃ Gene following mating and mutation
G+ Combinatorial optimization gene providing the closest match to experimental data
H Hessian matrix
J Jacobian matrix
M Number of generations in combinatorial optimization
ne Surface element normal vector
nn Surface node normal vector
NG Number of genes in combinatorial optimization
P Interface position parameter
P̂ Interface position corresponding to global minimum error
Ps Interface position for stacked geometry
Pc Interface position for concentric geometry
Ph Inclusion vertical position for spherical geometry
Pd Inclusion diameter for spherical geometry
P+ interface position parameter providing the closest match to experimental data
P̄ Reference point position across all image frames
P∆ Reference point motion path centroid
R Residual term
u Complex-valued displacement amplitude
uR Real component of displacement amplitude
uI Imaginary component of displacement amplitude
ũ Real-valued displacement amplitude
ū(t) Time-varying displacement
um Measured displacement amplitude
uc FE-simulated displacement amplitude
Greek Symbols
Γ Motion path error

γ Hessian matrix regularization term
NOMENCLATURE xxi
δ Gradient descent step size parameter

² Mechanical strain
ζ Damping ratio
θ Parameter combination
θ̂ Parameter combination corresponding to global minimum error
ν Poisson’s ratio
Πmate Mating probability
Πmutate Mutation probability
SA Π Simulated annealing probability
ρ Material density
σ Mechanical stress
Σ2 Variance of reconstructed solutions
σE 0 Standard deviation of reconstructed storage modulus
φ Linear basis function
ϕ Phase
Ψ1 2-parameter simulated reconstruction success metric
Ψ2 3-parameter phantom reconstruction success metric
Ω Non-linear squared error function
Ω1 Undamped phantom surface motion error
Ω2 Undamped/damped phantom surface motion error
Ω3 Damped phantom surface motion error
Ω4 Undamped simulated surface motion error
Ω̂ Global minimum surface motion error
Ω+ Motion error corresponding to the best-matching gene
ω Radial frequency
Abstract
Interest in elastographic techniques for soft tissue imaging has grown as relevant re-
search continues to indicate a correlation between tissue histology and mechanical stiff-
ness. Digital Image Elasto-Tomography (DIET) presents a novel method for identifying
cancerous lesions via a three-dimensional image of elastic properties. Stiffness recon-
struction with DIET takes steady-state motion captured with a digital camera array as
the input to an elastic property reconstruction algorithm, where finite element methods
allow simulation of phantom motion at a range of internal stiffness distributions. The
low cost and high image contrast achievable with a DIET system may be particularly
suited to breast cancer screening, where traditional modalities such as mammography
have issues with limited sensitivity and patient discomfort. Proof of concept studies
performed on simulated data sets confirmed the potential of the DIET technique, lead-
ing to the development of an experimental apparatus for surface motion capture from a
range of soft tissue approximating phantoms. Error studies performed on experimental
data from these phantoms using a limited number of shape and modulus parameters
indicated that accurate measurements of surface motion provide sufficient information
to identify a stiffness distribution in both homogeneous and heterogeneous cases. The
elastic reconstruction performed on simulated and experimental data considered both
deterministic and stochastic algorithms, with a combination of the two approaches
found to give the most accurate results, for a realistic increase in computational cost.
The reconstruction algorithm developed has the ability to successfully resolve a hard
spherical inclusion within a soft phantom, and in addition demonstrated promise in
reconstructing the correct stiffness distribution when no inclusion is present.
Chapter 1
Introduction and Background
Cancer is one of the most significant public health issues in the world, with one in four
deaths in the United States currently directly attributable to the disease. Amongst
all forms of cancers diagnosed in the female population, breast cancer is the most
common, with new cases in 2007 expected to represent a total of 26% of all female cancer
diagnoses. Although deaths from cancer of the lung and bronchus surpassed breast
cancer fatalities among American women in 1987, breast cancer remains the second
highest cause of female cancer death in the United States, with over 100 American
women per day currently dying from the disease [1].
The most recent statistics available from New Zealand indicate that breast cancer is
proportionally at least as significant an issue as in the United States. The 2,235 breast
cancer cases registered in New Zealand in 1999 represented a total of 28% of all female
cancer registrations, the highest proportion of any cancer among New Zealand women.
Significantly, and in contrast to American statistics, breast cancer is the leading cause
of cancer death for New Zealand females, with a fatality/case ratio of 29% [2]. A
related study of cancer mortality covering the 18,990 female breast cancer registrations
in New Zealand in the period 1994–2003 concluded that the majority of breast cancers
were diagnosed at a localized stage, where survival rates are at their highest [3]. The
increase in survival rate associated with early detection provides significant motivation
for screening programmes that are designed to detect breast cancer at the earliest
possible stage.
1.1 Basic Cancer Biology
Cancer initiates when cells within the body multiply at an abnormally high rate. These
groups of cells often form a hard, palpable mass, generically referred to as a tumor,
although other forms of cancer, such as leukemia, do not manifest in this manner. The
spread of cancerous cells beyond a localized area of the body is known as metastasis.
2 CHAPTER 1 INTRODUCTION AND BACKGROUND
The majority of cancer fatalities are a result of this metastatic spread to critical organ
systems, where malignant cells replace and inhibit the function of cells critical for
healthy organ function.
The biological mechanisms underlying cell transformation from healthy to can-
cerous are thought to be linked to changes in cell DNA, though the exact process is
not well understood. Hanahan and Weinberg proposed that cancerous cells acquire an
important set of functional capabilities that govern their transformation into malig-
nant cancers [4]. These capabilities are shown in Figure 1.1, and summarized in the
following.
Self-sufficiency in
growth signals
Evading Insensitivity to
apoptosis anti-growth signals
Sustained Tissue invasion

angiogenesis & metastasis
Limitless replicative
potential
Figure 1.1 The acquired capabilities that differentiate cancerous and healthy cells [4].
The natural process of cell division and growth is controlled by stimulatory signals
provided by the body. It is believed that cancerous cells are able to multiply at rates far
greater than healthy cells due to an apparent self-sufficiency in growth signals [4]. This
reduced dependence on external growth stimulation allows cancerous tissue to expand
rapidly and without reference to the mechanisms that usually ensure regulated cell
development. In addition to a reduced dependence on external stimulation for growth
and division, cancerous cells appear to have decreased sensitivity to the biological anti-
growth signals that are intended to keep the division and growth of normal cells within
healthy limits.
An additional factor leading to unchecked cancerous cell expansion is a lack of
response to apoptosis, or programmed cell death. After receiving an appropriate trigger,
the attrition of healthy cells is a key mechanism in maintaining a healthy balance of
1.2 BREAST CANCER 3
cell types within tissue. While the scientific understanding of the exact mechanisms
behind cell apoptosis is limited, it has been hypothesized that cancer cells are able to
ignore apoptosis signals and live beyond a regular cellular lifespan [4].
The first three acquired capabilities of cancer cells – growth signal autonomy, in-
sensitivity to anti-growth signals, and resistance to apoptosis – allow the growth mech-
anism of cancers to be de-coupled from the cellular signals in their environment. In
theory, this independence from cellular control mechanisms allows a limitless multipli-
cation of cancer cells within the body. It is this unconstrained multiplication, where
cancerous cells are able to invade large areas of healthy tissue, that can cause major
disruption to the functional systems within the body.
For cells to survive and perform their biological function, they must lie in close
proximity to blood vessels that provide essential nutrients and oxygen. Within healthy
human tissue, the ability to create new blood vessels, termed angiogenesis, is highly
regulated. This angiogenesis regulation ensures that the growth of abnormal lesions
is limited by the inability of the lesion to create a supporting vasculature. To sustain
the growth of macroscopic tumors, an intrinsic form of angiogenesis within cancerous
tissue is required, and has been observed in a range of human cancers.
Finally, it is the acquired ability of cancer to metastasize, or move around the body,
that is a key point of difference between cancerous and healthy cells. The exact pro-
cesses controlling this cellular migration are of critical importance to the development
of cancer treatments, but are complicated and not well understood. Survival rates
for almost all types of cancer drop considerably once metastasis has occurred, as the
ability to treat the disease locally, via methods such as radiation therapy, is removed,
and systemic treatments generalized to the entire body, such as chemotherapy, become
a requirement.
1.2 Breast Cancer
The adult female breast, shown in Figure 1.2, is a modified skin gland, comprised of
several different types of tissue. Between 15–20 glandular lobes form an approximately
conical projection from the chest wall, with the apex at the nipple. Included in this
region are lactiferous, or milk producing, ducts that open independently at the surface
of the nipple. These ducts form a complicated branching structure within the breast,
ending in Terminal Duct Lobular Units (TDLUs). Surrounding the central glandular
region of the breast is a layer of subcutaneous fat. The amount and distribution of this
fatty tissue can vary considerably, as can the appearance and position of the interface
with the underlying glandular structures. Both glandular and fatty elements of the
breast are held together by a connective network of fibrous tissue, known as mammary,
or Cooper’s, ligaments. Over time, these ligaments weaken and elongate, causing breast
tissue to sag as age increases. The observed decrease in breast tissue density with age
occurs as the breakdown of fibrous connective tissue means fat provides an increased
proportion of the total breast mass [5, 6].
Figure 1.2 Anatomy of the female breast [6].
The majority of lesions within the breast are believed to form within the TDLU,
with over 70% of breast cancers found at or near the interface between the glandular
structures and fatty tissue within the breast. Most breast cancers begin with an increase
in the number of cells within a localized area of the duct system, known as hyperplasia.
In many cases, this rapid cell multiplication becomes atypical and, if cancerous cells are
involved, is termed ductal carcinoma in situ (DCIS). Figure 1.3 shows a microscopic
view of a DCIS tissue specimen [7]. Cancers within the breast that initially present as
DCIS are estimated to represent 40% of all cancers detected via mammography [8].
The transition from localized to an invasive carcinoma occurs when cancer cells
break out from the ductal system into adjacent ducts, or other areas of the breast and
chest wall. As with a number of other types of cancer, in situ breast cancer, and even
invasive cancer contained within the breast, is not necessarily fatal. It is the metastatic
spread of the disease to other vital bodily systems, such as the lungs, liver, or bowel,
that ultimately cause death.
1.2 BREAST CANCER 5
Figure 1.3 A stained pathological specimen from ductal carcinoma in situ [7].
Following detection, the diagnosis of breast cancer is categorized based on the

size of the tumor, its invasiveness, and the involvement of other parts of the body.
The standard classification system for breast cancer is provided by the American Joint
Committee on Cancer, where the letters T, N and M are used to describe the tumor
size, underarm lymph node involvement, and metastasis state respectively [9]. The
range of classifications for each category in the TNM system is given in Table 1.1.
Alongside the TNM system, breast cancer is often given a stage classification in the
range 0–IV, providing a more straightforward method of conveying the current severity
of the disease. The stage number for a particular case is determined by the combination
of TNM values. DCIS, where the cancer cells have not spread to the surrounding breast
tissue, is referred to as Stage 0 breast cancer. Stage I describes a situation where the
tumor has expanded past the TDLU, is smaller than 2 cm, and has not spread to the
lymph nodes or other distant locations. Stage II covers tumors that are small and
have not spread into the lymph nodes, or larger tumors that are contained within the
breast. Tumors that are mid-sized and have spread to internal sites within the breast
and/or lymph nodes are classified as Stage III. If the cancer has metastasized and is
present in other organs, regardless of the size of any tumor, it is regarded as Stage IV,
where the disease is often terminal. The relative five year survival rate for women with
Stage IV breast cancer is less than 20%, compared with over 95% for localized cases in
Stages 0–II [3]. Determining the staging of breast cancer is a process that can involve a
range of procedures, including x-ray mammography, ultrasound, Magnetic Resonance
Imaging (MRI), and tissue biopsy via either needle aspiration or surgery.
Treatment options for women diagnosed with breast cancer are varied, although
the majority of cases involve surgery of some form. A lumpectomy involves the surgical
Table 1.1 The 2002 American Joint Committee on Cancer TNM system for breast cancer
classification [9].
Category Identifier and Description

Tumor TX Primary tumor cannot be assessed.
T0 No evidence of primary tumor.
Tis Pure carcinoma in situ.
T1 Tumor 2 cm or less in greatest dimension.
T2 Tumor more than 2 cm but not more than 5 cm in
greatest dimension.
T3 Tumor more than 5 cm in greatest dimension.
T4 Tumor of any size growing into the chest wall or skin.
Lymph Nodes NX Regional lymph nodes cannot be assessed.
N0 Cancer has not spread to regional lymph nodes.
N1 Cancer has spread to 1–3 lymph nodes under the arm.
N2 Cancer has spread to 4–9 lymph nodes under the arm.
N3 Cancer has spread to more than 10 lymph nodes under
the arm, or also involves lymph nodes in other areas.
Metastasis MX Presence of metastasis cannot be assessed.
M0 No evidence of metastasis.
M1 Metastasis is present.
removal of a specific area of breast tissue known to contain cancerous cells, and is usually
performed on women with localized breast cancers such as DCIS. This surgery is breast
conserving, and is accompanied in most cases with a period of radiotherapy to ensure
the destruction of all cancer cells in the affected region of the breast. For women with
more advanced cancer, including those with multiple regions of localized cancer within
a single breast, a lumpectomy may not be the most appropriate option. In these cases,
as a simple mastectomy, involving removal of the entire breast, is performed. This
procedure may also include the removal of some or all of the lymph nodes from the
underarm region, in a procedure known as a modified radical mastectomy [10].
Systemic treatment options are often used in addition to surgery in an attempt

to kill cancer cells that may have spread beyond the breast into other areas of the
body, even if there is no direct evidence of such spread. For women with breast cancers
that contain estrogen receptors, hormone therapy can be used to limit the growth of
the tumor by blocking the supply of estrogen to the affected region. Chemotherapy
involves the oral or intravenous application of cancer fighting drugs, and is often used in
combination with hormone therapy. Recent advances in genome research have led to the
1.3 BREAST CANCER SCREENING 7
increased popularity of gene therapy in the treatment of breast cancer. Gene therapy
involves using targeted drugs to combat the growth of specific types of breast cancer
using the body’s natural defence mechanisms. The side effects of systemic treatments
can be wide and varied, and are often controlled with the use of additional drugs.
1.3 Breast Cancer Screening
Breast cancer screening aims to reduce cancer mortality by detecting the disease at the
earliest possible stage. The intent is not to prevent breast cancer, but to reduce the
risk of dying from the disease. Screening programs involve widespread testing of the
general population, or at-risk groups, to identify individual cases that warrant further
investigation. Such programs require significant funding and resources, often sourced
from central or local government, to encourage target population compliance.
The use of breast self examination is widespread for detecting the presence of suspi-
cions lesions at an early stage [11]. This technique relies on the underlying mechanical
properties of breast tissue, where tumors often present as a discrete mass significantly
stiffer than the surrounding tissue [12]. However, such an examination is only able to
provide a qualitative assessment of tissue properties, limiting its usefulness for further
diagnosis of tissue histology. In addition, a number of studies have challenged the effec-
tiveness of the breast self exam as a useful tool for reducing the occurrence and fatality
rates of advanced stage breast cancer [13–15].
Twelve countries have currently implemented systematic breast cancer screening
programs [16]. The program in the United States has grown considerably in both size
and participation level over the past two decades, with the proportion of American
women over the age of 40 who have had a mammogram in the previous two years rising
from 29% in 1987 to over 70% in 2000. The American Cancer Society aims to have
90% of all women over the age of 40 in a screening program by 2008 [17].
In New Zealand, the first trials of government-funded breast cancer screening using
mammography were performed in 1991. Following this initial trial, a nationwide breast
cancer screening program was announced in 1995. BreastScreen Aotearoa was officially
launched in December 1998, and currently offers a free biennial mammogram to all
women aged between 45 and 69 with no previous history of cancer [18, 19].
The majority of literature supports the conclusion that screening for breast cancer
using mammography reduces mortality rates by detecting small tumors that would not
be identified by manual palpation. This early detection allows swift treatment while
the disease is still at a localized stage. A large study performed by Tabar et al. found
a 44% reduction in breast cancer mortality in the 40–69 age group after screening
was introduced for Swedish women in 1978 [20]. The majority of this reduction in
deaths was attributed to breast cancer screening. Seven other major trials have all
indicated that screening with mammography leads to a 20–30% reduction in breast

cancer mortality [16]. Though these trials have been critiqued by Olsen et al. [21], who
contended that none were of a high-quality, the weight of clinical evidence suggests
systematic breast cancer screening can save lives if performed correctly and accurately.
Two important metrics when assessing a cancer screening program are sensitivity
and specificity. Sensitivity refers to the proportion of the population with breast cancer
who will have a positive test using the screening modality. High sensitivity is desirable
for a screening modality in order to correctly identify those individuals who currently
have breast cancer. Specificity is a complimentary term that refers to the proportion of
people without cancer who test negative using the screening modality [22]. Sensitivity
is more important than specificity in a widespread screening program, as false positive
results can be compensated for by follow-up diagnosis. However, an ideal screening
method combines high sensitivity and high specificity, reducing false-positives, which
can have a negative psychological impact on the patient, and false-negatives, which
lead to increased patient mortality [16].
1.3.1 Mammography
Mammography is currently the most widespread clinical breast imaging method for
both screening and diagnostic purposes, and is currently regarded as the gold standard
for breast cancer screening [22]. The image contrast in a film mammogram is based on
the difference in propagation of x-rays through breast tissues. Various tissue properties
such as thickness, density and atomic composition alter the attenuation of the x-rays,
where areas on the resulting film that have had high x-ray attenuation appear lighter [5].
Standard mammographic procedure, shown in Figure 1.4(a) for a craniocaudal
mammogram, involves placing the breast between an x-ray source and a detector. The
breast is then compressed between paddles before the image is exposed. Compression
helps to even out the thickness of the breast over the image area, reducing the exposure
difference between the periphery and the center of the breast. Additionally, compression
reduces blur due to patient motion, and allows tissue close to the chest wall to be drawn
onto the projector. A standard screening mammogram involves both a craniocaudal
view in the horizontal plane, and mediolateral oblique view in the vertical plane. Both
images are shown for a healthy breast in Figures 1.4(b)-(c). The use of two views allows
examination of the internal structure of the breast from multiple perspectives, reducing
the chance of a false negative result. Following x-ray image development, a radiologist
examines the films and determines whether further diagnosis is required.
The interpretation of a mammogram requires skill, experience, and sound clinical
systems to reduce the incidence of false negative cases. A radiologist will often be
required to interpret screen-film mammograms in large batches, often with as little as
one minute available to analyze each image. With a low proportion of mammograms
X-ray
generator
X-ray beam
Compression
system
Detector plate
(a) Equipment and patient positioning for a craniocaudal mammogram.
(b) Craniocaudal projection. (c) Mediolateral oblique projection.
Figure 1.4 The equipment used and resulting images from a screening mammogram [23, 24].
having radiological abnormalities, fatigue or lack of attention introduces significant

potential for human error [16, 25].
Kriege et al. reported that the sensitivity and specificity of screening mammog-
raphy is 33% and 95% respectively, and that the sensitivity of mammographic breast
screening is less than ideal [26]. A separate study performed using professionals ac-
credited by the American College of Radiology indicated a variation of at least 11%
in the ability of a radiologist to discriminate between a mammograms from women
with and without breast cancer [27]. An Italian study by Ciatto et al. reported that
double-reading of screening mammograms, where each image is interpreted by multiple
radiologists, significantly reduced the incidence rate of false-negatives [28].
The low contrast of the mammographic image creates issues when identifying ab-
normalities, particularly in young women, where high breast tissue density has been
linked to a decrease in the effectiveness of mammography [29]. When the x-ray at-
tenuation of the surrounding tissue is largely the same as a tumor, and there are no
noticeable calcifications, it is possible for a tumor to be invisible in a mammogram.
This lack of image contrast accounts for 30–50% of all false-negative results [5].
Computer-aided detection and the use of digital mammography is becoming more

widespread as part of an effort to increase the sensitivity of mammography. Such
systems use computers to process images from digital mammography and automatically
identify suspicious regions [22]. Other techniques such as computer aided thickness
correction have been used to enhance the periphery of the mammographic image [30].
A significant reason for non-compliance with mammography screening programs is

the expected and/or experienced discomfort associated with breast compression. The
pain and discomfort experienced by women undergoing mammography has been inves-
tigated by several studies, yet is not well understood [31]. Using a visual analogue scale,
Ashgari et al. found that 92% of patients reported a mammogram as being painful, a
similar result to the 88% of painful mammography experiences reported by Kashikar-
Zuck et al. [32, 33]. In general terms, patients report pain during mammography at a
low to moderate level, although the pain levels experienced show high variability. The
cause of this variation in discomfort is not conclusively understood. Kashikar-Zuck et al.
found no significant relationship between breast density and pain levels, a conclusion
that conflicts with the findings of Kornguth et al., who observed a connection between
breast density and discomfort [34]. Although the percentage of women that reported
severe pain levels during mammography is low, a number of patients indicated they
would not have further mammography due to the pain they experienced. In addition
to reduced compliance among women who have experienced discomfort while having a
mammogram, the perceived discomfort amongst those women who have not enrolled
in a screening program can be a significant barrier to widespread participation.
1.3.2 Alternative Breast Imaging Modalities
MRI uses powerful magnetic fields to create images of the human body by measuring the
absorption and emission of radio frequency waves. When applied to soft tissue, MRI can
be used to generate highly detailed images of the internal structure of the breast. MRI
signal strength, and therefore image contrast, is determined by the density of hydrogen
protons within the imaged tissue and their response to rapidly changing radio waves
and magnetic fields. The image contrast of MRI systems can be significantly enhanced
using a range of chemicals ingested or injected prior to the scan [35, 36]. An example
of the equipment used for breast MRI and the type of image that can be obtained from
such a system is shown in Figure 1.5 [37, 38].
(a) Patient positioning for breast MRI, where the (b) A contrast-enhanced, color cor-
breasts hang pendant beneath the patient. rected breast MRI scan clearly in-
dicating a tumor.
Figure 1.5 MRI patient positioning, and an MRI image produced with a contrast-enhancing
agent [37, 38].
While not as prevalent as mammography, MRI has seen limited application as a

breast cancer screening modality. A large study in the Netherlands concluded that
breast cancer screening with MRI was more sensitive for at risk women than mammog-
raphy [26]. However, MRI was also found to have a lower specificity, indicating that
screening using MRI alone would lead to a much higher recall rate for false-positives.
Further evidence supporting the use of MRI for breast cancer screening was presented
by Warner et al., who concluded that breast MRI may be superior to mammography
for detecting small breast cancers in women at high risk [39]. However, the high cost of
MRI hardware and the difficulty in producing clear and consistent images has prevented
its adoption in widespread screening programs [40].
Medical imaging with ultrasound involves the use of high frequency acoustic signals
to form an image of subsurface tissue structure. Ultrasonic waves introduced using a
hand-held scanner are reflected at discontinuities within the tissue, such as the interface
between different tissue types. Precise measurement of these reflected waves allows
an image representing the internal tissue structure to be calculated and displayed.

Figure 1.6 shows an example of a clinical ultrasound examination, where the resulting
sonogram shows a large breast mass [41].
(a) Technician performing a breast ultrasound (b) A color-enhanced ultrasound image

exam. showing a large breast mass.
Figure 1.6 Diagnostic ultrasound of the breast [41].
Ultrasound is not currently approved in the United States for widespread screen-
ing in women with no previous history of breast cancer. However, it is often used as
a diagnostic tool in a follow-up examination if abnormalities are detected using mam-
mography. In these scenarios, ultrasound has proven useful as a tool for identifying the
possible malignancy of a detected mass [42]. Limitations of ultrasound systems include
a low spatial resolution and a limited ability to detect microcalcifications, which are
often an early warning sign of breast cancer [22].
The effect of breast tumors on the temperature of overlying skin was first identified
in the 1950s, when declassified military technology was adapted to produce accurate
thermal images of the human body [43]. Thermal imaging techniques measure the
infrared emissions from human skin, where the emitted signals have a wavelength of
approximately 5–15 µm. Using sensitive thermal imaging cameras and a carefully con-
trolled laboratory environment, accurately calibrated thermograms of the chest region
can be obtained, with examples shown in Figure 1.7. Early research in breast ther-
mography relied on entirely subjective image analysis, leading to large variations in the
effectiveness of such systems across a range of studies. The use of computers to assist
with image capture and processing in modern breast thermography has significantly
improved the accuracy of these systems.
Thermographic breast images are graded according to a thermobiological classi-
fication on a scale from TH1 (normal) through TH5 (highly abnormal), where the
abnormal surface temperature distribution in Figure 1.7(b) is indicative of a subsur-
face breast lesion. The use of breast thermography in conjunction with mammography
1.4 ELASTOGRAPHY 13
(a) A TH1 (normal) thermogram image. (b) A thermogram showing a severe abnormal-
ity (TH5) in the patient’s right breast.
Figure 1.7 Healthy and abnormal images from breast thermography [44].
is widespread, with both systems providing physicians with complimentary informa-

tion [44].
Electrical Impedance Tomography (EIT) is a novel breast imaging system that
uses low level electrical currents to create images of internal breast structure. EIT
systems are based on the observed correlation between tissue pathology and electrical
properties, where malignant breast tumors have shown lower electrical impedance than
healthy tissue [45]. Cherepenin et al. developed a three dimensional EIT system that
had potential to be used for breast cancer detection [46]. A similar system developed at
Dartmouth College is shown in Figure 1.8, and has been successfully used to identify
subsurface breast abnormalities [47]. Breast imaging using microwaves is a similar
area of research based on experimental data that indicate cancerous tissue has a lower
dielectric permittivity than healthy breast tissue [48, 49]. Both EIT and microwave
systems for breast imaging have shown promising initial results, but are yet to undergo
widespread clinical trials.
1.4 Elastography
Despite the questionable effectiveness of manual palpation, the mechanical property

contrast measured by breast self examination has inspired a range of research into
the accurate quantification of soft tissue elastic properties. One of the first significant
studies in quantitative analysis of soft tissue mechanical properties was performed by
Ophir et al., who defined the term elastography when referring to measurement of the
elastic properties of biological tissues. The resulting image, where tissue stiffness is
visualized in a two-dimensional plane through a specimen, is termed an elastogram.
Using an ultrasound scanner and cross-correlation techniques, elastographic eval-
uation was performed on a range of foam phantoms, and a block of porcine tissue,
(a) EIT imaging hardware, where the patient’s (b) Coronal image through the breast
breast is surrounded by electrodes within the cir- showing absolute permittivity values.
cular opening. A malignant tumor is visible in the up-
per right quadrant.
Figure 1.8 A prototype EIT system, and the resulting image of electrical properties [47].
shown in Figure 1.9(a). The sonogram resulting from an ultrasound scan of the tissue
was converted to an image of elastic modulus based on the correlation between sets
of images taken before and after a small sample compression. Although the resulting
elastogram, shown in Figure 1.9(b), has a low spatial resolution, the striated nature
of the tissue is apparent, with bands of soft fatty tissue identified as light horizontal
lines. The promising results from this and other similar studies demonstrated that
quantitative assessment of tissue elasticity was possible, and that the elastograms pro-
duced using an ultrasound system had potential uses in a wide range of applications,
including diagnostic medical imaging [50, 51].
(a) Porcine tissue sample, with visible stria- (b) The corresponding tissue elastogram,
tions of fat and muscle. where bright pixels represent regions with a
lower elastic modulus.
Figure 1.9 A tissue sample and elastogram from early research using ultrasound for elastog-
raphy [50].
1.4 ELASTOGRAPHY 15
1.4.1 Mechanical Properties of Soft Tissue
Mechanical property measurement of in vivo or ex vivo human soft tissue is a challeng-

ing field, with only a limited number of studies able to provide quantitative results using
traditional mechanical testing techniques. Biological soft tissue has features common
to both solid and liquid materials, making accurate characterization a difficult process
largely dependent on the assumptions about mechanical behavior made when perform-
ing testing. These assumptions regularly include treating the material as an elastic,
homogeneous, isotropic solid that is close to, or completely, incompressible, allowing the
stiffness of samples to be expressed as a single modulus value, where Young’s Modulus,
E, is commonly used.
Sarvazyan et al. provide a summary of early research in soft tissue elastography,
presenting both collated and experimentally generated mechanical testing results from
a range of human tissue types, including soft tissue and cartilage [52, 53]. The elastic
modulus results show large variation across the range of tissue types, and within each
tissue type, indicating the large experimental uncertainties often associated with testing
biological samples. However, an observed contrast in elastic modulus between healthy
and cancerous tissue of 1–2 orders of magnitude was postulated as the basis of an
imaging modality that could have particular relevance to breast and prostate tissue.
Krouskop et al. developed a mechanical testing system using a hydraulic servo
apparatus that allowed compressive viscoelastic mechanical testing of small tissue sam-
ples using a sinusoidally displaced indenter [54]. A total of 142 breast tissue samples
were tested at room temperature soon after excision in a surgical environment. These
samples represented a range of healthy fatty, glandular, and fibrous breast tissue, as
well as both intraductal and infiltrating ductal carcinoma. All samples were tested
at pre-compression values of 5% and 20%. For actuation frequencies between 0.1–
4 Hz, breast tissue was found to be only slightly viscoelastic, with modulus values not
strongly dependent on actuation frequency. When actuated at 1 Hz, fatty breast tissue
showed only small variations in elastic modulus over a large strain range, with elastic
modulus values of approximately 20 kPa. Results from glandular and fibrous breast
tissue samples indicated that both types were significantly stiffer than fatty tissue.
Glandular tissue modulus values were in the range 28–66 kPa, while fibrous breast tis-
sue was measured at between 96–244 kPa. Modulus values for glandular and fibrous
breast tissue displayed a high dependence on the amount of sample pre-compression,
indicating a non-linear elastic response to loading. A significant result of this study
was the consistently high measured elastic modulus values for cancerous tissue, with
contrast values up to 700% when compared to healthy fat and glandular tissue.
Wellman et al. developed a portable punch indenter system that allowed tissue
samples to be tested in the operating room immediately following excision [55]. A total
of 68 samples representing eight different types of breast and cancerous tissue were
statically tested to a strain of approximately 0.15. Fatty and glandular tissues were
found to have an approximately linear stress-strain relationship, with elastic modulus
values in the range 5–90 kPa at strain values less than 0.1. Carcinoma and other be-
nign lesions displayed non-linear elastic behavior, with a 500–1200% contrast observed
between carcinoma and fatty breast tissue. This study provides further evidence of a
significant correlation between tissue histology and observed stiffness.
Mechanical testing of excised tissue traditionally requires relatively large, homoge-
neous samples that are challenging to obtain in a surgical environment. The difficulty
in obtaining consistent samples was addressed by Samani et al., who developed a test-
ing system designed to compensate for biological tissue sample limitations, such as
small size, irregular shape, and inhomogeneity [56]. Small blocks of unconstrained ex-
cised tissue were indented in a quasi-static manner using a computer controlled loading
apparatus, with the resulting force measurements recorded. An approximate Young’s
Modulus value was obtained from force-displacement data using a conversion factor
based on sample geometry and calculated using a finite element (FE) model. Tissue
modulus values were observed to be non-linear over the small strains applied, with
Young’s Modulus values for fatty and fibroglandular breast tissues of approximately
2 kPa. In comparison, high grade ductal carcinoma had measured Young’s Modulus
values of approximately 12 kPa, a modulus contrast of approximately 600% when com-
pared to healthy breast tissue.
Subsequent refinement of the mechanical testing system developed by Samani et al.
allowed a more accurate characterization of tumor stiffness using larger tissue samples
containing intact tumors. This modified system was verified against small synthetic
samples of known stiffness [57]. A comprehensive study of 169 breast tissue samples was
performed using slightly different experimental setups for healthy tissue samples and
those containing tumors, based on the different sample geometries. The mechanical
testing results again indicated similar Young’s Modulus values for fatty and fibrog-
landular breast tissue at low strains, with both tissue types having observed modulus
values of approximately 3.5 kPa. Statistical analysis indicated that there was a signif-
icant difference in stiffness between a range of breast cancers and healthy tissue, with
Young’s Modulus values for high grade invasive ductal carcinoma more than 10 times
greater than healthy tissue. Breast fibroadenoma, which are usually benign, were found
to have stiffness values between those of healthy tissue and carcinoma. This result indi-
cates the potential of accurate stiffness characterization to differentiate between benign
and cancerous lesions within the breast. Such characterization could improve the speci-
ficity of elastographic imaging modalities by reducing the occurrence of false positive
results [58].
While experimentally tested soft tissue modulus values show large variations de-
pending on a range of factors specific to each testing methodology, a clear distinction is
apparent between healthy and cancerous breast tissue. This observed elastic contrast
1.4 ELASTOGRAPHY 17
has encouraged the development of a range of techniques that aim to determine the
distribution of elastic properties throughout the breast. A number of these approaches
are based on existing medical imaging modalities, where a modified application of the
hardware and/or software allows the generation of elasticity images that can assist in
determining tissue pathology.
1.4.2 Ultrasound Elastography
In addition to traditional clinical use, ultrasound technology has been implemented as

an elastographic breast imaging tool. Displacement of the breast during an ultrasound
scan allows the tissue displacement response to be expressed as a strain image. Early
research in ultrasound elastography used ultrasound scanning machines to measure
the mechanical response of both phantoms and soft tissue to harmonic actuation. By
assuming constant material stress throughout the material, the recorded strain infor-
mation was converted directly into a description of elastic modulus using basic Hookean
theory. Results from these studies indicated that ultrasound elastography had signifi-
cant potential to detect abnormal lesions within breast, prostate and liver tissue based
solely on tissue stiffness [59–62].
Recent results have clearly demonstrated the potential of ultrasound elastography
in diagnostic clinical medicine. Barr et al. developed a freehand ultrasound elastog-
raphy technique that was applied in conjunction with conventional ultrasound anal-
ysis [63]. This method used the relative size of suspicious lesions in sonograms and
elasticity images as an indicator of tissue pathology. As shown in Figure 1.10, cancer-
ous tumors with high stiffness contrast to the surrounding tissue appeared significantly
larger in the elastographic image.
A study performed on 80 patients with a total of 123 previously identified breast
lesions indicated this technique had a sensitivity of 100%, with all 17 malignant tumors
positively identified. A 99% specificity was also achieved, with 105 of 106 benign lesions
correctly identified based on their smaller appearance in the resulting elastographic
image. While ultrasound elastography studies such as these have shown significant
promise in diagnostic imaging, their widespread application in a breast cancer screening
program is limited by a number of factors, including a lack of off-axis resolution and
geometric repeatability.
The strain-based approach to ultrasound elastography relies on the assumption
of constant material stress in order to generate an elastic property image. Model-
based elastography is an alternative technique where stress values can vary across
the domain of interest, requiring the solution of an inverse problem to determine the
elastic property distribution [64]. Recent results have indicated that these forms of
inverse problems are ill-posed, meaning more than one exact mathematical solution
to the problem exists [65, 66]. However, studies in both transient and steady state
(a) A conventional ultrasound image, where (b) The corresponding elasticity image,
an invasive ductal carcinoma appears as a where the lesion appears significantly larger.
dark region.
Figure 1.10 Modern diagnostic ultrasound elastography of a biopsy-proven breast cancer [63].
model-based ultrasound elastography have yielded comparable results to the strain-

based approach, indicating that the mathematical challenge of non-uniqueness can be
overcome with careful choice of inverse algorithm [67–69].
1.4.3 Magnetic Resonance Elastography
In addition to the static images obtained in traditional medical diagnosis, MRI has the
ability to detect three dimensional motion within a displaced sample. Plewes et al. used
MRI to obtain quasi-static strain images from the breasts of volunteers, and used these
images to determine the bio-mechanical properties of previously-detected lesions [70].
Muthupillai et al. demonstrated a technique where an oscillating magnetic gradient
could be used to measure tissue amplitude when harmonically perturbed [71]. Strain-
based property estimation has been used to calculate measurements of tissue modulus
from the mechanical wavelength or velocity of the propagating shear waves in breast
tissue [72].
Van Houten et al. developed a harmonic Magnetic Resonance Elastography (MRE)
system for breast imaging [73, 74]. In combination with a sub-zone based inversion
algorithm, this system was able to reconstruct elastic property distribution of both
phantoms and in vivo breast tissue, with results from a breast with an high stiffness
inclusion shown in Figure 1.11. The mechanical properties reconstructed in this study
compared well with values from independent mechanical testing. Development of this
1.4 ELASTOGRAPHY 19
sub-zone approach using phantoms indicated that the system was capable of detecting
inclusions as small as 5 mm using a model-based elastic inversion algorithm [75, 76].
(a) MR detected displacement image. (b) The reconstructed elastic modulus distri-
bution, where a fibroadenoma is visible in the
lower right quadrant of the image.
Figure 1.11 Displacement and stiffness images from a MRE system [74].
1.4.4 Novel Elastographic Techniques
Several novel methods for elastographic tissue imaging are currently in development.
Kirkpatrick developed an optical elastography technique for measuring strain values
in biological tissue [77, 78]. This system used a CCD camera to capture the harmonic
motion of samples of chicken skeletal muscle. The observed change in speckle patterns
between subsequent images was used to calculate strain field data that could be used
to infer mechanical properties. While not currently intended for breast elastography,
such a system may has relevance for the measurement of strains as part of a surface
measurement based elastographic system.
Model-based computer-aided registration of breast images produced using tradi-
tional imaging modalities has shown potential for breast elastography. Rothney et al.
used a FE model with assumed elastic properties to displace images previously obtained
with a breast MRI system [79]. Using a range of image similarity metrics to compare
the original and displaced images, reconstruction of tumor location was demonstrated,
including situations with multiple inclusions. Barbone et al. have reported results
from a similar system, where an ultrasound scanner was used to generate images from
an experimental phantom, and image registration used to identify the presence of an
interior inclusion [80].
Recent work by Liu et al. has indicated the feasibility of elastographic property
reconstruction from a set of boundary measurements [81]. Such a system differs from
ultrasound or MRI-based elastography techniques in the limited amount of data avail-
able for elastic property reconstruction. Using an optimization-based reconstruction
algorithm, measurements of force and displacement at external and internal boundaries
within a simulated phantom model were used to reconstruct elastic properties. A sin-
gle static displacement mode was used for an isotropic elastic property reconstruction,
while anisotropic reconstruction required data from a combination of multiple inde-
pendent loading modes. While this method is currently limited to idealized simulated
data, the promising results indicate that elastographic reconstruction is possible using
only a small subset of the large data sets used for full-volume based reconstruction
techniques.
Alongside developments in data acquisition and registration, research into recon-
struction algorithms for solving the elastic inverse problem is ongoing. While many of
these algorithms are based on traditional approaches to non-linear reconstruction such
as gradient-descent methods, there is increasing interest in the use of stochastic meth-
ods for solving the large, highly non-linear problems often found in elastography. The
stochastic approach to non-linear optimization is summarized in the following section,
with particular focus on developments in soft tissue elastography.
1.5 Stochastic Optimization in Elastography
A stochastic process is one where the changes between consecutive states of a system
contain an element of randomness, and cannot be exactly predicted. Such systems differ
from deterministic processes, where subsequent system states can be predicted using
known information about the current and previous configurations. Stochastic opti-
mization algorithms include an element of random behavior in an attempt to overcome
limitations in traditional algorithms, such as a heavy reliance on starting conditions.
Combinatorial Optimization (CO) is a form of stochastic algorithm that combines
a set of possible solutions, termed a population, in a semi-random manner. CO aims
to minimize the value of an objective function, often represented by an error term.
Genetic Algorithms (GAs) are a form of CO that perform population combination in
a manner similar to the naturally occurring process of genetic evolution. The overall
‘fitness’ of the population increases over time, as strong solutions with a low error value
are retained across generations, while weak solutions are more likely to be eliminated.
Subsequent generations are created by combining the ‘genetic material’ of solutions
within the population in a manner similar to reproduction in the animal world.
In many GAs it is necessary that parameters for optimization are converted into a
standard form using a genetic encoding process. This encoding allows the optimization
1.5 STOCHASTIC OPTIMIZATION IN ELASTOGRAPHY 21
of several independent parameters with different units. Incorporating the naturally

occurring phenomena of genetic mutation into the reconstruction algorithm further
increases the diversity of solutions by introducing controlled, random variation to the
population [82–86].
Annealing is a metallurgical process involving several stages of controlled tempera-

ture variation that allows the microstructure of a material to reach a minimum energy
state. Kirkpatrick et al. first reported the application of annealing theory to problems
in CO in the early 1980s, describing a strategy where CO solutions that decreased the
overall system ‘energy’ were increasingly preferred over time [87]. Allowing a limited
number of solutions that increased the system energy provided possible additional path-
ways to a global minimum error state. One interesting conclusion from this work was
the limited aspects of artificial intelligence observed in such algorithms when applied
to the common ‘travelling salesman’ optimization problem [88].
Applications of CO in the field of elastic property reconstruction are limited, al-

though increasing in popularity as more powerful computer systems become main-
stream. Matsuoka et al. used a GA in combination with FE methods to solve for the
mechanical properties and internal structure of composite materials [89]. Here, the
inverse algorithm optimized the internal structure of a composite to obtain a desired
Young’s Modulus value. Olmi et al. performed a preliminary investigation into the
use of a GA for image reconstruction in EIT [90]. Here, a numerical model was used
to simulate the electrical properties of a simple geometry containing a high resistivity
inclusion. The GA used in this study to perform image reconstruction was found to
have significantly higher computational cost than more traditional Newton-Raphson
methods, though this increase was offset by a clear improvement in reconstructed solu-
tion accuracy. Advances in computer processing were seen as a pathway to the possible
implementation of such optimization techniques in realtime EIT systems.
The use of stochastic techniques in soft tissue elastography is a new and emerging
field. Zhang et al. developed a theoretical framework for the application of a GA to
problems in soft tissue elastography [91]. To ensure numerical stability, this method
constrained solutions using an assumed a priori property distribution. Simulating dis-
placements with a synthetic numerical model, stochastic methods were found to give
consistently accurate results for the elastographic reconstruction problem. Khalil et al.
developed a combined GA and FE approach to solve for elasticity values in vascular
soft tissue [92]. This study lumped material parameters into discrete regions in order to
simplify the genetic encoding process, and used a numerical model to simulate displace-
ments in a 2D model. Successful elastic property reconstructions were performed on
both homogeneous and heterogeneous models with multiple inclusions in the presence
of noise.
1.6 Digital Image Elasto-Tomography
Digital Image Elasto-Tomography (DIET) is a novel technique intended to detect and

identify areas of high tissue stiffness within the breast, where cancerous lesions can be
identified due to their high elastic modulus values. Utilizing digital imaging sensors,
image processing algorithms, and computer reconstruction techniques, DIET aims to
generate a three dimensional representation of elastic properties throughout the breast
volume. The system is targeted at providing an inexpensive, comfortable screening test,
where the associated hardware is compact enough to make portable screening units a
reality. The three-dimensional nature of the resulting elastogram should allow accurate
image registration between subsequent scans, assisting in tracking changes in internal
breast structure over time. The DIET system can be split into several key stages,
summarized in Figure 1.12. While the final configuration will integrate all stages into
a single unit, each step in the process has a degree of independence that allows limited
development in isolation from associated aspects of the system.
Breast with high

stiffness inclusion
Actuator
Control
Camera Control,
Image Capture
and Processing
Stiffness Tissue
Reconstruction Stiffness
Algorithm Image
Figure 1.12 Key components in the DIET system.
Mechanical breast actuation is used to induce a steady state, harmonic motion

throughout the patient’s breast volume. It is currently envisaged that the patient
1.7 FROM PROOF OF CONCEPT TO PHANTOM ELASTOGRAPHY 23
lies prone on a table with the examined breast hanging in the pendant position, and
actuated vertically from below by a small actuator placed against the surface of the
breast. Accurate control of the breast actuator allows a precise sinusoidal displacement
to be applied at a known frequency and amplitude. Once the motion throughout the
breast volume has reached a steady-state, imaging of the breast surface can commence.
The steady-state motion pattern at the surface of the breast is captured using a
digital imaging sensor array surrounding the breast surface. Calibration performed
prior to breast imaging allows a set of camera-specific parameters to be generated for
each sensor, describing the three dimensional position and orientation of each cam-
era with respect to a global origin. Strobe lighting allows consecutive images of the
steady-state surface motion to be captured at a lower frequency than the actuation,
eliminating the requirement for an expensive, high speed motion capture setup. Stereo
imaging techniques are used to convert the consecutive two-dimensional images from
all cameras into a three-dimensional vector description of the breast surface motion,
where displacements are measured at a set of reference points on the breast surface.
The final stage in the DIET process is a reconstruction that converts the surface mo-
tion provided by the image capture system into a three dimensional description of breast
tissue stiffness. Here, a non-linear reconstruction algorithm is used in conjunction with
a FE model, and minimizes the motion error between observed and computer-simulated
breast surface displacement fields. The output from this reconstruction algorithm is a
full-volume description of elastic breast properties. This stiffness image should clearly
indicate suspicious areas based on their high property contrast with healthy tissue.
1.7 From Proof of Concept to Phantom Elastography
This thesis outlines the progressive steps taken from the theoretical concept of surface
based mechanical property reconstruction through to the current implementation of the
DIET system hardware. Chapter 2 describes an initial proof of concept trial performed
to evaluate the DIET approach to stiffness reconstruction. This experiment involved a
basic simulation study that used FE methods to simulate breast model displacement,
with stiffness reconstruction performed using inverse algorithms based on previous re-
search in MRE. This study confirmed the feasibility of surface-based property recon-
struction, and identified several issues relevant to the experimental implementation of
DIET.
The phantoms and equipment used for experimental trials of the DIET system are
described in Chapter 3. The details of phantom development, including the progres-
sion from simple homogeneous phantoms to more realistic geometries containing high
stiffness regions, are considered. The actuation and motion capture system used for
experimental data collection is described,along with a brief consideration of the pho-
togrammetry required for three-dimensional point tracking. The iterative refinement

of the experimental system over time is discussed in approximately chronological order,
where improvements in phantom preparation and experimental hardware were made
across the collection of multiple phantom data sets.
Chapter 4 describes an investigation into the experimental surface motions obtained
from the tissue-approximating phantoms. These experiments consider the correlation
between FE-simulated motion and the motion captured from the experimental phan-
toms. The sensitivity of surface motion error metric values to variations in phantom
material properties is investigated, with the results providing useful information when
considering the design of a DIET stiffness reconstruction algorithm. In addition, this
chapter details the refinement of the FE model used for motion simulation across the
motion studies performed.
The reconstruction algorithms developed to solve the elastic inverse problem are
described in Chapter 5. Initial results are presented using traditional deterministic
non-linear reconstruction techniques, where the advantages and limitations of such
methods when applied to the DIET problem are considered. The development and
implementation of a novel phantom stiffness reconstruction algorithm is described,
where techniques from stochastic optimization theory are combined with traditional
methods to form a hybrid reconstruction algorithm. Results from both simulated and
experimentally obtained phantom data sets are presented, and show the improvement
in reconstruction accuracy with the refinement of the reconstruction algorithm. A
consideration of the computational costs associated with the range of reconstruction
algorithms investigated is also provided.
Finally, Chapters 6 and 7 provide a summary of the major findings of the thesis,
and an outline the areas of system improvement requiring focus in the future. Further
research in these areas will allow more accurate phantom stiffness reconstruction using
the DIET system, with an eventual aim towards a limited clinical trial on human
subjects. The implementation of the DIET system in such a trial will present a number
of significant challenges, but will also represent a major milestone for the project.
Chapter 2
Proof of Concept Studies
The DIET system was originally envisaged as an extension to previous research in MRE,
where positive results from both phantom studies and in vivo testing have provided solid
evidence that motion data sets can be used to determine the elastic properties within
a volume of soft material, specifically breast tissue. Using carefully controlled MRI
hardware and image processing algorithms, full volume data sets were realized, with
three dimensional information about displacement available at voxel resolution [73, 75].
The initial successes of MRE as an elastographic modality were tempered by the

requirement of full volume data sets for successful reconstruction. The DIET concept
was conceived as a method of enabling volumetric stiffness reconstruction using only
surface motion data, hence avoiding the need for costly MRI equipment to obtain data
sets. The reduced input data set for DIET reconstruction leads to a challenging inverse
problem, where local minima are expected to play a significant role in the performance
of any reconstruction algorithm. The work presented in this chapter represents the first
studies in DIET, where reconstruction techniques and algorithms for MRE are modified
and applied to data sets comprising only surface motion. This study was intended as a
proof of concept for an elastographic imaging system based entirely on surface motion
data.
2.1 Data Simulation
The lack of previous work in DIET meant that no experimental apparatus existed for
measuring surface displacement data from soft phantoms or in vivo tissue. Thus, the
proof of concept study was undertaken with data simulated using the FE method. FE
techniques allow simulation of harmonic displacements from three dimensional models,
using standard mathematical techniques, outlined in Appendix A. Sampling from the
surface motions calculated by FE analysis provides the data required for a DIET re-
26 CHAPTER 2 PROOF OF CONCEPT STUDIES
construction. Gaussian random noise was added to the simulated surface data sets to
account for inaccuracies in any feasible experimental measurement system.
To provide a reasonable estimation of the geometry involved, a computer model
was developed of approximately the same size as a female breast. The absence of
commercial meshing software at this stage required that the model was meshed using
crude mesh generation code not originally intended for use with the geometry in ques-
tion. Computational limitations placed an upper limit on the number of nodes in the
mesh, based on the memory required to perform the necessary matrix inversions in
the FE code. This limitation meant that the initial breast model used was a quarter-
hemisphere, with radius 50 mm. Meshing this computer model with linear tetrahedral
elements gave a total of approximately 2,700 nodes, where the meshed model geometry
is shown in Figure 2.1.
Figure 2.1 The FE mesh used for computer model simulation in the three-region proof of
concept study.
An attempt was made to approximate the heterogeneous nature of a human breast

by dividing the basic model into two regions with different elastic modulus values. A
basic representation of breast physiology divides the volume into an inner region of
fibroglandular tissue, with a layer of fatty tissue closer to the skin. However, to give
a reasonable approximation to published values, the fibroglandular and fatty tissues
were assigned Young’s Modulus values of E = 35 kPa and E = 22 kPa respectively [54].
The outermost 20 mm of the base breast model volume was categorized as fatty tissue,
with the remainder having the higher fibroglandular modulus value. Poisson’s ratio, ν,
and density, ρ, were set at 0.49 and 1,000 kg/m3 respectively, approximating the near-
incompressibility of soft tissue due to its high water content [70, 93, 94]. Validation
of the compressible material model used for FE simulation was required to confirm
that simulations near ν = 0.5 did not introduce significant numerical inaccuracy. This
2.1 DATA SIMULATION 27
validation was performed against simulations using an incompressible material model,

with details included in Appendix A.
A range of high stiffness inclusions were modelled using the basic FE mesh, with
a total of 27 separate models representing a combination of three inclusion locations,
sizes, and stiffness contrasts to the surrounding tissue. Variation of inclusion sizes and
positions allowed the sensitivity of the reconstruction method to be investigated. All
inclusions were spherical, with centers 15, 25 and 35 mm from the breast surface. These
positions represent shallow, middle, and deep inclusion locations. Inclusion diameters
were 5, 10 and 15 mm, representing small, medium and large inclusions. Figure 2.2
shows the range of internal geometries used for computer simulation. The stiffness
contrast values chosen for the inclusion were relative to the fibroglandular model region,
and were chosen to represent a range of inclusion stiffness values around those obtained
in relevant clinical testing [54, 56]. Contrast values of 5x, 10x and 100x gave inclusion
modulus values of 174 kPa, 348 kPa and 3,480 kPa respectively.
Inclusion Depth Model Geometry Inclusion Diameter
5mm (small)
15mm (shallow)
10mm (medium)
25mm (mid)
35mm (deep) 15mm (large)
E = 35kPa E = 22kPa
Figure 2.2 The geometry used to generate simulated data sets in the three-region study.
The FE code used to simulate model motion was written in Fortran 77, and assumes
an undamped, linear elastic material model, in line with previous work in MRE [74].
Stiffness values within the mesh were assigned nodally based on the three dimensional
position of the node with respect to the stiffness distribution being simulated. The
boundary conditions applied during FE simulation were an approximation to a possible
clinical implementation of a DIET system, where harmonic vertical displacement was
applied over a small semi-circular area at the highest point of the model. An actuation
frequency of 100 Hz was chosen as it led to mechanical waves with a wavelength of
approximately 30 mm, allowing the motion to be clearly defined on the model surface.
The amplitude of actuation was 10 mm. The bottom face of the model had displacement
constrained to zero in all directions, representing a rough approximation of the chest
wall attachment. The quarter-hemisphere model used in the initial study required
additional boundary conditions to constrain the model in a symmetric manner, ensuring
realistic displacements. These additional boundary conditions were applied at the

two vertical internal faces of the model, and prescribed no in-plane forces or normal
displacement.
The output from FE simulation is a full vector of steady-state motion amplitudes
for each orthogonal motion component, x, y and z, at every node within the model.
The simulated motion observed at the surface of the mesh for a selection of the cases
simulated is shown in Figure 2.3. Here, the large amplitudes observed are directly pro-
portional to the amplitude of actuation for the linear elastic model assumed. Therefore,
clinically unrealistic displacements in the simulated data could be reduced with a re-
duction in the actuation amplitude. Thirty two randomly selected nodes on the outer
surface of the model were used as reference points, simulating the measurement of
motion data at points on the surface of the breast only. An assumed value for ex-
perimental measurement error was added to all reference point displacements during
motion processing. This noise was normally distributed about the mean surface dis-
placement amplitude, with a standard deviation of 5% of the average scale of motion.
The resulting simulated data set comprised 96 orthogonal displacement measurements
for each of the 27 scenarios simulated.
2.2 Three Region Reconstruction
The reduced number of known displacement values used as input to the reconstruction
algorithm requires a corresponding reduction in the number of reconstructed parame-
ters. Reconstruction of a modulus value at all 2,700 nodes within the model using only
96 known displacement values presents a highly ill-posed problem with little chance of
success, particularly given the noise added to the input data. The number of recon-
structed parameters for the inverse problem was reduced by dividing the FE model into
regions of constant modulus value. These regions were based on the internal geometry
configuration used for forward FE simulation, resulting in a total of three indepen-
dent stiffness values, one each for the fibroglandular, fatty, and inclusion regions of the
model.
Dividing the domain into fixed regions would be highly impractical in a real life ap-
plication due to the necessary a priori knowledge of internal tissue structure. However,
it was considered appropriate for the initial study, where the success of the method re-
gardless of significant assumptions was unknown. The large reduction in parameter
space for the inverse problem also reduced computational cost, allowing rapid simula-
tion at this early stage.
The inverse elastic reconstruction algorithm used in the three-region study was
based on the gradient descent method, and comprised Fortran 77 code developed from
algorithms already in use for MRE data [74, 94]. Here, the reconstruction aimed to
2.2 THREE REGION RECONSTRUCTION 29
0 Displacement Amplitude (mm) 29 0 Displacement Amplitude (mm) 141
(a) Shallow, 5 mm diameter, 5x contrast. (b) Middle, 10 mm diameter, 10x contrast.
0 Displacement Amplitude (mm) 30 0 Displacement Amplitude (mm) 38
(c) Deep, 5 mm diameter, 100x contrast. (d) Deep, 15 mm diameter, 100x contrast.
Figure 2.3 Surface displacement amplitudes for four of the inclusion scenarios simulated in
the three-region study.
minimize an error term representing the difference between the displacements simulated
at the target stiffness distribution and those simulated using the current values for the
three stiffness parameters. The initial Young’s Modulus guess for all three model regions
was 24 kPa, chosen as an approximation to the mean nodal stiffness value across the
entire model. Reconstructions on each simulated scenario were run to 300 iterations,
regardless of algorithm convergence. Each iteration required approximately 90 seconds
to complete on a Linux workstation with a 1.4 GHz CPU.
Figure 2.4 shows the convergence of the reconstruction algorithm for a selection of
simulated scenarios. Figures 2.4(a) and 2.4(d) indicate a failed reconstruction, where
the inclusion stiffness is reconstructed as softer than the surrounding regions. Fig-
ures 2.4(b)-(c) represent reconstructions considered successful, where the modulus value
for the inclusion is correctly identified as greater than the surrounding material. In ad-
dition to successful and unsuccessful reconstructions, cases where the fatty or glandular
modulus values are reconstructed at unrealistic values were considered inconclusive.
The lack of convergence observed in a large number of scenarios was not of significant
concern due to the unsophisticated gradient descent reconstruction algorithm used.
Figure 2.5 is a summary of the reconstruction result obtained in all 27 scenarios

tested in the three-region study. The success rate of the algorithm is nearly constant
across the full range of inclusion stiffness contrasts. Of particular significance is the
algorithm’s identification of a stiff inclusion in all the deep, low-contrast cases simulated.
These are situations that present a significant challenge to conventional screening and
diagnosis modalities. The algorithm has limited success with cases where the inclusion
is large, even though these would initially appear to be the easiest cases to reconstruct.
In addition, the algorithm was found to be very sensitive to the initial parameter
stiffness guess, and it appeared that in a number of the large inclusion cases that using
the same initial guess of 24 kPa was no longer suitable, as it led to convergence at an
incorrect solution representing a local error minimum.
The reconstructed results from the initial proof of concept study indicated that a
surface motion based reconstruction algorithm had the potential to successfully identify
a high stiffness inclusion given only a selection of surface displacement data. However,
significant issues were raised that required further investigation. Prevalent among these
concerns was the tendency of the inversion algorithm to converge to solutions that only
represented a local error minimum, and not the global optimum parameter values.
Issues with mesh resolution and accuracy were likely causes for the unrealistically large
simulated displacements observed, most probably as a result of resonance within the
model. These issues indicated that a further study would be of benefit, using a refined
FE mesh, and a more realistic approach to the reconstruction of internal stiffness that
did not rely on prior knowledge of the internal model geometry.
2.2 THREE REGION RECONSTRUCTION 31
28 200
Fatty
Fibroglandular
Young’s Modulus, E (kPa)
21

150 Inclusion
14 100
7 Fatty 50
Fibroglandular
Inclusion
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Iteration Iteration
(a) Shallow, 5 mm diameter, 5x contrast. (b) Middle, 10 mm diameter, 10x con-

trast.
29
1000
Fatty
Fibroglandular
22
750 Inclusion
500 15
250 7 Fatty
Fibroglandular
Inclusion
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Iteration Iteration
(c) Deep, 5 mm diameter, 100x contrast. (d) Deep, 15 mm diameter, 100x con-
trast.
Figure 2.4 Convergence of the three-region reconstruction algorithm for a selection of inclu-
sion scenarios, where the target modulus values for the fibroglandular and fatty regions were
E = 35 kPa and E = 22 kPa respectively.
Inclusion Diameter and Stiffness Contrast

5mm Diameter 10mm Diameter 15mm Diameter
5x 10x 100x 5x 10x 100x 5x 10x 100x
Shallow
ûüüüüüû ? ?
Middle
üüüüüü ? û û
Deep
üûüüüüüû û
Figure 2.5 A summary of the reconstruction results from the three-region study, showing
successful, unsuccessful and inconclusive reconstructions.
2.3 Dual-Mesh Reconstruction
A significant limitation of the initial three-parameter reconstruction study was the

required a priori knowledge of the model’s internal structure when formulating the
inverse problem. Because the three parameters described in Section 2.2 were based
on known model physiology, the reconstruction algorithm had no way of indicating an
inclusion in a more generalized case, where internal structure was not known. As this
general case is the intended target of the DIET system, a second phase of work was
undertaken that aimed to allow a more sophisticated stiffness reconstruction through
the use of an increased number of reconstructed parameters. This additional study also
allowed issues concerning mesh resolution and accuracy to be addressed.
The availability of a commercial FE meshing software package and an increased

computational capacity at the time of this study allowed creation of a significantly
refined FE mesh, shown in Figure 2.6. The mesh contained approximately 8,700 nodes,
with the more advanced meshing software able to produce significantly smoother exter-
nal surfaces. Here, the external model geometry remained the same as that described
in Section 2.2.
Figure 2.6 The fine mesh used for FE displacement simulation during dual-mesh reconstruc-
tions.
A single inclusion scenario was modelled for this experiment, where a 10 mm di-
ameter spherical inclusion was located 30 mm from the surface of the model. The
target material properties for the ‘healthy’ regions were the same as the initial study,
with an inner fibroglandular stiffness of E = 35 kPa, and an outer fatty region where
E = 22 kPa. The inclusion stiffness was set at E = 175 kPa, representing a stiffness
contrast of approximately seven times the surrounding material.
2.3 DUAL-MESH RECONSTRUCTION 33
Boundary conditions applied to the mesh were the same as those described for
the three-region study, with actuation amplitude limited to 1 mm in an attempt to
avoid the unrealistically large surface displacements observed with the previous mesh.
Surface displacement amplitudes resulting from FE simulation of the model at 100 Hz
are shown in Figure 2.7. Here, the symmetric appearance of the motion field indicates
the mesh is sufficiently refined to avoid significant numerical errors.
0 Displacement Amplitude (mm) 1
E = 22kPa
E = 35kPa
E = 175kPa
(a) The target stiffness distribution along a plane (b) Simulated surface displacement at
passing through the center of the inclusion. the target stiffness distribution when
actuated at 100 Hz.
Figure 2.7 The stiffness distribution and surface motion from the target model in the dual-
mesh study.
Increasing the number of reconstructed parameters for this study required a cor-
responding increase in the number of surface displacement values used in the inverse
problem in order to maintain the posedness of the reconstruction. Promising advances
in the concurrent development of the DIET motion tracking system indicated that a
high point resolution on the breast surface was possible [95], allowing all 980 nodes
on the exterior surface of the mesh to be considered as reference points for motion
measurement. Normally distributed noise was again applied to the simulated motion
values, with a standard deviation of 10% of the mean surface motion amplitude value.
The desire for an increased number of parameters in the reconstruction problem

led to the development of a dual-mesh scheme for reconstruction of the model stiffness
distribution. Here, the displacements were discretized on a fine mesh to ensure accurate
capture of the model physics, while the elastic property distribution was supported
on a coarse mesh, which allowed for gradual stiffness variation while maintaining a
reasonable match between the number of measurements and the number of unknowns.
The conversion from a stiffness distribution supported on this coarse mesh to the fine
mesh was performed by interpolation, where the value of the stiffness at each fine node
was the linear interpolation of the coarse node stiffness values surrounding it. This
interpolation method required that the coarse mesh elements completely enclosed all
fine mesh nodes to prevent numerical errors.
A two dimensional example of a dual mesh scheme is shown in Figure 2.8, where
the material property distribution is interpolated from the coarse (black) nodes to the
fine (red) mesh used for FE simulation. Using a coarse mesh scheme gave the inverse
problem the flexibility to reconstruct a range of internal stiffness distributions. Here,
a variable number of coarse mesh points can increase or decrease the spatial resolution
of the reconstructed solution.
Figure 2.8 A two-dimensional representation of a dual mesh scheme, where displacements

are calculated using the fine mesh (red), and material properties are described at the coarse
nodes (black).
Reconstructions were performed using a range of coarse meshes varying from 20

nodes through to over 800. The gradient, or Jacobian, matrix calculation routine within
the gradient descent algorithm was modified to account for the added interpolation
required for the dual mesh scheme. Figure 2.9 shows the best results obtained from
the dual mesh reconstruction study. Results from the vast majority of reconstructions
failed to correctly localize the hard inclusion, although many were able to detect the
presence of a high stiffness region. These results indicated that local minima were
severely hampering the reconstruction performance as the number of reconstructed
parameters was increased.
2.4 Summary
The first proof of concept studies in DIET applied modified theory and computational
algorithms from MRE to the surface motion based reconstruction problem. Initial
results from the three region simulation and reconstruction study indicated that re-
construction of a simplified internal modulus distribution was possible using the DIET
2.4 SUMMARY 35
10 Young's Modulus, E (kPa) 210 10 Young's Modulus, E (kPa) 122
(a) Coarse material property mesh with (b) Coarse material property mesh with
20 nodes. 839 nodes.
Figure 2.9 The best reconstructed stiffness distributions achieved during the dual-mesh re-
construction study, where the image plane passes through the center of the target inclusion
position.
method. Limitations in the accuracy and flexibility of both the forward FE model
and the reconstructive approach led to a dual mesh study with an increased number
of reconstructed parameters. Although the dual mesh study also contained promising
results, the majority of reconstruction attempts failed to reconstruct the correct target
modulus distribution. Analysis of the reconstruction algorithm code confirmed it was
performing as intended, indicating that the limitations of the approach had more to do
with the forward FE model, and the relationship between internal stiffness distribution
and observed surface displacements.
Difficulties in analyzing the limited success of the reconstruction algorithms led to
a decision to shift focus to soft phantom based studies. Such studies would allow more
thorough evaluate the overall DIET system in a controlled experimental environment.
Concurrent research had reached a point where an experimental actuation and motion
system was ready for testing [96]. This research provided the hardware and software
required to extend the evaluation of the DIET concept into experimental phantom
studies.
Chapter 3
Experimental Methods
An important step in the development of any soft tissue imaging modality is a testing
phase using tissue approximating phantoms. In such testing, experimental data col-
lection and processing can occur in a controlled environment, allowing detailed system
analysis prior to attempting ex vivo or in vivo testing. The first experimental studies
using the DIET system were performed on data obtained from a series of cylindrical
phantoms.
This chapter describes the methodology behind the selection of suitable phantom
materials, and the creation of the phantoms. Also detailed is the experimental system
developed for phantom actuation and motion capture, including the algorithms that
allow comparison of measured data with FE-simulated displacement fields. Develop-
ments in the major facets of the experimental method are presented chronologically,
detailing the improvements in experimental technique over the collection of a number
of experimental data sets.
3.1 Phantom Preparation
Previous soft tissue imaging research has been supported by the use of tissue approxi-
mating phantoms created from a range of both natural and artificial materials, includ-
ing gelatine and agar based gels, and polymer materials [53, 97]. Experimental data
collection for the DIET system utilized two distinct classes of phantom material. The
first experimental efforts used gelatine based phantoms. An investigation of alternative
materials led to soft silicone gels being used for subsequent experiments.
A cylindrical phantom geometry was chosen to provide a simple shape for mold
manufacture using commonly available hardware components. In addition, when actu-
ated along the major axis, the cylindrical shape was expected to provide a symmetric
mechanical deformation pattern on the phantom surface. This symmetry would allow
38 CHAPTER 3 EXPERIMENTAL METHODS
for a simplified camera system that had been developed to capture surface deformation
from only a small area of the phantom surface.
3.1.1 Gelatine Phantoms
Gelatine is a translucent solid manufactured primarily from collagen, a connective tissue

found in animals. Raw materials for gelatine manufacture are sourced primarily from
animal processing plants, and include skin, cartilage and bone [98]. After pre-treatment
to sanitize and stabilize the raw material, gelatine is extracted in a multi-stage process
using hot water. Filters and separators in a purification stage separate gelatine from
any fat and other fibres remaining in the mixture. The resulting material is a soft gel
containing hydrolyzed collagen. Drying, grinding, sifting and blending this gel produces
small granules suitable for transport and storage, shown in Figure 3.1.
Figure 3.1 Granules of processed gelatine, a raw ingredient in the first experimental phan-
toms.
A gelatine-based hydrogel was chosen for the initial phantom experiments because
of its linear elastic behavior and similar mechanical properties to soft human tissue [97].
In addition, it is commonly used in testing medical imaging modalities [76, 97, 99]. The
major steps in the manufacturing process for the gelatine phantoms used in this research
are outlined below:
1. Distilled water was heated to 70 ◦ C and mixed with 37% 2-Propanol. The addition
of Propanol altered the sound speed in the finished phantom. While this property
was not imperative for the DIET system, as the testing did not involve ultrasound,
the addition of Propanol ensured the compression wave speed was similar to that
of soft tissue.
3.1 PHANTOM PREPARATION 39
2. A small amount of titanium dioxide (TiO2 ) powder was added to the hot liquid
mixture. This gave the final phantom a white color, ensuring the surface motion
was easily measured during actuation.
3. Granulated 20A grade gelatine was slowly added to the water mixture until com-
pletely dissolved. The mixture was stirred for at least ten minutes to ensure
complete dissolution of the gelatine granules. The final stiffness of the phantom
was determined by the amount of gelatine added at this stage.
4. After waiting for the temperature of the mixture to decrease to approximately

40 ◦ C, a small amount of Formaldehyde was added to initiate the cross-linking
process. Continued stirring ensured the Formaldehyde was evenly distributed
throughout the mixture.
5. The warm liquid was poured from the mixing vessel into a mold, and left for 24
hours at room temperature, to allow the cross-linking process to complete and
form a congealed solid. After setting, the resulting phantom was removed from
the mold and stored in an airtight container, to prolong its shelf life.
Experiments with gelatine composition and manufacturing methods resulted in

two different types of gelatine being used to create phantoms, with ingredients listed
in Table 3.1. A ‘soft’ gelatine was created using 20 g of gelatine granules, with 40 g of
raw gelatine used for a second, ‘hard’, gelatine material.
Table 3.1 The ingredients used to make to make the two types of gelatine used for phantom
experiments.
Identifier
Ingredient Soft Hard
Distilled Water 300 ml 300 ml
2-Propanol 25 ml 25 ml
TiO2 0.1 g 0.1 g
Gelatine 20 g 40 g
Formaldehyde 3 ml 3 ml
The high water content of the gelatine material made the phantoms highly sus-
ceptible to the environment in which they were left to set. Qualitative examination of
gelatine phantoms created using a range of mold designs indicated that exposure to air
when setting had a negative impact on the resulting phantom. Areas of the gelatine
exposed during setting were found to have partially dehydrated and hardened after
exposure to a room temperature environment. This sensitivity required careful mold
design to ensure minimal exposure to the air while the gelatine set.
The final mold design used for gelatine phantom preparation is shown in Figure 3.2.
The mold’s circular top and bottom (1) were cut from 15 mm thick plastic sheet. A
thin sheet of clear acetate film (2) wrapped around the end caps formed the cylindrical
mold wall. Using clear sheet allowed the inside of the mold to be visible when filling
with liquid gelatine. Fastening the acetate sheet to the end caps with double sided
tape created a watertight seal that prevented gelatine leakage.
Figure 3.2 The cylindrical mold developed for gelatine phantom preparation, with end caps
(1), clear sidewalls (2), and pouring ‘chimney’ (3) indicated.
The top of the mold was extended using a second plastic disc (3), with a small
circular hole drilled through both layers. This ‘chimney’ functioned as an opening
through which the liquid gelatine mixture was poured into the mold. In addition,
filling the mold to the top face indicated by (3) prevented gelatine shrinkage during
the setting process from leaving a depression in the top of the final phantom. All mold
components were given a thin coating of a mold release spray before assembly to assist
the phantom de-molding process.
After setting for twenty-four hours, the mold was carefully removed from around
the phantom. Any gelatine in the chimney was removed using a scalpel. Two of
the many test phantoms created during the mold development process are shown in
Figure 3.3.
Two homogeneous phantoms created from gelatine were used for experimental
motion capture. The first phantom was made using the soft gelatine ingredients listed
in Table 3.1. After results had been obtained from this phantom, and experimental
techniques advanced, a subsequent phantom was created using the hard gelatine recipe.
Figure 3.3 Two gelatine phantoms created during the mold development process.
Both phantoms had a number of reference points added to their exterior surface using
an oil-based permanent marker prior to experimental motion capture.
Several significant limitations with the gelatine material were observed during man-
ufacturing and qualitative evaluation of the gelatine phantoms. The use of highly toxic
Formaldehyde to initiate cross-linking within the gelatine mixture required that parts
of the manufacturing process were performed within a fume hood, while wearing protec-
tive equipment. In addition, to avoid any potential health risk, the gelatine phantoms
were only handled while wearing latex gloves.
Phantom desiccation was observed to be a significant problem after de-molding,
where the majority of the phantom surface was exposed to air. Dehydration of the
phantom surface had the potential to alter the material properties of the gelatine,
with the exposed surface hardening as the moisture content decreased. In addition
to a complicated, airtight mold, phantom dehydration limited mechanical testing and
surface motion capture to a time window of between one and two days after phantom
manufacture.
The high temperature required for manufacture posed difficulties when considering
heterogeneous gelatine phantoms. When the DIET system was sufficiently advanced,
the natural progression from simple homogeneous phantoms would be to heterogeneous
models with distinct hard and soft regions. However, tests performed indicated that
any existing, set, gelatine in a mold began to melt when it came into contact with
additional warm liquid gelatine, making it difficult to determine the appearance and
location of the interface position between the hard and soft regions.
3.1.2 Silicone Phantoms
Silicone is a form of synthetic elastomer made from a cross-linked polymer reinforced

with silica, and available in many forms, from liquid to solid [100]. Room Tempera-
ture Vulcanizing (RTV) soft silicone gel was identified as a suitable phantom material,
having several advantages when compared with the gelatine previously used:
• Silicone gels were available in food grade standards, indicating many types were
non-toxic, and completely safe to handle.
• The silicone manufacturing process was extremely simple and rapid, requiring
only a mixture of two constituent parts.
• The manufacturing process could be performed entirely at room temperature.
• Achieving a range of phantom colors was possible using intrinsic pigments.
• The cured silicone gel was stable in air, resilient, and had an indefinite shelf
life, allowing phantoms to be repeatedly tested over a period of time with no
significant changes in material properties.
• The availability of silicone gels with a range of hardness values meant that creat-
ing heterogeneous phantoms from both hard and soft silicone would not present
significant difficulties.
• The raw materials for silicone manufacture were affordable, and could be pur-
chased in the small quantities required.
The silicone gel used for phantom manufacture was sourced from an American
supplier to the prosthetic and special effects industry (Factor II Inc.). Dialogue with
sales consultants identified two types of silicone likely to be suitable for approximating
both soft and hard tissue, designated A-341 and LSR-05 respectively. Both products
were Platinum catalyzed RTV-2 silicone gels, indicating a two part mixture that would
cure in total confinement if required. Material Safety Data Sheets (MSDS) for both
A-341 and LSR-05 silicones are included as Appendix B. The manufacturing process
for both soft and hard silicone materials involved the same major steps:
1. Silicone components Part A and B were weighed according to manufacture-

specified ratios into a plastic mixing container.
2. A small amount of intrinsic pigment was added to give the desired final phantom
color.
3. Ingredients were thoroughly mixed by hand to ensure uniform distribution of

silicone components and pigment.
4. The mixing container was placed into a small venturi vacuum chamber, shown
in Figure 3.4, where a light vacuum was applied for 10–15 minutes. This stage
removed any entrained air bubbles from the mixture.
5. The final silicone mixture was poured into a mold, and left to set at room tem-
perature for 6–12 hours.
1
3
Figure 3.4 The venturi vacuum setup used to remove air from the silicone mixture, where
the water supply (1), venturi (2), and chamber (3) are indicated.
Experimentation with both A-341 and LSR-05 indicated that minor variations in
the ratio of Part A to Part B for each type of silicone had no significant effect on
the stiffness of the resulting material. Hence, silicone stiffness was controlled using
a combination of A-341 silicone and LSR-05 silicone, where the final stiffness was
determined by the ratio of the two types. Combining the two silicones did not have
an adverse effect on the resulting material as both types were chemically similar. The
weight percentages of each type of silicone used to create the final ‘soft’ and ‘hard’
silicone materials are listed in Table 3.2.
The straightforward silicone manufacturing process meant that the cylindrical mold
developed for gelatine phantoms in Section 3.1.1 could be simplified for silicone phan-
tom manufacture. This final silicone mold developed is shown in Figure 3.5. The
vertical sides of the mold (1) were created from 80 mm nominal diameter PVC pipe,
providing an accurate circular cross-section and straight sides. This pipe section was
split longitudinally into two pieces, allowing easy disassembly. A shallow groove ma-
chined around the inside of the pipe indicated a phantom height of 75 mm.
Table 3.2 The weight percentage composition of the two types of silicone used for phantom
preparation.
Identifier
Ingredient Soft Hard
A-341 Part A 90% 36%
A-341 Part B 10% 4%
LSR-05 Part A - 30%
LSR-05 Part B - 30%
Figure 3.5 The mold used to cast silicone phantoms, with mold wall (1), clamping screw (2),
and fastening glue (3) indicated.
To ensure the mold was watertight, a pipe clamp (2) was used to secure the sides
of the mold together, and sticky tape was used to seal the internal seams. Hot glue
was used liberally (3) to attach the pipe to a clean, flat plastic base plate. The wa-
tertightness of the mold was checked by filling with water prior to silicone casting. A
lubricating silicone spray was applied to the interior surface of the mold before intro-
ducing the liquid silicone to assist the de-molding process. After curing, the phantoms
were removed by disassembling the mold and gently prising the silicone material free.
The phantoms were then stored at room temperature in a box to prevent accidental
damage.
A total of five cylindrical phantoms were created using the soft and hard silicone
materials identified in Table 3.2. The first two phantoms were homogeneous, with
3.2 MATERIAL TESTING 45
one made entirely from soft silicone, and another from hard silicone. A further two
heterogeneous phantoms were subsequently created for more advanced DIET system
testing. A ‘stacked’ phantom was created with a layer of hard silicone beneath a layer of
soft silicone, while a ‘concentric’ phantom contained a cylindrical core of hard silicone
surrounded by soft silicone. The first four silicone phantoms created for experimental
motion capture are shown in Figure 3.6. The reference points visible on the surface of
these phantoms were created from small pieces of paper and glitter. The final silicone
phantom created comprised a hard sphere within a soft cylinder, and is discussed in
more detail in Chapter 5.
3.2 Material Testing
Independent knowledge of material properties is important for verification of an elas-

tographic imaging system. In the experimental work with the DIET system, measured
mechanical property values provided approximate material property values that were
used to validate reconstructed results. This section describes the static and dynamic
mechanical testing attempted on both gelatine and silicone phantom materials. All me-
chanical testing was performed using a Dynamic Mechanical Analysis (DMA) machine
(Pyris Diamond DMA, Perkin Elmer Instruments), shown in Figure 3.7. The highly
pliable nature of the phantom materials ruled out tensile material testing, meaning
compressive testing was used for all mechanical measurements.
Mechanical testing samples were created from small batches of gelatine and silicone
produced using the same method as the materials used for phantom manufacture. All
samples were tested at room temperature approximately 24–48 hours after preparation.
Gelatine samples were prepared using a sharp hollow cylindrical cutter of approximately
10 mm diameter to take cylindrical core samples from a block of gelatine. The extracted
gelatine cylinder was then cut into 3–5 mm lengths with parallel ends using a razor
blade, producing the small compression samples required. This technique ensured
that the gelatine samples did not come from material exposed to the air when setting,
avoiding the issue of surface dehydration described in Section 3.1.1. A mold for silicone
mechanical testing samples was created by drilling 10 mm diameter holes in a 6 mm
thick plastic sheet. Sealing the bottom of each hole and casting small amounts of
liquid silicone created small, highly regular, cylindrical compression samples.
Material density estimates were made by weighing the homogeneous phantoms us-
ing digital scales. Measurements of the gelatine phantoms indicated a material density
between 1000–1100 kg/m3 . The measured density of the soft and hard silicone ma-
terials was ∼975 kg/m3 and ∼1045 kg/m3 respectively. All density calculations were
performed assuming the phantoms had perfectly cylindrical geometry.
(a) ‘Soft’ homogeneous phantom. (b) ‘Hard’ homogeneous phantom.
(c) ‘Stacked’ heterogeneous phantom. (d) ‘Concentric’ heterogeneous phantom.
Figure 3.6 The four silicone phantoms used for experimental motion capture.
3.2 MATERIAL TESTING 47
Figure 3.7 The Pyris Diamond DMA machine used for mechanical testing of phantom ma-
terials.
Static compressive tests were performed at room temperature on a range of samples

from both gelatine and silicone materials to a strain of approximately 0.1. A low strain
rate of 100 µm/min was used to approximate a static loading scenario. Stress-strain
plots from static testing of both gelatine and silicone samples are shown in Figure 3.8.
The stress-strain relationship observed for both gelatine and silicone materials in-
dicated an approximately linear elastic material behavior over the strain range tested.
The apparent non-linearity at low strain for the hard silicone sample in Figure 3.8(b)
is the result of incomplete contact between the upper testing platen and the top of
the sample, meaning it was not completely constrained until a strain of approximately
0.02. The two small discontinuities in the hard silicone loading path are due to small
amounts of slip, where the sample end faces moved slightly against the testing platens.
Approximate Young’s Modulus values, E, were calculated for all samples using Hooke’s
Law, where E is the gradient of the stress-strain plots in Figure 3.8:
σ
E= . (3.1)
²
5
Hard Gelatine
Soft Gelatine
4
Stress, σ (kPa)
0
0 0.02 0.04 0.06 0.08 0.1
Strain, ²
(a) Gelatine stress-strain plots.
12
Hard Silicone
Soft Silicone
8
Stress, σ (kPa)
0
0 0.02 0.04 0.06 0.08 0.1
Strain, ²
(b) Silicone stress-strain plots.
Figure 3.8 Static testing results from the gelatine and silicone materials used for phantom
manufacture.
3.3 ACTUATION AND MOTION CAPTURE 49
The calculated Young’s Modulus values for the four phantom materials tested are given
in Table 3.3.
Table 3.3 Young’s Modulus values obtained from static testing of phantom materials.
Material E
Soft Gelatine 9 kPa
Hard Gelatine 26 kPa
Soft Silicone 32 kPa
Hard Silicone 135 kPa
During static testing it was observed that there was little or no slip between the
material sample and the testing platens. This type of boundary condition gives a
non-uniform strain distribution within the sample, meaning Young’s Modulus values
calculated using Equation 3.1 may be overestimated. A stiffness correction factor can
be calculated via FE simulation of the static testing scenario both with and without
a no-slip condition at the boundaries. These additional simulations indicated that the
measured static modulus was overestimated by up to 10%, depending on the material’s
Poisson’s ratio.
Assessing soft material viscoelastic parameters at relatively high frequencies is a
challenging task, meaning ex vivo and in vivo tissue testing has focused on static
modulus values as a point of comparison between healthy and cancerous tissue [58, 73].
Because the modulus values of the phantom materials tested were at the lower limit of
the DMA’s capabilities, the results of dynamic material testing were highly variable,
and considered unreliable.
3.3 Actuation and Motion Capture
Experimental motion capture from soft phantoms requires the development of a phan-
tom actuation and motion capture system. The individual components within the
experimental DIET system changed as it was developed. However, the basic functional
elements remained the same throughout, and are shown in Figure 3.9. An example
of the real hardware used in the system is shown in Figure 3.10. What follows is
a generalized description of the experimental system used for actuation and motion
capture.
The experimental actuation and image capture system performed two major func-
tions, each controlled by separate computers. A ‘control’ computer was used for actu-
ator control, motion feedback, and lighting tasks, while ‘imaging’ computers were used
for motion capture from each of the cameras. A dSPACETM modular hardware control
Blackout Curtain
Camera 2
Strobe Phantom
Camera 1
Actuator
Control Position
Trigger Trigger Image Trigger Image
Signal Signal
dSPACE Hardware Imaging Imaging

Computer Computer
Control Computer 1 2
Figure 3.9 The experimental setup for phantom motion capture.
Figure 3.10 A photo of the typical setup for phantom displacement imaging.
system (dSPACE GmbH, Germany) attached to the control computer allowed accurate
control of the actuator and stroboscope [96]. This system comprised a MATLABTM
and SimulinkTM model incorporating specific dSPACE hardware control blocks. Once
compiled, the control system code was downloaded from the control computer to the
dSPACE hardware. An analog output port on the dSPACE unit provided a trigger
signal to a stroboscope used for scene illumination. The dSPACE unit also provided
frequency and amplitude control signals for the actuator. These signals were amplified
to drive the electromagnets within the voice coil actuator. A velocity signal returned
from the actuator provided the feedback path to complete a closed loop actuator con-
trol system. Control software and hardware was run in real time, allowing visualization
of all input and output signals using the dSPACE Control Desk environment shown in
Figure 3.11.
2
2
Figure 3.11 The dSPACE Control Desk software interface, where actuation parameters and
strobe phase are set manually (1), and the displacement response of the actuator is displayed
for the user (2).
Proprietary software limitations preventing simultaneous image capture required

two separate computer workstations to control the image capture cameras. Each imag-
ing computer was connected to a digital camera using a USB interface, and ran remote
image capture software provided by the camera manufacturer. The image capture
process for each camera was manually triggered using the software interface on the
respective host computer.
All experimentally tested phantoms were placed on a flat circular surface on top
of the actuator, which was configured to provide vertical sinusoidal motion. Both
imaging cameras were set up with a field of view including the all reference points on
the phantom surface. The stroboscope was situated to give clear and even illumination
of the reference points from the perspective of both cameras. As any ambient light
would upset the imaging process, the actuator, strobe and cameras were covered with
a blackout curtain for the duration of the imaging process.
Figure 3.12 is a schematic showing how controlled strobe lighting was used to
capture the full range of phantom motion, despite the high speed of actuation relative
to the longer shutter speed and exposure time required to generate each digital image.
Each row of the figure demonstrates the capture of an individual frame, with time the
horizontal co-ordinate. The sinusoidal line represents the steady state oscillation of a
reference point on the phantom surface.
The time between strobe illumination was constant, and matched the period of the
actuator. This synchronization meant that at each strobe offset, the reference point
appeared at a single, static location, allowing camera exposures longer than one motion
cycle. After both cameras had downloaded their respective images, the phase offset
between the strobe and actuation signals was adjusted. In the actual experimental
motion capture, the phase increment was in the range 10–20◦ .
After shifting the strobe phase, the phantom was illuminated at a different position
in its motion cycle. This process allowed the motion of any reference point to be
tracked across the full range of its steady-state motion. Experimental motion capture
was typically performed until at least one complete motion cycle had been captured.
A total of seven motion data sets were obtained from the experimental setup during
its development. These data sets fell into two main stages, based on the components
used in the system at the time. The first motion data sets were obtained from the ho-
mogeneous gelatine phantoms described in Section 3.1.1. An improved motion capture
system was used to measure surface motions on all of the silicone phantoms described
in Section 3.1.2. Both systems use the same fundamental hardware and software ap-
proach.
3.3.1 Motion Capture from Gelatine Phantoms
The first attempt at experimental motion capture used a soft homogeneous gelatine
phantom that was 70 mm in diameter and 50 mm in height. The actuation system used
for this experiment comprised a linear voice coil actuator (VP5M, Derritron Electronics
Ltd) driven by a 300 W amplifier (DCP3, Derritron). Actuation was at a frequency
of 50 Hz, with an amplitude of 0.6 mm. Two consumer-level digital cameras (Olympus
C-4040 Zoom and Canon PowerShot G6) were used for image acquisition. A Dawe
Transistor Stroboflash (Type 1209C) provided the illumination required for motion
capture. Actuation plate motion was measured using a Helium-Neon laser and an
interferometer (Model OFV-512, Polytec) coupled to a vibrometer controller (Model
Time
Strobe Harmonic
Camera Exposure Captured
Phase Point Strobing Image
Offset Motion
0±
45 ±
±
90
±
135
180 ±
225 ±
±
270
±
315
±
360
Figure 3.12 A schematic detailing the timing of the image capture process, where in this
example the strobe offset between consecutive frames is 45◦ .
OFV-5000, Polytec). The actuation plate velocity values provided by the interferometer
were converted into a real-time position within the dSPACE control system.
Two columns of black dots on the surface of the soft gelatine phantom formed the
fourteen reference points that were motion tracked with the camera system. Simultane-
ous left and right images were captured at 20◦ phase offsets throughout the sinusoidal
motion cycle. The process was repeated until two complete cycles of motion had been
captured. The still frames captured by the right camera across one steady state motion
cycle are shown in Figure 3.13.
The second round of gelatine phantom imaging was performed using the homoge-
neous phantom created using the hard gelatine ingredients in Table 3.1. This phantom
had a diameter of 75 mm and a height of 70 mm. A single column of fifteen black dots
formed the surface reference point locations for motion tracking. Actuation was at a
frequency of 50 Hz, with an amplitude of 0.975 mm. The actuation and control system
used was identical to the previous soft gelatine phantom testing. However, both cam-
eras used for displacement imaging were Canon PowerShot G5 models, allowing a more
consistent approach to camera setup and image capture. The phase offset between se-
quential still images was 18◦ , giving a total of 20 images across one steady-state motion
cycle.
3.3.2 Motion Capture from Silicone Phantoms
The third stage of experimental data collection was from the silicone phantoms de-
scribed in Section 3.1.2, and using a new actuation system designed specifically for the
DIET phantom testing [101]. This actuator incorporated a Linear Velocity Displace-
ment Transducer (LVDT) that simplified the actuation system by circumventing the
requirement for the laser interferometer system used in prior tests. Initial testing indi-
cated the new actuator was accurately controllable to within 0.05 mm. Both cameras
used for motion capture were the same Canon Powershot G5 models used for the hard
gelatine phantom experiments.
Prior to motion imaging, all silicone phantoms had a single column of fifteen dots
applied to the surface, such that they were equally visible to both cameras. The color
of these reference markers was chosen to give the greatest possible contrast with the
underlying silicone material. Thus, black markers were used for the homogeneous
phantoms, and white markers for the heterogeneous phantoms. All silicone phantoms
were actuated at 100 Hz, with an amplitude of 0.5 mm. A total of twenty frames were
captured across one steady-state motion cycle from each phantom.
Figure 3.13 Motion frames from the right camera during actuation of the soft gelatine phan-
tom.
3.4 Photogrammetry
Experimental motion capture results in sequential still images of all surface reference
points across the complete phantom motion cycle. Because the DIET reconstruction
approach is based on three dimensional surface motion information, a conversion of
reference point position from the two dimensional image co-ordinate system into a
three dimensional, ‘world space’, co-ordinate system was required. The mathematics
used for this conversion are beyond the scope of this work, but are covered in detail in
a number of texts, including those by Klette, and Trucco [102, 103].
Calibration of the camera system was performed prior to actuation and motion
capture, using a specifically developed calibration object [104]. Images of the calibration
cube from the right and left cameras are shown in Figure 3.14. Between calibration and
phantom imaging, it is imperative that the camera setup is not altered. The output
from the calibration process was a set of camera specific parameters that were used
to calculate a three dimensional position in world space for every reference point over
every frame i, termed P̄i .
(a) Left camera. (b) Right camera.
Figure 3.14 Left and right camera views of the object used for calibration of the imaging
system.
Figure 3.15 demonstrates the result of the conversion process, where left and right
images of a reference point motion track are converted into a single three dimensional
point path. The upper panels in this figure are based on experimental data from
a silicone phantom, where the ellipsoidal shape of each point location is due to the
angle between the reference point and the camera lens. The output from experimental
motion capture was a vector for each reference point, P̄, describing its three dimensional
location across all frames.
3.4 PHOTOGRAMMETRY 57
y y
x x
(a) Two-dimensional reference point posi- (b) Two-dimensional reference point posi-
tions as viewed from the left camera. tions as viewed from the right camera.
x y
(c) Calculated reference point positions in world

space, with the dimensions of the point path bound-
ing box shown.
Figure 3.15 An example of reference point position conversion from two dimensional image
co-ordinates to the three dimensional world space coordinate system.
3.5 Displacement Data Correlation
A fundamental principle of the DIET system is the observed correlation between a

measured set of displacements and a displacement field simulated using FE analysis. To
calculate an error metric value quantifying this displacement correlation, a relationship
between the reference point locations and the FE mesh used for displacement simulation
is required. The range of solutions developed to provide this relationship evolved
alongside developments in the underlying modelling strategy, and are described in this
section.
3.5.1 Mesh Alignment
During the initial experimental work, measured surface motions were compared to FE-
simulated displacements at surface nodes on the FE mesh. This approach required that
the FE mesh had nodal positions matching the world space locations of the reference
points used for experimental motion capture. The advantage of the mesh alignment
approach was that the measured motion data could be directly compared with the
output from FE simulation, providing a rapid method for evaluating the correlation
between the two. However, mesh matching required a separate FE mesh, with custom
surface node locations, for every experimental data set. An enlargement of a FE mesh
created using this technique is shown in Figure 3.16, where the fourteen reference points
from the soft gelatine phantom (red) all correspond to unique nodes on the surface of
the mesh.
Figure 3.16 An example of the mesh alignment technique, where surface nodes are manip-
ulated so their position corresponds to the dots located at the reference point locations (red).
3.5 DISPLACEMENT DATA CORRELATION 59
3.5.2 Surface Projection
A point projection and interpolation scheme was developed to allow an accurate and
flexible approach to relating reference point locations to FE mesh geometry. FE meshes
were, in general, created using the same dimensions as the phantom being studied.
However, it was not possible to obtain an exact match between the surface of a FE
mesh and the surface of the phantom. In particular, the phantom size and geometry
could not be exactly determined due to small measurement errors. Additionally, ge-
ometry discretization into linear tetrahedral mesh elements introduces small errors by
representing curved model surfaces as small straight lines, as illustrated in Figure 3.17.
Figure 3.17 An example of the error introduced when a curved surface (grey) is discretized
with linear elements (black), where the error is particularly pronounced for curves of small
radius.
To allow multiple phantoms with the same external geometry to be represented

with a single mesh, a point projection algorithm was developed in MATLAB. Inputs to
this algorithm were a node location index, and element incidence lists defining both the
tetrahedral mesh elements, and the triangular surface elements on the exterior surface
of the mesh. The experimentally measured motion data calculated in Section 3.4 was
loaded from a text file. Figure 3.18 is a pictorial representation of the major steps in
the surface projection algorithm, where a reference point in three dimensional space
(red) is projected to the surface of a FE mesh. Additionally, Figure 3.19 introduces
the co-ordinate system and terminology used within the algorithm.
3.5.2.1 Reference Point Centroid
Surface projection of a reference point position involves first identifying the centroid
of the reference point’s motion path, defined P∆ . The extent of the motion path in
this work was determined by enclosing the measured point position over all frames, P̄,
(a) Elevated reference point location. (b) Surface element normal vectors.
(c) Surface node normal vectors. (d) Projected surface element.
(e) Reference point projection. (f) Projected reference point location.
Figure 3.18 A graphical representation of the steps in the surface projection process.
3.5 DISPLACEMENT DATA CORRELATION 61
(x 20,y 20,z 20)

(x 30,y 30,z 30)
P∆ = (Px∆,Py∆,Pz∆ )
A0
n 2n
n 3n
(x 10,y 10,z 10)

(x 2,y 2,z 2)
(x 3,y 3,z 3)
ne n 1n
A
y x
(x 1,y 1,z 1)
Figure 3.19 A reference point (red) and a triangular surface element (bold) in a three di-
mensional coordinate system, where the three element vertices lie on the x-y plane.
within a rectangular volume. The side lengths of this volume, ∆x, ∆y and ∆z, are
shown in Figure 3.15(c), and were used to calculate the path centroid,
 
∆x/2
 
P∆ =  ∆y/2  , (3.2)
∆z/2
where P∆ is assumed to be the neutral position of the reference point motion in world
space.
3.5.2.2 Surface Element Normal Vectors
Projecting a reference point centroid, P∆ , to the underlying surface mesh, first involves
the calculation of a surface element normal vector, ne , for all surface elements in the
mesh. This vector can be calculated for any surface element as the cross product of any
two vectors that lie on the plane of the element. For the element shown in Figure 3.19,
   
 x2 − x1
 
  x3 − x1
 

e
n = y2 − y1 × y3 − y1 . (3.3)

  
  

z2 − z1 z3 − z1
where the x, y and z indices refer to the surface node locations indicated in the figure.
Calculated values for ne for all surface elements were normalized by vector length, and
are shown in Figure 3.18(b) as red arrows.
3.5.2.3 Surface Node Normal Vectors
Following the calculation of surface element normal vectors, a nodal normal vector,
nn , is calculated for all surface nodes. For any surface node i, this process first uses
the surface element incidence list to identify the k surface elements in the mesh that
contained the node. The surface node’s normal vector, nni , is then defined as the mean
of the k connected surface element normal vectors, so that
k
1 X ¡ e¢
nni = nj , (3.4)
k
j=1
Surface node normal vectors were also normalized by length, and are shown in Fig-
ure 3.18(c) as blue arrows.
3.5.2.4 Linear Basis Function Values
After the calculation of surface node normals, nn , for all surface nodes, all surface
elements were projected along their constituent nodal normals until the element plane
passed through the reference point centroid, P∆ . The surface element with a projec-
tion bounding P∆ was then identified, and is shown in Figure 3.18(d). Values from
interpolating polynomials within the projected element containing P∆ were used to
convert the reference point location back to the surface of the mesh, shown pictorially
in Figures 3.18(e)-(f). The relationship between the projected surface element vertices
and P∆ was calculated using linear basis functions φ, where
£¡ 0 ¢¤
φ1 = x2 y3 − x03 y20 ) + P∆ 0 0 ∆ 0 0
x (y2 − y3 ) − Py (x2 − x3 ) /2A
0
£¡ 0 ¢¤
φ2 = x3 y1 − x01 y30 ) + P∆ 0 0 ∆ 0 0 0
x (y3 − y1 ) − Py (x3 − x1 ) /2A , (3.5)
£¡ 0 ¢¤
φ2 = x1 y2 − x02 y10 ) + P∆ 0 0 ∆ 0 0
x (y1 − y2 ) − Py (x1 − x2 ) /2A
0
for the projected surface element shown in Figure 3.19. The calculated φ values could
then be used to interpolate any FE-simulated displacement field to the point on the
surface of the mesh that best matched the reference point location. This corresponding
surface location is shown in Figure 3.18(f).
3.6 Measured Data Fitting
To make a meaningful comparison between the discrete experimentally measured mo-

tion at a reference point, P̄, and a FE-simulated displacement field interpolated to the
3.6 MEASURED DATA FITTING 63
surface location matching P∆ , the experimental data must be expressed as orthogonal

components of a steady state harmonic displacement vector. Figure 3.20(a) shows an
experimentally measured reference point motion track in a world space co-ordinate sys-
tem, including noise and any computational error. The approximations to this motion
from the range of fitting methods developed are shown in Figure 3.20(b)-(d).
P∆
P0
z
x y
(a) Measured mo- (b) Simple vector. (c) Basic ellipse. (d) Refined ellipse.
tion.
Figure 3.20 The evolution of measured motion fitting techniques.
3.6.1 Steady State Measured Motion
The steady state oscillation of a point about an origin can be expressed using a time-
harmonic displacement vector, defined at any time t as
n o
ū(t) = < ueiωt , (3.6)
where ω is the frequency of the system, and the amplitude term contains both real and
imaginary components, u = uR − iuI . The measured position of a reference point at
time ti , P̄i , can be approximated as the addition of a harmonic displacement, ū(ti ), to
the reference point centroid, P∆ , as
P̄i = ū(ti ) + P∆ . (3.7)
Alternatively, reference point motion can be described using a real-valued ampli-

tude, ũ, and a phase, ϕ, yielding
n o
ū(t) = < ũei(ωt+ϕ) . (3.8)
The motion description in Equation 3.8 can be converted to the form of Equation 3.6 by
relating the real-valued damped amplitude, ũ, and phase, ϕ, to a complex amplitude,
u = uR − iuI , as follows:
n o n o
< ueiωt = < ũei(ωt+ϕ)
n o n o
< ueiωt = < ũeiωt eiϕ
u = ũeiϕ
uR − iuI = ũ cos(ϕ) − iũ sin(ϕ). (3.9)
Experimentally measured damped motion behavior can be approximated by obtaining

an undamped motion amplitude, u, from the real component of the expression in
Equation 3.9, so that
u ≈ uR = ũ cos(ϕ). (3.10)
3.6.2 Undamped Fitting
When assuming that the measured reference point motion path, P̄, is generated by an
undamped material, this motion can be described by Equation 3.6, where imaginary
displacement components were ignored, allowing u to be approximated by uR . The
components of this real-valued amplitude vector were estimated using the bounding
box method described in Section 3.5.2.1, where
 
∆x/2
 
u =  ∆y/2  . (3.11)
∆z/2
An example of the estimated motion vector generated by this undamped assumption

is shown in Figure 3.20(b).
3.6.3 Damped Fitting
When assuming the measured motion track, P̄, is generated by a material that has some
internal damping, Equation 3.8 can be used to generate a motion fit defined using a
damped amplitude, ũ, and phase, ϕ. A two stage algorithm was developed to optimize
the amplitude and phase values, where the first stage created initial estimates, and a
second stage refined the estimates using non-linear squared error optimization.
Equation 3.11 can be used to generate a basic estimate of damped amplitude, ũ,
using an undamped assumption. A three-element vector containing an initial phase
3.7 SURFACE ERROR CALCULATION 65
estimate for each orthogonal direction, ϕ, can then be made using the measured ref-
erence point location at time t = 0, P̄0 , and the centroid of the bounding rectangle,
P∆ ,
³¯ ¯ ´
ϕx = sin−1 ¯P̄0x − P∆
x
¯ /ũ x
³¯ ¯ ´
ϕy = sin−1 ¯P̄0y − P∆ ¯
y /ũy . (3.12)
³¯ ¯ ´
ϕz = sin−1 ¯P̄0z − P∆
z
¯ /ũz
Here, the translation P̄0 − P∆ allows the measured point location to be related to the
origin.
The sign of each component in the estimated phase value, ϕ, can not be determined
based on P̄0 alone, as the velocity of the reference point at this location is unknown.
Therefore, the evaluation of eight possible combinations of ±ϕx , ±ϕy and ±ϕz is re-
quired before an initial phase estimate ϕ can be found. Each phase combination was
considered by first generating an estimated point path at the discrete time intervals
used for motion capture ti using Equation 3.8. The disagreement between this esti-
mated path, ū(ti ), and the measured motion track for the point, P̄, was measured by
the error term
r°
X ° ¡ ¢°
°2
Γ= °ū(ti ) − P̄i − P∆ ° , (3.13)
i
where the translation P̄i − P∆ expressed the measured point positions relative to the
world space origin. The phase combination with the lowest value of Γ provided the
initial estimates of ϕx , ϕy and ϕz . An estimated motion track generated using these
initial damped amplitude and phase values is shown as a blue ellipse in Figure 3.20(c).
The second stage of the damped motion estimation refined the initial estimates
of damped amplitude and phase with a simple least-squares optimization, where the
objective function for minimization was taken from Equation 3.13. The result of this
optimization was a closely fitted motion estimate containing damped amplitude, ũ,
and phase, ϕ, for each reference point, expressed as orthogonal components. The red
motion track in Figure 3.20(d) was generated using these optimized values.
3.7 Surface Error Calculation
Once the measured motion data was represented in the same coordinate space as the
FE model, the quantitative match between the simulated and measured motions could
be evaluated. The error metric used varied depending on the assumed FE material
model and the methods used for reference point projection and motion fitting.
3.7.1 Soft Gelatine Phantom Surface Error
The undamped FE model used for soft gelatine phantom simulations calculated the
time-harmonic reference point motions in the form
n o
ū(t) = < ucR eiωt , (3.14)
where ucR represents a calculated, real-valued displacement amplitude that may be

negative. The undamped real motion fitting described in Section 3.6.2 allowed the
measured motion to be expressed as a corresponding real displacement amplitude,
um
R . Using the mesh alignment method described in Section 3.5.1, the measured real
amplitude values at each reference point, um
R , could be directly compared with the
real amplitude values output from FE simulation, ucR . In this case, the error metric,
Ω1 , was calculated at all N reference points as the error between the measured and
simulated motion amplitudes,
1 X ¯¯ m ¯
N
c ¯
Ω1 = ¯ |uR | − |uR | ¯ . (3.15)
N i
i=1
By only comparing the magnitude of the displacement vectors, this error calculation
avoids any possible errors due to the lack of phase information.
3.7.2 Hard Gelatine Phantom Surface Error
A second variant of the surface error metric was used for the hard gelatine experimental
data set described in Section 3.3.1. In this case, FE simulation again assumed an
undamped model, and provided real amplitude ucR . However, the measured motion
was fitted using the damped vectors for amplitude, ũ, and phase, ϕ, described in
Section 3.6.3. Here, the surface error metric Ω2 was defined
N
Ãq !
1 X ° °2
°um − uc °
Ω2 = R R (3.16)
3N
i=1 i
where the equivalent undamped measured motion, um

R , was approximated from ũ and
ϕ using Equation 3.10 before the error metric was calculated.
3.7.3 Silicone Phantom Surface Motion Error
A final motion error metric value was used for the motion data sets obtained from
the silicone phantoms described in Section 3.3.2. For these simulations, the FE model
included damping, and provided calculated displacement amplitudes containing both
real and imaginary components, uc = ucR + iucI . The measured motion, um , was fitted
using damped amplitude, ũ, and phase, ϕ, and was expressed as real and imaginary
3.8 SUMMARY 67
components, um = um m
R + iuI , using Equation 3.9. In this case, the simulated displace-
ment amplitude, uc , was compared to the measured displacement amplitude, um , at
all N reference points, using
1 X ³¯¯ m ¯´
N
¯ ¯
Ω3 = uR − ucR ¯ + ¯um
I − uc¯
I . (3.17)
3N i
i=1
3.8 Summary
The experimental methods described in this chapter provide both real experimental
data, and the methods required to compare this data to corresponding FE-simulated
displacements. As each experimental data set was obtained, a series of further experi-
ments were performed in an attempt to estimate or reconstruct the material properties
of the phantoms. These experiments are described in the following chapters.
Chapter 4
Phantom Surface Motion Error Studies
The experimental methods described in Chapter 3 provided surface motion data from
a range of both homogeneous and heterogeneous soft tissue approximating phantoms.
To further understand the relationship between internal material properties and exper-
imentally observed motion, an error study was performed on the experimental phantom
displacement sets to investigate the suitability of surface motion error as an indicator
of internal stiffness distribution. This work involved comparing the measured phantom
motion data collected in Chapter 3 with large sets of FE-simulated displacements for
a variety of property distributions. The distribution with the lowest resulting surface
motion metric could then be compared with the known properties of the phantom to
determine the potential effectiveness of the DIET system and method. The experi-
ments performed in this chapter represent an evaluation of surface motion error as an
indicator of internal stiffness. Should such an approach give favorable results, an op-
timization algorithm will be required to automate the process of finding the minimum
error value.
The surface error studies involved FE simulation of phantom displacement using

a mesh based on the measured phantom geometry. The internal stiffness distribu-
tion of the FE model was described using a number of different parameters. All FE
simulation used computer code written in a combination of Fortran and Linux shell
scripting, with code compiled and executed on Linux workstations. The matrix inver-
sion required during FE simulation was performed using routines from LAPACK, and
the matrix solver MUMPS, a direct sparse matrix inversion software package written
in Fortran 90 [105, 106]. In addition, the BLAS library developed by Kazushige Goto
allowed matrix algebra to be performed in parallel where the workstation had multiple
processors available for computation [107].
70 CHAPTER 4 PHANTOM SURFACE MOTION ERROR STUDIES
4.1 Homogeneous Gelatine Phantoms
Experimental data sets were collected from homogeneous gelatine phantoms created
during the work described in Section 3.1.1. These experimentally measured motion
sets represented the first successful data collection using the DIET actuation and mo-
tion measurement system. The surface error studies performed on this experimental
motion data aimed to estimate a value for Young’s Modulus, E, for the gelatine used to
create each phantom. Error studies for the soft and hard gelatine phantoms were per-
formed sequentially, but shared a common objective and approach, and are described
simultaneously in this section.
4.1.1 Measured Phantom Displacement
The actuation and surface motion capture of both gelatine phantoms was performed
using the experimental setup described in Section 3.3.1. Figure 4.1 shows the surface
deformation of both hard and soft gelatine phantoms during actuation, along with the
location of each reference point centroid across the motion cycle. The tracked position
of all reference points on the soft gelatine phantom was fitted using the undamped
motion assumption described in Section 3.6.2, resulting in real amplitude components
at each reference point centroid.
Reference point motion on the surface of the hard gelatine phantom was estimated
using the more accurate fitting algorithm described in Section 3.6.3, where point motion
was described about the motion centroid using damped amplitude and phase values.
A selection of measured reference point locations over time, P̄, on both phantoms, are
compared with the corresponding fitted motion estimates, ū(t), in Figure 4.2.
4.1.2 Finite Element Simulation
The assumed homogeneity of the gelatine phantoms allows a single value for each mate-
rial property to be assigned across the entire model when simulating displacements us-
ing FE methods. Computational limitations constrained the gelatine phantom surface
error investigations to variations in Young’s Modulus, E. Both hard and soft phan-
tom models had material density, ρ, set at 1,000 kg/m3 , and Poisson’s Ratio ν = 0.49,
representing a close approximation to water, the main ingredient used in gelatine man-
ufacture.
Computer models of both cylindrical phantom geometries were created and meshed
using the software program GambitTM (Version 2.2.30, Fluent Inc.), where key model
dimensions were taken from physical measurement of the actual phantoms. The result-
ing soft phantom model had a diameter of 70 mm, and a height of 50 mm, while the
hard phantom was 75 mm diameter, and 69 mm tall. The FE meshes for the soft and
4.1 HOMOGENEOUS GELATINE PHANTOMS 71
(c)
(a) Soft gelatine motion capture. (b) Tracked reference point locations. (c)
(f)
(d) Hard gelatine motion capture. (e) Tracked reference point locations. (f)
Figure 4.1 The experimentally observed motion of the homogeneous gelatine phantoms.
z z
x y 0.5mm x y 0.5mm
(a) Soft phantom, surface reference point 1. (d) Hard phantom, surface reference point 1.
0.2mm 0.3mm
(b) Soft phantom, point 6. (e) Hard phantom, point 7.
0.1mm 0.3mm
(c) Soft phantom, point 12. (f) Hard phantom, point 14.
Figure 4.2 A selection of measured (black) and fitted (red) reference point motion from the
soft and hard homogenous gelatine phantoms.
hard models contained approximately 37,000 and 120,000 linear tetrahedral elements
respectively, and are shown in Figure 4.3.
(a) Soft phantom mesh. (b) Hard phantom mesh
Figure 4.3 The FE meshes used to simulate soft and hard gelatine phantom motion.
The increase in mesh density for the hard model was made possible by an increase
in computational resources. Surface node locations on the soft phantom mesh were
manipulated to match the reference point locations on the phantom surface according to
the mesh alignment process outlined in Section 3.5.1. The hard gelatine phantom model
did not require any data-specific mesh manipulation due to the subsequent development
of the surface projection algorithm described in Section 3.5.2.
Boundary conditions were applied to both FE models that closely matched the
experimental conditions. The bottom face of the mesh was displacement-constrained
in the horizontal x-y plane, based on the assumption of no slip between the base
of the phantom and the actuation plate. In addition, the bottom surface of each
model was harmonically displaced in the z direction at frequency and amplitude values
corresponding with the experimental data collection. The soft phantom actuation was
at an amplitude of 0.6 mm, and frequency 50 Hz, while the hard phantom had an
amplitude of 0.975 mm, and a frequency of 50 Hz. All other external faces on both
models had free surface boundary conditions, where displacements were not specified,
and boundary forces were set to zero in all orthogonal directions. Table 4.1 lists the
FE model specifications used when simulating the motions of both the soft and hard
gelatine phantoms.
A ‘sweep’ of FE simulations was performed on both the soft and hard computa-
tional models, where full volume displacements were simulated at a range of Young’s
Modulus values. The modulus range for each model was chosen based on the expected
material properties from the mechanical testing in Section 3.2 and Table 3.3. Soft
Table 4.1 Specifications of the hard and soft gelatine models used for FE simulation.
FE Model
Specification Soft Hard
Height (mm) 50 69
Diameter (mm) 70 75
Elements (approx.) 37,000 120,000
Frequency (Hz) 50 50
Amplitude (mm) 0.6 0.975
Density (kg/m2 ) 1,000 1,000
Poisson’s Ratio 0.49 0.49
model simulations were performed between E = 5–20 kPa, and hard model simulations
between E = 5–40 kPa. Both sets of simulations were performed at stiffness incre-
ments of 100 Pa, providing a good balance between resolution and overall simulation
time. The resulting sets of motion data characterized the simulated surface deforma-
tion over the prescribed stiffness range, where the surface deformation at a selection of
Young’s Modulus values is shown for both phantoms in Figure 4.4.
4.1.3 Stiffness Estimation
The displacement fields generated using FE simulation formed a database of solutions

that were used to evaluate surface motion error over the range of stiffness values con-
sidered. Surface motion error metrics from Section 3.7 were used to compare each
simulated displacement field to the corresponding measured motion. An estimate of
the phantom stiffness value was taken from the Young’s Modulus value correspond-
ing with the lowest error metric, and could be compared with the measured gelatine
stiffness values described in Section 3.2.
The surface error metric Ω1 defined in Equation 3.15 was used to evaluate the
surface displacement error for the soft gelatine phantom. Use of the mesh alignment
technique from Section 3.5.1 in conjunction with the undamped FE solution allowed
measured reference point amplitudes to be directly compared to simulated amplitudes
at the corresponding mesh nodes. A plot of the surface error metric over the 5–20 kPa
stiffness range is shown in Figure 4.5.
Areas of high motion error on the plot correspond to resonant stiffness values in the
FE model where simulated displacements were large and did not match the measured
motion. The lowest motion error metric on the figure corresponds to a Young’s Modulus
value of E = 12.5 kPa, though the stiffness corresponding to global minimum error is
not clearly defined, with other significant local minima at 6.1 kPa and 9.5 kPa. The
(a) Soft phantom model, E = 8.5 kPa. (d) Hard phantom model, E = 10 kPa.
(b) Soft model, E = 9.5 kPa. (e) Hard model, E = 17.5 kPa.
(c) Soft model, E = 12.5 kPa. (f) Hard model, E = 27.5 kPa.
Figure 4.4 FE-simulated displacement for the soft and hard gelatine phantom models across
a range of Young’s Modulus values.
Motion Error Metric, Ω1 (mm) 100
10.0
1.00
E = 9.5kPa
0.49
E = 6.1kPa Minimum Error, E = 12.5kPa
5 10 15 20
Figure 4.5 Motion error metric (on a logarithmic scale) across a range of homogenous stiffness
values for the soft gelatine phantom, where the location of minimum error is indicated.
50
Point 1
Point 8
Point
Height Measured
(mm) Simulated,
E = 12.5kPa
Point 7 Point 14
0
0 1 2 0 1 2
Displacement Amplitude (mm) Displacement Amplitude (mm)
Figure 4.6 A comparison of reference point displacement amplitudes between the experi-
mental soft gelatine phantom and FE-simulated motion at E = 12.5 kPa.
simulated displacement amplitudes at E = 12.5 kPa are compared with the measured
amplitudes at all fourteen reference points in Figure 4.6, where the reference point
heights are plotted to scale.
FE simulation of the hard gelatine phantom model generated undamped displace-

ment amplitudes at Young’s Modulus values in the range 5–40 kPa. The error between
the simulated and measured displacements was defined by the metric Ω2 in Equa-
tion 3.16, where the measured motion’s fitted estimates of damped amplitude and
phase were converted to an estimated undamped amplitude prior to error calculation.
Values for Ω2 across the stiffness range simulated are shown in Figure 4.7, where the
minimum motion error corresponds to a Young’s Modulus value of E = 27.1 kPa. Here,
the stiffness corresponding to global error minimum has increased definition versus the
local minima when compared to the soft gelatine case. The match between the com-
ponents of simulated displacement at E = 27.1 kPa and the measured displacement is
shown for all reference points in Figure 4.8.
100
Motion Error Metric, Ω2 (mm)
10.0
1.00
0.38
Minimum Error, E = 27.1kPa
5 10 15 20 25 30 35 40
Figure 4.7 Motion error metric across a range of homogenous stiffness values for the hard
gelatine phantom, where the location of minimum error is indicated.
4.1.4 Discussion
The Young’s Modulus values estimated using the stiffness sweep method showed promis-
ing correlation with the mechanically measured stiffness values obtained in Section 3.2.
Measured and estimated Young’s Modulus values are summarized in Table 4.2, where
the estimated modulus value was within 40% of the measured modulus value for the
soft phantom, and within 10% for the hard phantom.
The experimentally measured motion from the soft phantom was highly erratic,
and often not a good representation of the expected smooth reference point motion
path. This irregular motion can be seen in Figure 4.2, particularly for points 6 and 12.
69
Point 1
Point
Height
(mm)
Point 15
Measured
Simulated, E = 27.1kPa
0
{2 0 2 {2 0 2 {2 0 2
x y z
Orthogonal Displacement Amplitude (mm)
Figure 4.8 The match between FE-simulated displacements at E = 27.1 kPa and experimen-
tal displacements for the hard gelatine phantom.
Table 4.2 A comparison of surface error estimated and measured Young’s Modulus values
for the soft and hard gelatine.
Young’s Modulus, E
Material Measured Estimated
Soft Gelatine 12.5 kPa 9 kPa
Hard Gelatine 27.1 kPa 26 kPa
Obtaining accurate positions for the reference points on the soft phantom was made
difficult by a combination of noise within the actuation and strobing system, and the
large, non-uniform appearance of the reference points on the phantom surface. Ad-
ditional error was introduced after motion capture through the use of an undamped
amplitude vector to approximate the reference point motion path, where no phase
information was considered.
The observed surface motion of the hard gelatine phantom in Figures 4.1 and 4.2
demonstrates the improvement in the motion capture system subsequent to the soft
gelatine phantom experiments. These improvements in the experimental setup, in con-
junction with the damped motion path fitting algorithm, led to the much closer match
between measured and estimated reference point paths visible in Figure 4.2. Here,
the clearly ellipsoidal nature of the measured motion indicated that the gelatine ma-
4.2 HOMOGENEOUS SILICONE PHANTOMS 79
terial was damped, and therefore not ideally suited to simulations using an undamped
material model.
FE simulation results from both phantom models indicated resonance issues at a

large number of simulated stiffness values. Calculated surface motions at these resonant
stiffness values were unrealistically large and would prevent a close match between the
measured and simulated motion should the phantom’s actual Young’s Modulus lie at or
near a resonant peak. The reduction in observed resonance around the minimum error
location for the hard gelatine phantom was partly attributable to the longer surface
wavelengths resulting from higher model stiffness values. Circumventing these reso-
nance issues would require the incorporation of material non-linearity and/or damping
into the FE model, but at the cost of added material parameters for material property
reconstruction.
The aim of the gelatine phantom surface error study was to investigate the behavior
of the linear FE model over a range of stiffness values, and evaluate the potential of
the DIET method as a stiffness estimation tool when used with real data. The sweeps
performed indicated that limited surface motion error measurements have the potential
to be used as an indicator of stiffness in a realistic experimental situation, validating
the potential of the overall DIET approach.
The clear identification of a best-fit stiffness value was found to be dependent to

a large extent on the accuracy of the experimental motion capture and fitting algo-
rithms. This effect is particularly evident when considering the improvement in both
the appearance of the error domain and the minimum error value for the hard gelatine
simulations. Resonant behavior was identified as a major obstacle to more generalized
reconstruction, requiring special attention if robust methods in elasticity imaging using
DIET are to be achieved. Despite the issues identified, this straightforward approach
to estimating elasticity was able to clearly differentiate between soft and hard material
with a contrast as low as 3x, indicating the potential of a surface error based system
to be clinically useful after appropriate refinements.
4.2 Homogeneous Silicone Phantoms
The stiffness sweeps and resulting modulus estimations performed in Section 4.1 used
an undamped FE model, where all simulated motions were treated as being in phase.
While the stiffness estimations made using this undamped model showed good agree-
ment with independent mechanical property testing, the ellipsoidal reference point
motion observed experimentally indicated that the real phantom displacements were
generated by a viscoelastic material. Including these viscoelastic effects in a damped
FE material model would allow a more thorough consideration of the underlying me-
chanical properties, with the added benefit of reducing the resonance issues observed
when performing the gelatine model simulations.
The decision to use silicone phantoms for this study arose from the difficulties en-
countered using gelatine for phantom manufacture, as outlined in Section 3.1.1. The
new silicone material was expected to have similar viscoelastic properties as the gela-
tine material used previously, confirming the need for a damped material model when
simulating displacements. The first silicone surface error studies were performed on
data from both soft and hard homogeneous phantoms, with ingredients and measured
material properties identified in Sections 3.1.2 and 3.2.
The homogenous silicone error studies focused on optimizing a total of three in-
dependent material parameters, where each parameter was assumed to be constant
over the entire model. The stiffness of a viscoelastic material undergoing steady-state
harmonic displacement can be described by a complex modulus,
E ∗ = E 0 + iE 00 , (4.1)
where the storage modulus E 0 relates strains to stored elastic energy, and the loss modu-
lus, E 00 , describes energy lost to internal attenuation during vibration. The derivation
of this complex-valued elastic modulus description is included in Appendix A. This
stiffness description was parameterized for the damped material model so that the
loss modulus was calculated as the product of a storage modulus E 0 and a damping
parameter ζ, such that
E 00 = ζE 0 . (4.2)
In addition to the viscoelastic material modulus represented by E 0 and ζ, Poisson’s

ratio, ν, was also included as an independent parameter as the actual value of ν for the
silicone was unknown. Variations in all three material parameters were used to esti-
mate property values for the soft and hard silicone material using the same parameter
sweeping method used for the gelatine phantoms in Section 4.1.
Experimental motion was captured from both hard and soft homogeneous silicone phan-
toms using the system described in Section 3.3, where an improved actuation system
had been implemented following the experimental work with gelatine phantoms. Fig-
ure 4.9 shows single frames from the motion capture of both hard and soft silicone
phantoms, along with the tracked reference point motion across the steady state mo-
tion cycle for each phantom. All reference point motion was observed to be continuous
and steady-state, confirming the accuracy of the enhanced actuation and motion cap-
ture system. Using the damped motion estimation algorithm from Section 3.6.3, a best
fit estimate of the amplitude and phase values was calculated for each of the observed
ellipsoidal reference point positions across all frames.
(c)
(a) Soft silicone motion capture. (b) Tracked reference point locations. (c)
(f)
(d) Hard silicone motion capture. (e) Tracked reference point locations. (f)
Figure 4.9 The experimentally observed motion of the homogeneous silicone phantoms.
A FE model matching the cylindrical phantom geometry was generated and meshed
using Gambit. The surface projection technique described in Section 3.5.2 allowed a
single FE mesh to represent both silicone phantom models, as both phantoms had the
same external dimensions of 74 mm diameter, 74 mm height. The final model contained

approximately 92,000 linear tetrahedral elements, with an average element edge length
of 3 mm, and is shown in Figure 4.10. The complex-valued stiffness matrix terms in
a damped FE simulation required twice as much computer memory as an undamped
model, leading to the element size increase when compared to the undamped hard
gelatine mesh used in Section 4.1.
Figure 4.10 The FE mesh used for all silicone phantom simulations.
To match the experimental setup, the FE model was displacement constrained in

all three axes on the lower face, with x and y displacements set to zero, and the z
direction displacement described by an 0.5 mm amplitude, 100 Hz sinusoid. The model
density was set at 1,000 kg/m3 as an approximation to the measured density values for
both phantoms. The single density value used was within five percent of the measured
values for these phantoms, and avoided the need for an additional independent material
parameter.
Storage modulus ranges of E 0 = 10–50 kPa for the soft model and E 0 = 80–150 kPa
for the hard model were covered in a grid search using modulus increments of 2 kPa.
The damping ratio, ζ, ranged from 0.1–0.3 in steps of 0.02 for both models. Simulation
sets were generated for each model at Poisson’s ratio values between ν = 0.40–0.49, the
estimated range for the silicone material. The match between a FE-simulated data set
interpolated onto the reference points and the measured motion data at the same refer-
ence points was represented by the error term Ω3 previously defined in Equation 3.17.
In this case, the metric compared both real and imaginary displacement components
of the damped motions.
4.2.3 Mechanical Property Estimation
Following FE simulation, the error metric Ω3 was calculated for both hard and soft
silicone phantom simulations across the ranges of storage modulus E 0 , damping ratio ζ
and Poisson’s ratio ν considered. Surface error results for the soft phantom simulations
are shown in Figure 4.11(a), where a plane of constant ν passing through the location
of global minimum error is plotted. Motion error varies from blue (low) to red (high)
on the plot, where the minimum error metric corresponds to a storage modulus value
E 0 = 28 kPa, damping ratio ζ = 0.22, and Poisson’s ratio ν = 0.48. Variation in
the error metric with changes in Poisson’s ratio at constant E 0 and ζ are shown in
Figure 4.11(b).
0.109 Motion Error Metric, Ω3 (mm) 0.602 0.336
30
Damping Ratio, ζ (%)
22
0.109 Minimum Error, ν = 0.48
10
10 28 50
0.40 0.49
Storage Modulus, E 0 (kPa) Poisson’s Ratio, ν
0
(a) Parameter sweep results at ν = 0.48, (b) Variations in Ω3 with ν, where E = 28 kPa, and
where the location of global minimum er- ζ = 0.22.
ror is indicated (◦).
Figure 4.11 Displacement sweep surface error results for the soft homogeneous silicone phan-
tom.
Displacement error results for the hard silicone phantom are shown in Figure 4.12,
where the global error minimum corresponds to parameter values E 0 = 103 kPa, ζ =
0.20, and ν = 0.41. Figure 4.12(b) shows variations in Ω3 with Poisson’s ratio at
constant E 0 and ζ. All motion components from the closest simulated matches to the
measured motion from the soft and hard phantoms are shown in Figure 4.13, where
the vertical location of each reference point is plotted to scale.
4.2.3.1 Storage Modulus Frequency Dependence
Assuming the material storage modulus, E 0 , is independent of actuation frequency as-

sumes that it is comparable to the static measurement of Young’s Modulus, E. There-
fore, estimated values for E 0 should not vary when performing an error analysis at
varying frequency values. To test the frequency dependence of the estimated value
for E 0 , an additional experimental data set collected from the soft silicone phantom
at a frequency of 50 Hz was considered. A parameter sweep was performed where the
0.098 Motion Error Metric, Ω3 (mm) 0.519 0.132

30
Damping Ratio, ζ (%)
20
Minimum Error, ν = 0.41
0.984
10
90 103 150
0.40 0.49
Storage Modulus, E 0 (kPa) Poisson’s Ratio, ν
(a) Parameter sweep results at ν = 0.41, (b) Variations in Ω3 with Poisson’s Ratio, where E 0 =
where the location of global minimum er- 103 kPa, and ζ = 0.20.
ror is indicated (◦).
Figure 4.12 Displacement sweep surface error results for the hard homogeneous silicone
phantom.
FE-simulated amplitude and frequency matched the 50 Hz experimental data, and E 0

was varied from 10–50 kPa. Values for damping and Poisson’s ratios were held constant
for these simulations at the previously estimated values of ζ = 0.22 and ν = 0.48 for
the soft silicone. The resulting error metric values over the range of E 0 considered
are shown in Figure 4.14. Here, the storage modulus value corresponding to minimum
motion error is E 0 = 29 kPa, a difference of only 1 kPa from the E 0 = 28 kPa found
with 100 Hz actuation.
4.2.4 Discussion
The motion error plots in Figures 4.11–4.12 provide important information about the
sensitivity of the motion error metric Ω3 to variations in the three material parameters
considered. The error contours in Figure 4.11(a) indicate the soft phantom motion
error has a high sensitivity to storage modulus, while variations in damping ratio have
a smaller influence. The wide contour spacing in Figure 4.12(a) indicates that at higher
stiffness values, the sensitivity of Ω3 to variations in E 0 and ζ is more balanced and
less significant. The relatively high error sensitivity to parameter variations in the soft
phantom is most likely due to the large proportional change in shear wavelength in
the softer material when stiffness is varied. Motion error sensitivity to Poisson’s ratio
followed a similar trend, where variations in Ω3 with ν were much more significant in
the soft model than in the hard model.
The addition of damping to the material model significantly reduced the effects
of resonance that were readily visible in the previous undamped studies, particularly
at low stiffness values. Resonant behavior was almost completely eliminated with the
addition of a viscoelastic damping effect for the silicone simulations, resulting in contin-
uous error domains where the parameter values corresponding to the location of global
Measured Measured
74
Simulated Simulated
Point 1
Point
Height
(mm)
Point 15
0
{1 1 {1 1 {1 1 {1 1 {1 1 {1 1
x y z x y z
Real Displacement (mm) Imaginary Displacement (mm)
0
(a) Soft phantom, simulated displacement at E = 28 kPa, ζ = 0.22, ν = 0.48.
Measured Measured
74 Simulated Simulated
Point 1
Point
Height
(mm)
Point 15
0
{1 1 {1 1 {1 1 {1 1 {1 1 {1 1
x y z x y z
(b) Hard phantom, simulated displacement at E 0 = 103 kPa, ζ = 0.20, ν = 0.41.
Figure 4.13 The best fitting FE-simulated cases compared with measured motions for both
soft and hard silicone phantoms.
0.582
Minimum Error, E 0 = 29kPa

0.203
10 20 30 40 50
Storage Modulus, E 0 (kPa)
Figure 4.14 Motion error plotted against stiffness for the soft silicone phantom actuated at
50 Hz.
error minimum had clearer definition than in the undamped results from the gelatine
phantoms. The distinctness of this global minimum error location is important when
considering the ability of reconstruction algorithms to prefer global optimal parameter
values over values representing local minima.
The static mechanical testing results from Section 3.2 were a close match to the
estimated modulus values based on Ω3 for the soft silicone, with the estimated value
for E 0 within 15% of the measured static modulus value shown in Table 4.2. The
stiffness of the hard phantom was under-estimated by approximately 30 kPa, or 25%,
when compared with the mechanical test result. One possible source of this error was
a reduction in observed stiffness due to viscoelastic heating of the phantom during the
relatively high frequency actuation. Such heating would be more prevalent in the hard
silicone, assuming a constant damping ratio. In addition, the static testing boundary
conditions discussed in Section 3.2 were found to overestimate the material stiffness
measurement by up to 10%. Despite the hard modulus discrepancy, the error sweeps
were clearly able to differentiate silicone material stiffness values with a contrast as
low as 400–500%. This latter result is more important for the DIET system concept
than the exact characterization of material parameter values in hard material, given
clinically reported contrasts of up to 1,200% [54–58].
The modulus values in Table 4.3 show that the estimated storage modulus value
did not significantly change when the error study was performed on soft phantom
experimental data at 50 Hz. This result provides evidence that the estimated storage
modulus of the silicone was not strongly dependent on frequency, validating the use
4.3 HETEROGENEOUS SILICONE PHANTOMS 87
of a static mechanical test as an independent measurement of the silicone stiffness.

Due to a similar chemical composition, the hard silicone was be expected to display
comparable behavior.
Table 4.3 Elastic modulus values for the soft and hard silicone phantoms, where modulus
values estimated using a surface error metric are compared to the measured values from inde-
pendent static testing.
Elastic Modulus
Method Frequency Soft Silicone Hard Silicone
Surface Error 100 Hz 28 kPa 103 kPa
Surface Error 50 Hz 29 kPa -
Measured (DMA) static 32 kPa 135 kPa
Geometric non-linearity is one possible explanation for the trend of slightly under-
estimated displacement amplitudes observed in Figure 4.13. Large measured displace-
ments lead to an increased mismatch with the small strain linear elastic assumption
which underpins the FE simulations. That the match between measured and simulated
motions is closer toward the top of the phantoms, as evident in Figure 4.13, where vi-
bration amplitudes are lower is further evidence of this effect. Another possible cause
of the small displacement mismatch is drive miscalibration, where the actuator dis-
placement may have had a slightly greater amplitude than the prescribed waveform.
The error study performed on the homogeneous silicone data showed that stiffness
estimation using the DIET system is possible using a damped material model, with
good agreement to mechanical testing results, particularly at lower stiffness values.
The method was able to accurately characterize damped material motion on the sur-
face of silicone phantoms to create fitted displacement estimates. Material parameter
estimations made for both hard and soft silicone phantoms showed clear differentia-
tion between the storage modulus of hard and soft materials, where optimal parameter
values were clearly identified, and resonance was not observed to be a significant issue.
4.3 Heterogeneous Silicone Phantoms
Following the successful estimation of silicone material properties using homogeneous

phantoms and a damped material model, a surface error study was performed using the
experimental data captured from the two heterogeneous silicone phantoms described
in Section 3.3.2. Here, the aim was to evaluate the ability of a surface error metric
to estimate a discrete variation in the internal material properties of heterogeneous
phantoms. The simple phantom geometries allowed the internal stiffness distribution
to be described by a limited number of shape-based parameters, facilitating the use of

the same parameter sweep techniques applied to the homogenous phantoms.
The material properties estimated for both the hard and soft silicone during the
heterogeneous surface error study in Section 4.2 are summarized in Table 4.4. The
best estimates of damping ratio, ζ, and ν, for the soft and hard silicone were fixed for
all heterogeneous simulations, allowing the study to focus on storage modulus values.
Clinically, such a simple approach would occur when seeking a tumor at a fixed, large
stiffness contrast to healthy tissue values around 20–30 kPa.
Table 4.4 Silicone material properties estimated from the homogenous silicone phantoms
using surface error.
Material E0 ζ ν
Soft Silicone 28 kPa 0.22 0.48
Hard Silicone 103 kPa 0.20 0.41
Motion capture from the heterogeneous silicone phantoms was performed using the
same experimental setup as the homogeneous silicone phantoms, where the compo-
nents used are described in Section 3.3.2. The tracked reference point positions in
three dimensional world space were estimated using the damped fitting algorithm from
Section 3.6.3. The observed surface motion for the fifteen reference points on the surface
of each heterogeneous phantom are shown in Figure 4.15.
With material property estimates for the soft and hard silicone in hand, a range of
simulations were performed on FE models of the stacked and concentric geometries.
The identical size and shape of all silicone phantoms allowed the mesh used for the
homogenous silicone phantoms in Section 4.2 to be used for all heterogeneous model
simulations. All boundary conditions applied to the models matched the experimental
testing conditions, where the actuation was from below in the vertical direction, at
0.5 mm amplitude, 100 Hz frequency.
A two-parameter, shape-based property distribution was used for the heteroge-
neous phantom error study, where the stiffness distribution within each phantom was
defined by independent parameters representing the interface position, P , and stiffness
contrast, C, between the soft and hard silicone regions. For the stacked geometry,
the interface position, Ps , characterized the ratio of the hard/soft interface height to
the total phantom height. For the concentric geometry, the interface position, Pc , was
defined as the ratio of the hard inclusion diameter to the total phantom diameter. The
measured values of Ps and Pc for both silicone phantoms manufactured are shown in
Table 4.5.
(c)
(a) Stacked phantom motion capture. (b) Tracked reference point locations. (c)
(f)
(d) Concentric phantom motion cap- (e) Tracked reference point locations. (f)
ture.
Figure 4.15 The experimentally observed motion of the stacked and concentric heterogeneous
silicone phantoms.
Surface error sweeps in this study thus considered variations in both shape-based
parameters P and C. Performing FE simulations over a wide range of both parameters
allowed surface error values to be correlated to phantom material property distribu-
Table 4.5 The measured interface position parameter, P , for both heterogeneous silicone
phantom geometries.
Geometry P
Stacked Ps = 47%
Concentric Pc = 42%
tions. In line with previous silicone phantom error studies, the experimental evidence
for or against a particular property distribution was calculated over all reference points
using the surface error metric Ω3 from Equation 3.17.
4.3.3 Model Parameter Estimation
The heterogeneous study considered two fundamentally different approaches to esti-

mating the phantom stiffness distributions. These two methods are referred to as
corroborative and contradictive methods. Both methods are referenced to a null stiff-
ness distribution, representing the case where the phantom being tested is entirely
comprised of soft material.
The corroborative method assumes that this soft material modulus is known, and
seeks experimental evidence in the form of matching displacement data that confirms
the presence of any hard region. If the experimental evidence corroborates this pres-
ence, a region of hard material can subsequently be added with an estimated interface
position and hard stiffness described by P and C. Conversely, the contradictive method
assumes that the hard material modulus is known, and seeks to estimate a soft modulus
and interface position supported by the experimental evidence. Should the experimen-
tal evidence refute any distribution including a high stiffness region, the null result of
entirely soft stiffness can still be returned by setting the size of the hard region to zero,
P = 0%, and estimating the soft stiffness value alone using the parameter C.
Two stiffness contrast parameters, C, were used during the course of the study. For
the corroborative method, where the soft silicone storage modulus, Es0 , was assumed to
be known, the focus was on optimizing the stiffness contrast parameter C1 = Eh0 /Es0 ,
where Eh0 represents the hard silicone storage modulus. For the contradictive method,
where the hard modulus was assumed known, the contrast parameter was defined as
C2 = Es0 /Eh0 . The definition of these contrast parameters was consistent across both
geometries.
4.3.3.1 Corroborative Method
To test the ability of the corroborative method to estimate a heterogeneous phantom

stiffness distribution, a set of FE simulations were performed assuming a constant soft
silicone modulus value, previously estimated as Es0 = 28 kPa. For both models, the
interface position parameter, P , was varied from 0–100%, and the stiffness contrast
parameter, C1 , was varied to generate a range of hard modulus values in the range
Eh0 ≈ 80–200 kPa. The resulting values of the error metric Ω3 from the parameter
sweep performed on each geometry are shown in Figure 4.16.
0.152 Motion Error Metric, Ω3 (mm) 0.418 0.143 Motion Error Metric, Ω3 (mm) 0.504
100 100
Interface Position, P (%)
45
15
0 0
84 116 196 84 196
Hard Modulus, Eh0 (kPa) Hard Modulus, Eh0 (kPa)
(a) Stacked phantom. (b) Concentric phantom.
Figure 4.16 Parameter sweep results for the heterogeneous silicone phantoms, using the
corroborative method with Es0 = 28 kPa, where the location of minimum error on each plot is
indicated (◦).
The estimated stiffness contrast value in Figure 4.16(a) indicates a hard silicone
modulus of Eh0 = 116 kPa for the stacked geometry. The estimated interface position
for this geometry was within 3% (relative) of measured position of 47%. The location of
minimum error for the concentric model was along the edge of the parameter domain,
with neither Pc nor Eh0 successfully estimated. In this case, the parameter combination
leading to overall minimum motion error was not bounded by the range of parameters
simulated.
The sensitivity of the two-parameter corroborative method to variations in the as-
sumed soft modulus value was tested by modifying Es0 by ±15% to 24 kPa and 32 kPa,
and repeating the range of simulations. Results from these sweeps are shown in Fig-
ure 4.17. The estimated value for Ps for the stacked geometry varied by only 3%
(absolute) across the stiffness range, with both estimates of Ps within 10% (relative) of
the measured value. Stiffness contrast values corresponding to minimum Ω3 indicated
a hard storage modulus of Eh0 = 116–122 kPa for the stacked model.
Results from the additional concentric model corroborative sweeps had varied suc-
cess. When Es0 = 32 kPa, the error metric based approach again failed to identify the
100 100

46
43
0 0
82 116 197 83 122 198
(a) Stacked phantom, Es0 = 24 kPa. (b) Stacked phantom, Es0 = 32 kPa.
100 100
34
0 0
82 106 197 83 198
(c) Concentric phantom, Es0 = 24 kPa. (d) Concentric phantom, Es0 = 32 kPa.
Figure 4.17 Corroborative method sensitivity analysis results for the heterogeneous phantom
data, where the location of minimum error on each plot is indicated (◦).
presence and location of the hard inclusion, with minimum motion error corresponding
to an interface position of P = 0%. However, in the case where Es0 = 24 kPa, the
interface position was estimated to within 20% (relative) of the measured value, with
the hard modulus estimated at Eh0 = 106 kPa.
4.3.3.2 Contradictive Method
Parameter estimation using a contradictive method was tested by analyzing parameter

sweeps based on the previously estimated hard silicone modulus, Eh0 = 103 kPa. Stiff-
ness contrast values, C2 , were varied to cover a soft silicone storage modulus range of
approximately Es0 = 10–50 kPa, and the interface position, P , was varied from 0–100%.
The resulting motion error from these parameter sweeps is shown in Figure 4.18 for
both geometries considered.
100 100
47
35
0 0
10 25 49 10 24 49
Soft Modulus, Es0 (kPa) Soft Modulus, Es0 (kPa)
Figure 4.18 Parameter sweep results for the heterogeneous silicone phantoms, using the
contradictive method with Eh0 = 103 kPa, where the location of minimum error on each plot is
indicated (◦).
Stacked and concentric motion error plots indicate a soft silicone storage modulus
of Es0 = 24–25 kPa. The estimated interface position for the stacked geometry was an
exact match to the measured value of 47%, while the concentric geometry interface
position was estimated as 35%, which is within 20% (relative) of the measured value
of Pc = 42% from Table 4.5. The minimum motion error metric values using the
contradictive method were lower for both geometries than in the corroborative case,
indicating a closer match to measured displacements using this approach.
The sensitivity of the contradictive method to variations in the assumed hard
modulus value was tested by varying Eh0 by ±15%, to 88 kPa and 118 kPa, and repeating
all simulations. Figure 4.19 shows the result of these parameter sweeps. Here, results
are consistent with the Eh0 = 103 kPa case, with all simulations continuing to indicate
a soft modulus of Es0 = 24–25 kPa. The interface position for the stacked geometry was
estimated within 7% (relative) of the measured value in both cases, while the concentric
interface position was estimated to within 20% (relative) of the measured value.
100 100

50
46
0 0
11 25 48 12 24 47
(a) Stacked phantom, Eh0 = 88 kPa. (b) Stacked phantom, Eh0 = 118 kPa.
100 100
33 35
0 0
11 25 48 12 24 47
(c) Concentric phantom, Eh0 = 88 kPa. (d) Concentric phantom, Eh0 = 118 kPa.
Figure 4.19 Contradictive method sensitivity analysis results for the heterogeneous silicone
phantom data, where the location of minimum error on each plot is indicated (◦).
The best match to the experimental data for both geometries was obtained us-
ing the contradictive method with Eh0 = 118 kPa. Figure 4.20 is a comparison of the
simulated motion at the best fit parameter values when Eh0 = 118 kPa, and the mea-
sured motions, for both stacked and concentric geometries. The motion error metric Ω3
for the stacked and concentric geometries in these cases was 0.141 mm and 0.125 mm
respectively. These values represent an average orthogonal amplitude error per refer-
Ω3
ence point of 2 ≈ 0.06–0.07 mm, which less than 20% of the average surface motion
amplitude.
Inherent to the contradictive parameter estimation method is the assumption that
if the null stiffness is represented by an entirely soft phantom, the correct stiffness
distribution will be returned. To test this hypothesis, measured motion data from the
Measured Measured
74
Simulated Simulated
Point 1
Point
Height
(mm)
Point 15
0
{1 1 {1 1 {1 1 {1 1 {1 1 {1 1
x y z x y z
(a) Stacked phantom, simulated displacement at Eh0 = 118 kPa, Es0 = 24 kPa,
Ps = 46%.
Measured Measured
74 Simulated Simulated
Point 1
Point
Height
(mm)
Point 15
0
{1 1 {1 1 {1 1 {1 1 {1 1 {1 1
x y z x y z
(b) Concentric phantom, simulated displacement at Eh0 = 118 kPa, Es0 = 24 kPa,
Pc = 35%.
Figure 4.20 The best fitting FE-simulated cases compared with experimental motions for
both heterogeneous silicone phantoms.
soft homogeneous silicone phantom was used as reference motion data, and compared
to sweep displacements using the contradictive method, where Eh0 = 103 kPa. The
resulting error domains for the stacked and concentric cases are shown in Figure 4.21.
The estimated interface position value for both geometries is P = 0%, corresponding
to a homogenous phantom, where estimated Es0 values indicate a soft modulus of 27–
28 kPa.
100 100
0 0
10 27 49 10 28 49
Figure 4.21 Parameter sweep results from the homogenous soft silicone phantom, using the
contradictive method with Eh0 = 103 kPa, where the location of minimum error on each plot is
indicated (◦).
4.3.4 Discussion
The two parameters describing the stiffness distribution of the stacked geometry were
consistently estimated correctly using both corroborative and contradictive approaches.
Results from the concentric case showed the stiffness distribution can be consistently
estimated for this more challenging geometry only when adopting the contradictive
approach. The error plots in Figures 4.16 and 4.17 indicate that a corroborative ap-
proach to estimation of stiffness distribution is sensitive to any error in the assumed
value for Es0 . This sensitivity is more significant in the concentric case, where assumed
soft modulus values of Es0 = 28 kPa and Es0 = 32 kPa led to failed parameter estimation.
Because the a priori silicone stiffness values were estimated based on phantoms created
at a different time, this sensitivity to Es0 renders the corroborative method unsuitable
in a clinical sense.
Figures 4.18 and 4.19 provide evidence that parameter estimations using a con-
tradictive approach have very little sensitivity to variations in Eh0 , as expected from
the relatively long shear wavelength in the harder material. Estimated values for soft
modulus and interface position showed relative variation of less than 10% across both
geometries over a 30% range in the assumed value of Eh0 , indicating a sensitivity of ap-
proximately 33%, or a robust result. All six instances of the contradictive estimations
indicated a consistent soft modulus value of 24–25 kPa.
The estimated silicone stiffness values using both corroborative and contradictive
estimations did not agree with the previously estimated homogeneous silicone modulus
values. In comparison with E 0 = 28 kPa estimated from the homogeneous studies,
all contradictive estimations indicated a soft modulus value of Es0 = 24–25 kPa, and
the lowest motion error values using the corroborative method came when assuming
Es0 = 24 kPa. Similarly, the lowest error metric values when using the contradictive
method came when Eh0 = 118 kPa, and all corroborative estimations indicated a value
for Eh0 above the previously estimated value of E 0 = 103 kPa.
The range of values estimated for silicone stiffness indicates that the modulus
values for the soft and hard silicone varied between homogeneous and heterogeneous
phantoms, most likely due to the different batches of silicone used when preparing
these phantoms. Figure 4.22 presents a summary of the range of silicone modulus
values estimated and measured during the study, where the different colors indicate
batches of silicone prepared at different times. The clustering of results visible in this
graphic indicates that there is a degree of variability in the modulus value of the silicone
based in part on the experimental uncertainties when preparing the silicone material.
Additionally, the values from DMA testing were overestimated by up to 10% due to
the boundary condition effects discussed in Section 3.2.
Soft Modulus, E 0s (kPa)

20 25 30 35
24 25 27 28 29 32
24 25 28
24 25
Silicone Heterogeneous Homogenous DMA Testing

Batch Phantoms Phantoms Samples
116
103 106 116 122 135
100 110 120 130 140

Hard Modulus, E 0h (kPa)
Figure 4.22 The range of modulus values estimated and measured for both soft and hard
silicones, where color is used to differentiate batches of silicone prepared at different times.
In general terms, the error metric for the stacked geometry had a higher sensitiv-
ity to variations in the interface position parameter than for the concentric case, as
evidenced by highly ellipsoidal motion error contours, for example in Figure 4.16(a).
This heightened sensitivity is expected as the stacked geometry has reference motion
measurement located directly on the hard and soft silicone regions, allowing the in-
terface position to be accurately identified due to the step change in shear wavelength
across the two materials. The wavelength change between the lower and upper portions
of the stacked phantom is visible in Figure 4.20(a). The concentric geometry had a
significantly lower error sensitivity to interface position, and consequently less accurate
interface position estimates. The physical separation of the measurement locations
from the stiffness interface provides evidence as to why this loss of accuracy occurs.
The success of the contradictive estimation approach when applied to the soft ho-
mogenous silicone phantom data shown in Figure 4.21, confirmed that this method
was able to successfully return the null distribution when applied to a phantom with
constant stiffness. The estimated value for the soft silicone modulus across both geome-
tries of 27–28 kPa matched the corresponding value estimated from the soft phantom
in Section 4.2, while the interface position value of 0% confirmed the absence of any
hard inclusion in both cases.
The comparison between the best-fit displacements and the measured motion shown
in Figure 4.20 indicates that when measured displacements are large, the equivalent
simulated motion is of a smaller amplitude. This result is consistent with the homoge-
neous silicone phantom study, where the largest differences in amplitude were observed
to occur close to the actuation plate, where amplitudes are large. It is possible that
geometric non-linearity causes displacements to be under-estimated by FE simulation,
an effect that could be minimized by reducing the actuation amplitude.
Though the shape based approach to property distribution estimation involved

very simple geometry in this study, the underlying principles allow such a method to be
extended to more clinically relevant scenarios. In such cases, an increase in the number
of parameters would allow the identification of irregular shaped inclusions using a more
sophisticated geometry description. Such an increase in parameters would require non-
linear methods for parameter reconstruction, as the problem domain becomes too large
for exhaustive exploration using FE simulation. An extension to the heterogeneous
study where Es0 , Eh0 , and P are estimated independently may provide a more complete
method for reconstructing the elasticity distribution of the concentric phantom, which
returned less accurate interface position results in this experiment.
The heterogeneous surface error study performed demonstrated that the DIET
method can estimate the elasticity distribution of heterogeneous phantoms using a sim-
ple two parameter, shape-based approach with limited surface motion data as input.
Trials of both corroborative and contradictive techniques indicated that the contradic-
tive method is more robust in the presence of error in any a priori stiffness assumptions.
Adopting a contradictive approach to estimating elasticity distribution has the poten-
tial to reduce the number of parameters required to perform shape-based elasticity
reconstructions. In such a method, the stiffness of any inclusion is fixed, as the exact
stiffness used has very little effect on the observed surface motion. Such an approach
represents a potentially new avenue to inclusion imaging in elastography, as it is also
applicable to problems characterized by a larger number of parameters, allowing more
complicated geometries to be reconstructed with less computational effort.
Chapter 5
Non-Linear Reconstruction Algorithms
The studies performed in Chapter 4 provide important information about how experi-
mentally observed and FE-simulated phantom surface motions related to the internal
distribution of material properties. In addition, the error sweeps performed also al-
lowed estimates of the phantom material properties to be made using a limited number
of material parameters. Performing these error sweeps required an assumed a priori
knowledge of the material property distribution and computational resources sufficient
to perform FE simulations across the entire range of independent, unknown parameters.
While limited a priori knowledge may be available in a more realistic clinical sce-
nario, an exhaustive evaluation of the parameter space is not likely to be possible given
the time and computational constraints that would be applied to any clinical applica-
tion of a DIET system. Computational considerations will become more significant as
the number of unknown parameters is increased to evaluate more realistic geometries.
The work described in this chapter aimed to reconstruct the material property distribu-
tion of simulated and experimental phantoms using a range of non-linear reconstruction
algorithms.
Inverse problems in the field of elastography are highly non-linear, as the ob-
served displacement field of a solid body, and therefore any error function based on
this displacement, has a complex, high order differential relationship to the underlying
mechanical properties of the object. A range of strategies have been developed to solve
the elastic inverse problem, the majority of which apply an iterative approach based on
a gradient descent strategy to minimize the error between a set of model-simulated mea-
surements and corresponding experimental observations. Though novel reconstruction
techniques based on a stochastic approach, such as Genetic Algorithms (GAs), have
been applied to inverse problems where the FE method is used for system simulation,
the application of such stochastic methods to the tomographic elastography problem,
particularly when using only exterior surface motion measurements from experimental
phantoms, has not been considered. In this study, both deterministic and stochastic
102 CHAPTER 5 NON-LINEAR RECONSTRUCTION ALGORITHMS
classes of reconstruction algorithm are considered with regards to their suitability for
inclusion as a reconstruction algorithm in a DIET system.
The investigation of reconstruction algorithms in this chapter was performed con-
currently with the surface error studies presented in Chapter 4. Therefore, the inverse
algorithms were tested on a range of phantom data in conjunction with the development
of the experimental and FE simulation techniques. Surface error domains similar to
those used in Chapter 3 were often developed prior to the evaluation of reconstruction
algorithms in this work. This detailed error information provided an additional tool
used to allow a more thorough investigation of algorithm performance. However, it is
not required for practical, clinical implementation of any of the algorithms considered.
5.1 Gradient Descent Optimization
The minimization of a non-linear error function, Ω, using the gradient descent method,
is a deterministic process where an iterative approach is used to optimize a parameter
combination, θ. A gradient descent algorithm aims to determine a best fit parameter
combination by calculating the gradient of the error function at the current parame-
∂Ω
ter estimate, ∂θ , and updating the estimate by moving along this direction towards
decreasing error.
To perform gradient descent reconstruction, the gradient of the error function at
any parameter combination must be obtainable, requiring that the error function Ω(θ)
be continuous. This gradient calculation can be made using either analytic or finite-
difference methods. For the tomographic reconstruction problem, the error function Ω
must be calculated from available experimental and/or simulated data each iteration.
Figure 5.1 is a basic example of gradient descent reconstruction in a two-parameter
case, where optimal values for the independent parameters θ = (x, y) are θ̂ = (x̂, ŷ).
The colored contours on the plot define constant values of the error function Ω(θ).
Over a number of iterations, the algorithm’s estimates of x and y are updated towards
the combination that minimizes Ω.
Non-linear elastic property optimization aims to reconstruct any number of param-
eters, represented by θ, that describe the internal stiffness distribution of a model. Gra-
dient descent optimization in this case involves the minimization of an error function,
Ω, that compares a set of measured displacements, um , to FE calculated displacements
based on the current estimate of the model stiffness parameters, uc (θ). The range of
motion error metrics developed for the DIET system are described in Section 3.7, and
are represented here in general terms as a squared vector error,
Ω = kum − uc (θ)k2 . (5.1)

5.1 GRADIENT DESCENT OPTIMIZATION 103
y^
Start
0
0 x^ x
Figure 5.1 An example of a gradient descent algorithm, where successive iterations minimize
the value of a two-dimensional error function Ω(x, y).
The gradient of the error function with respect to θ will be zero at the parameter
∂Ω
combination corresponding to a minimum error value. Setting ∂θ = 0 generates a
non-linear system of equations,
µ ¶T ³ ´
∂Ω ∂uc
= −2 um − uc (θ) = 0, (5.2)
∂θ ∂θ
where the operator T denotes the matrix transpose. Equation 5.2 can be solved itera-
tively using the Gauss-Newton method. In this case, the formulation is written
µ ¶−1 µ ¶
∂2Ω ∂Ω
θr+1 = θr − δr (5.3)
∂θ2 ∂θ
for iteration r + 1, where δr is chosen to influence the step size of the algorithm.
∂2Ω
Development of the Hessian matrix term, H = ∂θ2
, leads to
µ ¶T µ ¶ µ ¶T ³ ´
∂2Ω ∂uc ∂uc ∂uc
=2 −2 um − uc (θ) , (5.4)
∂θ2 ∂θ ∂θ ∂θ
¡ ∂uc ¢T
where the term 2 ∂θ (um − uc (θ)) is assumed to be negligible, as um − uc (θ) is
small when θ is close to the actual parameter distribution of the medium being imaged.
The full iterative formulation for the problem can thus be written by expanding the
form of Equation 5.3,
Ã µ ¶ µ ¶ !−1 " µ ¶ #
∂uc T ∂uc ∂uc T m
θr+1 = θr − δr 2 + γI −2 (u − uc (θ)) , (5.5)
∂θ ∂θ ∂θ
where γ is a regularization term added to the Hessian matrix for numerical stability.
5.1.1 Reconstruction Algorithm Structure
Mechanical property reconstruction using a gradient descent algorithm was first per-
formed on the hard gelatine phantom. The surface error study performed on this
phantom in Section 4.1 indicated that the gelatine material used in the phantom had
a Young’s Modulus value of E = 27.1 kPa. This estimate was also corroborated by
independent mechanical testing.
When formulating the inverse problem for this study, a homogeneous, isotropic
case was assumed, where ν and ρ were assumed known at 0.49 and 1,000 kg/m3 respec-
tively. These assumptions limited the reconstruction to a single parameter representing
Young’s Modulus, E. The undamped material model used for FE simulation in Sec-
tion 4.1 was used for all displacement simulations. A flow chart showing the basic
structure of the gradient descent algorithm developed to perform the one parameter
reconstruction is shown in Figure 5.2, and is described in the following.
START
E0
FE Simulate to
generate uc (E )
Calculate Ω
Stopping
STOP
critera
FINISH
CONTINUE
E+
Calculate and H
Update E
Figure 5.2 A schematic of the gradient descent algorithm used for elastic reconstruction.
The reconstruction of E using the gradient descent algorithm began with an initial
guess for the stiffness parameter at iteration zero, E 0 . FE simulation using this value
and the constant values of ρ and ν previously assumed gave a vector of calculated
displacements at all nodes within the model, uc . The transformation from these nodal
amplitude values to the equivalent reference point displacements was performed accord-
ing to the point projection and damped fitting techniques described in Sections 3.5.2
and 3.6.3. The error between the FE-simulated displacement field and the measured
motion was calculated using the metric Ω2 previously defined in Equation 3.16, where
the damped simulated motion was transformed into an undamped approximation prior
∂Ω
to error calculation. Following the calculation of ∂θ , and the approximate Hessian ma-
trix, regularization was applied to the Hessian matrix diagonal based on the method
described by Marquardt [108].
An updated estimate for E was calculated using Equation 5.5, where the parameter
step size, δr , increased from 0.1–1.0 over the first ten iterations. The new stiffness value
was subsequently used to simulate a new displacement vector uc . Stopping criteria for
the gradient descent algorithm were based on the value of the residual term from
Equation 5.5,
µ ¶T
∂uc
R= (um − uc (θ)) . (5.6)
∂θ
In this study, the reconstruction algorithm was halted when R < 1 × 10−10 m4 N−1 ,
indicating a situation where the algorithm step size was sufficiently small to consider
the reconstruction as having converged. The output from GD reconstruction was the
stiffness value providing the closest match to the experimental data, E+ .
5.1.2 Stiffness Reconstruction Results
Stiffness reconstruction for the hard gelatine phantom using the gradient descent algo-
rithm was split into two stages. The first application of the reconstruction algorithm
was to a simulated data set, where FE-simulated motion at the previously estimated
phantom stiffness value of 27.1 kPa represented the experimentally measured data.
Testing the reconstruction algorithm using this noise free simulated data would con-
firm the ability of the algorithm to reach a global minimum in an ideal case. Following
the simulated data study, the same set of reconstructions were performed using the real
experimentally measured motion data from the gelatine phantom.
In both the simulated and experimental data cases, knowledge of the error domain
across the stiffness range of interest was obtained prior to reconstruction. While this
information was not required to perform the reconstructions, it allowed for a more
detailed analysis of the results in each case. The stiffness range selected for the re-
construction study was E = 20–30 kPa, where the surface error study from Section 4.1
was refined so that FE-simulated displacement data was available every 1 Pa over this
range.
5.1.2.1 Simulated Data
The error metric Ω2 resulting from comparison of the FE simulations between E = 20–
30 kPa to the simulated data set at E = 27.1 kPa is shown in Figure 5.3(a). Several
resonance stiffness values are apparent, where the value for the error metric is highly
discontinuous. As expected, the optimal value for Young’s Modulus was Ê = 27.1 kPa,
where Ω2 = 0 in the noise free case considered.
A total of 21 gradient descent reconstructions were performed on the simulated
data, with initial guesses E 0 at 500 Pa increments in the range 20–30 kPa. The perfor-
mance of these reconstructions over the first ten iterations is shown in Figure 5.3(b),
where each reconstruction progresses vertically downwards. The shaded upper part
of the plot represents the continuity of the error function over the range E = 20–
∂Ω
30 kPa using the gradient term ∂E , where areas of discontinuity appear as dark vertical
bands. Of all the reconstructions performed in the simulated case, a total of 7 failed to
indicate stiffness values within 5% of the target modulus. The remaining fourteen suc-
cessful reconstructions all converged to modulus values within 5% of the target value,
E = 27.1 kPa.
5.1.2.2 Experimental Data
The error metric Ω2 from comparison of the FE-simulated displacement database to

the experimental data from the hard gelatine phantom is shown in Figure 5.4(a). Initial
estimates of E 0 for the reconstructions were the same as in the simulated case, where a
total of 21 reconstructions were performed. The performance of these reconstructions
over the first ten iterations is shown in Figure 5.4(b), where the separate upper section
of the plot again represents the error gradient, indicating the continuity of the error
function over the stiffness range. Of the 21 initial guesses applied to reconstruction of
the experimentally measured data, 5 failed to converge to the correct solution. Similar
to the simulated data reconstructions, the remaining reconstructed modulus values
converged to within 5% of the target value.
5.1.3 Discussion
Application of a gradient descent reconstruction algorithm to the ideal simulated data

highlighted significant limitations in the method. In particular, one third of all re-
constructions failed to identify the single stiffness parameter correctly when starting
from initial guesses within 8 kPa of the target value. Analysis of Figure 5.3 provides
clear evidence as to why this issue occurs, with all failed reconstructions converging
2
0 Minimum Error, E = 27.1kPa
20 22 24 26 28 30
(a) Surface motion error metric variations with stiffness.
∂Ω
Motion Error Gradient, ∂E
Iteration
10
20 27.1 30
(b) Gradient descent reconstruction results over the first ten iterations,
where algorithm convergence is indicated (•).
Figure 5.3 Gradient descent reconstruction results when assuming that simulated motion at
E = 27.1 kPa represented an experimental data set.
2
0.375 Minimum Error, E = 27.1kPa
20 22 24 26 28 30
(a) Surface motion error metric variation with stiffness.
∂Ω
Motion Error Gradient, ∂E
Iteration
10
20 27.1 30
(b) Gradient descent reconstruction results, where algorithm convergence

is indicated (•).
Figure 5.4 Gradient descent reconstruction results from the real experimental data captured
from the hard gelatine phantom.
on stiffness values at or very near resonant values. These resonance bands are visi-
ble as peaks in Figure 5.3(a), and vertical lines of high error gradient in the upper
part of Figure 5.3(b). Despite these resonant bands being comparatively narrow, they
were sufficient to cause the reconstruction to prematurely converge when evaluating a
stiffness value close to a resonance peak.
An example of a local minimum caused by a resonant stiffness value is shown
in Figure 5.5. The clear global error minimum in Figure 5.3(a) led to all but one
successful reconstruction converging to the correct stiffness value within five iterations.
The remaining case showed clear preference for this stiffness value after ten iterations,
with its oscillation about the target stiffness value reducing over successive iterations.
0.36
Local Minimum, E = 25.281kPa

Global Error Minimum
0.24
25 25.6
Figure 5.5 An enlarged view of the error domain from Figure 5.3(a), detailing the local
minimum caused by a resonant stiffness.
Results from the application of the gradient descent algorithm to the experimen-
tally measured phantom data were very similar to the simulated case, with a slightly
higher proportion of reconstructions identifying the 27.1 kPa stiffness value correspond-
ing to minimum motion error. All failed reconstructions again converged to stiffness
values corresponding with areas of motion error discontinuity that created local min-
ima. The number of iterations required for convergence in the successful real data
reconstructions was greater than in the simulated data reconstructions due to the shal-
low gradient surrounding the global error minimum visible in Figure 5.4(a).
The high proportion of failed gradient descent reconstructions in this simple one
parameter case indicated that even highly localized error minima can cause significant
problems for a deterministic reconstruction algorithm based on local gradient calcula-
tions. Expanding the stiffness range investigated past the 20–30 kPa considered here
would almost certainly increase the number of local error minima, providing a signif-
icant challenge to reconstruction. While adding more sophistication to the gradient

descent algorithm to avoid small resonance minima may increase the success rate when
applied to the one-dimensional error domain in this study, the expansion of that domain
to include multiple parameters introduces additional local minima issues that gradient
descent methods would struggle to overcome even if resonance issues could be reduced
with a damped material model.
The initial gradient descent study indicated that an investigation into stochastic
reconstruction algorithms for the non-linear reconstruction of stiffness was potentially
warranted, particularly when considering the more realistic clinical case that will even-
tually involve multiple independent parameters. The degree of randomness inherent
to stochastic algorithms, such as GAs, presents a possible method for overcoming the
limitations of the gradient descent approach when dealing with local minima.
In addition, a hybrid combination of deterministic and stochastic reconstruction
algorithms for solving the elastic inverse problem also warrants investigation based on
these results. When determining the effectiveness and suitability of such alternative
methods for use in a DIET system, a comparison to a traditional gradient descent
approach should be considered. Performance differences would be measured by the
balance of parameter reconstruction effectiveness and computational cost.
5.2 Hybrid Reconstruction Simulation Study
The use of stochastic algorithms for non-linear optimization has increased with the
availability of fast and affordable computer hardware that can offset the large compu-
tational costs associated with such methods. Stochastic reconstruction methods such as
a GAs vary in their application based on simulation techniques, the type and quantity
of measured data, and the method used to encode and/or parameterize the optimization
problem. Hence, this class of algorithms are highly application specific.
A first study using stochastic algorithms for parameter reconstruction in a DIET
system was performed using FE-simulated data to represent experimentally measured
motion in a controlled fashion. The intent of this study was to create and evaluate a
stochastic algorithm using reference data free from unknown and highly variable exper-
imental error, before performing reconstructions on real phantom data. The algorithm
was intended to be a hybrid approach combining both traditional Gradient Descent
(GD) methods with novel Combinatorial Optimization (CO) techniques. As part of
the algorithm development, the three non-linear reconstruction methods summarized
in Table 5.1 were evaluated for their performance when applied to the elastographic
inverse problem.
The two shape-based parameters P and C introduced in Section 4.3.2 were used to
describe the internal stiffness distribution of the simple heterogeneous model geometries
5.2 HYBRID RECONSTRUCTION SIMULATION STUDY 111
Table 5.1 The reconstruction algorithms considered during development of the hybrid recon-
struction method.
Algorithm Identifier Class

Gradient Descent GD Deterministic
Combinatorial Optimization CO Stochastic
Hybrid Algorithm Hybrid Combination
considered in this study, where both geometries are shown in Figure 5.6. The interface
position parameter Ps represented the ratio of the hard/soft interface height to the
total model height for the stacked geometry, and Pc the ratio of the hard inclusion
diameter to the total diameter in the concentric case. The stiffness contrast parameter
C was defined for both geometries as Eh /Es , representing a corroborative approach
according to the definition in Section 4.3.3.
E 0s = 20kPa E 0h = C.E 0s E 0s = 20kPa E 0h = C.E 0s
(a) Stacked model geometry. (b) Concentric model geometry.
Figure 5.6 The geometry of both phantom models used in the hybrid algorithm simulation
studies.
Surface motion data sets required for the study were generated by FE simulation, us-
ing a mesh containing approximately 169,000 linear tetrahedral elements. The model
size of 75 mm diameter and 75 mm height was chosen as a close approximation to the
silicone phantoms that were prepared subsequent to this work. The bottom face of
the cylinder was constrained in the horizontal plane and actuated vertically with an
amplitude of 1 mm at a frequency of 50 Hz. Both geometries were simulated using the
same mesh, with Young’s Modulus values assigned nodally based on location relative
to the specified values of P and C. An undamped material model was used to simplify
the reconstruction problem by reducing the number of parameters required for simu-
lation. While the results from the gelatine phantom gradient descent reconstructions
in Section 5.1 indicated an undamped model would not give a close match to exper-
imental phantom displacements, such a model was sufficient in this case for testing
reconstruction algorithms where all data was simulated.
The Young’s Modulus value for the soft material, Es , was fixed at 20 kPa, chosen
to be of the same order as Young’s Modulus measurements of fat and glandular tissue
in the human breast, and the soft phantom stiffness values estimated in Chapter 4.
Poisson’s ratio, ν, and density, ρ, were fixed at 0.45 and 1000 kg/m3 respectively, ap-
proximating the nearly incompressible nature and high water content of human tissue.
The target parameters P̂ and Ĉ that represented the real data set for each ge-
ometry are shown in Table 5.2. Also summarized are the relevant material properties
required for FE simulation of the experimental motion. All 3,215 nodes on the ver-
tical exterior of the cylinder were chosen as surface motion reference points for the
reconstruction problem. This sampling method provided a simple approximation to
a more advanced DIET system, where it is hoped motion data from closely spaced
points across the entire surface of the breast will be available [109]. Noise added to
the two simulated experimental data sets was normally distributed with a mean of zero
and a variance of 10% of the average reference point displacement amplitude, giving a
controlled approximation of experimental noise.
Table 5.2 The parameter values representing real data for the reconstructions performed
using simulated data.
Geometry
Model Properties Stacked Concentric
Shape Parameters P̂ 40% 50%

Ĉ 400% 300%
Modulus Values Es 20 kPa 20 kPa
Eh 80 kPa 60 kPa
Material Properties ν 0.45 0.45
ρ 1,000 kg/m3 1,000 kg/m3
In addition to the target data set, FE simulation was also used to generate a mo-
tion database for both geometries across a range of values for P and C. Simulations
were performed on both stacked and concentric geometries at interface positions P
ranging from 10% to 90% in increments of 1%, and stiffness contrasts C from 100%
to 500% in increments of 5%. These parameter limits defined the problem domain for
the reconstruction algorithms, and encompassed a total of 6,561 FE simulations for
each geometry. This database of simulations allowed rapid testing of the reconstruc-
tion algorithms as well as evaluation of the surface error appearance over the problem
domain.
5.2.2 Reconstruction Algorithms
The three non-linear reconstruction algorithms considered all aimed to minimize the
squared error between a set of reference displacements, um , simulated at the target
parameter values P̂ and Ĉ, and the corresponding simulated displacements, uc , calcu-
lated using the current estimates of parameters P and C. The error metric that formed
the cost function for all reconstructions in this study was a representation of the mean
squared motion error across all reference points, and was defined
s
PN m
i=1 kui − uci k2
Ω4 = , (5.7)
3N
where N is the number of reference points. The minimum value of the error metric Ω4
occurred when simulations were performed at P̂ and Ĉ for a particular geometry, and
was defined as Ω̂, representing the lowest possible error metric value any reconstruction
could attain.
Figure 5.7 shows the error metric from Equation 5.7 across the entire problem
domain for both geometries, where the ‘measured’ displacements were those simulated
at the target parameters from Table 5.2, with the addition of noise. The minimum
error value Ω̂ for the stacked geometry was 0.39 mm, and for the concentric geometry
0.24 mm. Indicated on both plots is the target parameter combination corresponding
to parameter values P̂ and Ĉ. The error information in these plots provided both data
and context to the reconstruction algorithms subsequently tested. The three non-linear
reconstruction algorithms considered in this study are described in this section in the
order in which they were evaluated.
5.2.2.1 Gradient Descent Reconstruction
The GD reconstruction algorithm used in the simulation study was based on the one
parameter algorithm developed in Section 5.1. For this study, the algorithm was modi-
fied to reconstruct the two independent parameters P and C using the error metric Ω4
defined in Equation 5.7. For the relatively small number of reconstruction parameters
used in this study, the Jacobian matrix, which traditionally relies on parameter val-
ues being nodally distributed, is most efficiently calculated through a finite-difference
approximation in which terms take the form
90 90
Interface Position, Ps (%)
Interface Position, Pc (%)

50
40
10 10
100 400 500 100 300 500
Contrast, C (%) Contrast, C (%)
(a) Stacked geometry. (b) Concentric geometry.
Figure 5.7 The displacement error metric from Equation 5.7 evaluated across the parameter
domain for both heterogeneous silicone geometries, where the value of Ω4 has an upper limit
of 5 mm, and the location of minimum error, Ω̂, is indicated (◦).
uc (θ + ∆θ) − uc (θ)
J= , (5.8)
∆θ
where θ represents either P or C. As previously described, inversion of the approximate

Hessian matrix, H = J T J, allows the calculation of updates applied to both parameters.
Gradient descent reconstructions were performed with a range of initial parameter

guesses covering the entire problem domain. Initial parameter guesses were generated
for P every 2% between 12% and 88%, and for C every 10% between 110% and 490%,
resulting in a total of 1,521 reconstructions for each geometry. A GD reconstruction
was stopped when any of the following criteria were satisfied:
1. The surface motion error metric Ω4 was within 1% of the value for Ω̂ for the
geometry considered.
2. The calculated algorithm step size was zero for both P and C parameters for
more than 2 consecutive iterations.
3. A total of 100 GD iterations had been performed.
The first criteria represented a definite successful reconstruction. The second was
a case where the gradient was either so shallow the algorithm halted, or so steep that
the reconstruction continually attempted to step outside the problem domain, and the
update was ignored. The third case covers a failure to converge within the allocated
number of iterations.
5.2.2.2 Combinatorial Optimization
CO involves efficiently exploring problem search domains to solve large and complex
non-linear optimization problems faster than an exhaustive search [84]. The optimiza-
tion algorithm used in this study combined features of two popular CO techniques –
Genetic Algorithms (GAs) and Simulated Annealing (SA).
The CO algorithm shown in Figure 5.8, and described in the following, contains
parameter pairs, Gji , defined as genes, where the subscript refers to the ith gene in
the population, 1 ≤ i ≤ NG , and the superscript, j, indicates the gene’s generation,
where 0 ≤ j ≤ M . Error metric values, Ωji , are defined using the same convention.
Values for both NG and M can be chosen arbitrarily, with an upper limit usually
reflecting increased computational cost. In the CO algorithm used for this study, a
total of ten genes were used to define the population, NG = 10, and a maximum of
100 generations was specified, M = 100. Within an entire population of genes at a
particular generation, termed Gj , the gene providing the closest match to experimental
data is described as Gj+ , and has corresponding error metric value, Ωj+ .
CO begins with a population of NG genes, each containing an initial value for the
two parameters, G0i = {P, C}0i . Initial values for all parameters within all genes are
usually randomly assigned within a range based on a priori knowledge. Before entering
the CO algorithm proper, FE simulations are performed to obtain starting values of
the error metric, Ω0i , for all genes in the population.
Mating in CO involves interchanging genetic material between genes, with the in-
tent of improving the fitness of the population as a whole [82]. Pairs of genes are
identified for mating using roulette wheel selection, where the probability of gene se-
lection is directly proportional to the value of its error metric, Ωji [110]. Genes selected
as one of a mating pair have their constituent parameter values modified by swapping
a proportion of their binary expression with bits from the corresponding parameter
in their mating partner. The chance of the mating operation being performed is set
with the probability value Πmate , where for this study Πmate = 0.9. Parameter values
output from the mating process form a new, temporary gene pool.
Random mutation modifies the value of any parameter within any gene by flipping
the value of a bit from the parameter’s binary expression. This process increases the
diversity of the population by introducing new genetic material that is unlikely to be
introduced through mating alone. Mutation is potentially performed on every bit within
the temporary population after mating, where the likelihood of a bit changing value
is governed by the probability Πmutate . In this study Πmutate = 0.02. The resulting
population is defined as G̃ji , where the tilde indicates that the mating and mutation
operations have taken place. FE simulation at the parameter values provided by these
potential genes allows corresponding error metric values, Ω̃ji , to be calculated.
START
G 10 = {P, C }01 , G 20 = {P, C }02 , ... GN0 = {P, C }N0
Perform mating Perform mutation

across all G on all P, C
FE simulate at FE simulate at FE simulate at

{P, C }1 {P, C }2 {P, C }N
Calculate Ω1 Calculate Ω2 Calculate ΩN
Screen potential offspring

Find G+ leading to
using Simulated
minimum error, Ω+
Annealing (SA)
CONTINUE
Stopping FINISH
STOP
critera G+ = {P, C }+
Figure 5.8 A schematic of the CO reconstruction algorithm.
SA takes its theoretical approach from the statistical mechanics of physical an-
nealing, where solutions that decrease the global system ‘energy’, here the error term
Ω, are increasingly preferred as a system-wide temperature variable is decreased over
time [87]. In the CO algorithm described here, SA provides a mechanism for determin-
ing which genes from the potential offspring population, G̃j , will become part of the
subsequent generation, Gj+1 . Error values from each potential gene, Ω̃ji , are compared
with the corresponding value from the current population, Ωji . The chance of the off-
spring gene, G̃ji , replacing the original gene, Gji , in the subsequent population is set by
the probability i
SA Πj , defined
Ã !
j −Ω̃ji /Ωji
SA Πi = exp , (5.9)
Tj
where T j is the value of the system-wide temperature variable at generation j. De-

creasing T each generation means potential genes increasing the population error have
a decreasing chance of being selected as generations progress. In this experiment, the
temperature was decreased logarithmically between values of 10 and 0.5 over a total
of 100 generations. After completing SA screening, parameter values from all genes
chosen for the subsequent generation are logged, along with their error values, and the
gene displaying the best match to the experimental data, Gj+ , noted.
The CO algorithm in this study had two stopping criteria:
1. The error metric from the best-fitting member of the current population, Ωj+ , was
less than a pre-determined minimum based on the value of Ω̂ for the geometry.
This criterion represented a successful reconstruction.
2. The maximum number of generations for the reconstruction, M , was reached.

This criterion represented a failed reconstruction.
Upon halting the CO algorithm, the parameter values providing the closest match
to the measured motion, G+ = {P, C}+ , with corresponding error metric, Ω+ , were
output.
5.2.2.3 Hybrid Algorithm
A hybrid algorithm for non-linear reconstruction of shape-based parameters was identi-

fied as a possible solution to the DIET elastography problem in Section 5.1. The simple
hybrid algorithm investigated in this study involved first running a CO algorithm for a
total of 100 generations. The second stage of the algorithm was a GD reconstruction,
where the initial guess for the algorithm was taken from the output of the CO recon-
struction, G+ = {P+ , C+ }. This approach was intended to test the potential of such
a method to avoid the issues with local minima previously identified when using GD
alone to perform parameter reconstruction.
5.2.3 Parameter Reconstruction Results
The overall success of each type of reconstruction algorithm across all starting values
for a particular geometry was represented by a cumulative, normalized parameter error
value, defined
ÃP PN !
N
i=1 |P̂ − (P+ )i | i=1 |Ĉ − (C+ )i |
Ψ1 = 1 − + /2N, (5.10)
P̂ Ĉ
where P̂ and Ĉ are the target parameters from Table 5.2, and N is the total number of
reconstructions performed on the geometry in question. Equation 5.10 returns a value
between 0–1, expressed as a percentage, where 100% represents all reconstructions

reaching the global optimal solution, where (P+ )i = P̂ , and (C+ )i = Ĉ, for all i.
Hence, the metric provides an easy to understand measure of how close a particular set
of reconstruction results are to the target parameter combination.
The distribution of the 1,521 GD reconstructed solutions is shown in Figure 5.9.

The success metric for the GD algorithm when applied to the stacked and concentric
geometries was Ψ1 = 71.2% and Ψ1 = 78.2%, respectively. In both cases, the majority
of failed reconstructions were those that attempted to step to parameter values outside
the problem domain. These reconstructions can be seen at lower left of Figure 5.9(a),
and upper left of Figure 5.9(b), where the imposed limits on parameter values direct
the algorithm to a corner of the domain.
<1% Percent of Reconstructed Solutions (N = 1521) 45% <1% Percent of Reconstructed Solutions (N = 1521) 34%
90 90
50
40
10 10
100 400 500 100 300 500
Stiffness Contrast, C (%) Stiffness Contrast, C (%)
Figure 5.9 The distribution of results from 1,521 GD reconstructions performed on both het-
erogeneous geometries, where the intersection of the dashed lines indicates the target parameter
combination.
To allow statistical comparison with the GD algorithm, the same number of CO

reconstructions were performed on each geometry, beginning with random parameter
guesses within the problem domain. Figure 5.10 shows the distribution of these 1521
reconstructed solutions for both geometries considered. The success metric Ψ1 for the
stacked and concentric geometries when using CO was 97.8% and 91.2%, respectively.
Here the reconstructed solutions are observed to be largely confined to a smaller area
or band adjacent to the target parameter combination.
The final reconstruction testing performed the same 1,521 instances using the hy-
brid reconstruction method, where the results from the previously performed CO were
used as the first phase of the algorithm. The reconstructed solution density plots for
both geometries are given in Figure 5.11, where success metrics for the stacked and
concentric geometries were Ψ1 = 98.9% and Ψ1 = 94.2%, respectively. A clear re-
duction in solution spread is apparent when comparing Figure 5.11 with Figures 5.9
90 90

50
40
10 10
100 400 500 100 300 500
Figure 5.10 The distribution of results from 1,521 CO reconstructions performed on both
heterogeneous geometries, where the intersection of the dashed lines indicates the target pa-
rameter combination.
and 5.10. The success metric Ψ1 for all three reconstruction algorithms across both
heterogeneous phantom geometries considered is summarized in Table 5.3.
90 90
50
40
10 10
100 400 500 100 300 500
Figure 5.11 The distribution of results from 1,521 instances of the hybrid algorithm per-
formed on both heterogeneous geometries, where the intersection of the dashed lines indicates
the target parameter combination.
5.2.4 Discussion
The plots in Figure 5.7 provide a clear indication of the error metric trends across the
problem domain for each geometry, where the large areas of resonance observed are a
result of the undamped linear model used for FE simulations. The stacked model had
a more complicated resonance pattern, as the geometry places the interface between
hard and soft material perpendicular to the mechanical waves generated by actuation
of the model.
Table 5.3 The reconstruction algorithm success metric, Ψ1 , across both stacked and concen-
tric geometries.
Geometry
Algorithm Stacked Concentric
GD 71.2% 78.2%
CO 97.8% 91.2%
Hybrid 98.9% 94.2%
While the areas of resonance indicate simulated displacements becoming unrealis-

tically large, they also provide an challenging problem domain with clear local minima,
ideal for reconstruction algorithm evaluation. The large area of low error surrounding
the target parameter combination for the concentric geometry was due to the low sen-
sitivity of Ω4 to P or C in this region, and makes accurate elastographic reconstruction
of this geometry especially challenging.
The large spread of solutions for the GD reconstructions shown in Figure 5.9 indi-
cate a high probability that a randomly-selected starting position within the parameter
domain will not lead to a successful reconstruction. Analysis of failed reconstructions
showed that reconstructions initiated on the opposite side of a resonance band from
the target solution were not able to reach the correct solution. This result confirms
the behavior identified in Section 5.1 when reconstructing in a one parameter scenario,
where a GD method converges prematurely at a local minimum. However, the GD
solver performed very well when initiated within the vicinity of the global solution,
supporting its inclusion in a hybrid algorithm with a good initial value, given the small
number of FE simulations required for convergence.
The results in Figure 5.10 demonstrate that CO has a greater ability to avoid local
minima and reach solutions surrounding the globally optimal parameter combination.
However, the CO approach had a limited ability to differentiate between the globally
optimum solution and one nearby, resulting in a spread of solutions at low error metric
values. This limitation constrained the effectiveness of the algorithm when applied
to parameter domains where the error metric has low sensitivity to the parameters
used, such as the concentric geometry in this case. However, a qualitative comparison
of Figures 5.9(b) and 5.10(b) clearly shows an obvious clustering about the correct
solution for the CO reconstruction that is reinforced by the significant increase in the
Ψ1 metric values in Table 5.3.
The hybrid algorithm was clearly the most accurate and consistent of the three
approaches considered, with success metrics above 94% for both geometries. The re-
duction in solution spread visible in Figure 5.11 demonstrates the potential for a re-
construction algorithm combining the advantageous aspects of GD and CO to have

markedly greater success than either algorithm in isolation. Reconstructed solutions
from both CO and hybrid algorithms were, in general terms, clustered about the min-
imum error solution, potentially allowing the use of statistical methods to characterize
not only reconstructed parameter values, but also confidence metrics for these recon-
structions.
The number of forward FE simulations required for each reconstruction algorithm
is shown in Table 5.4, and provides an estimate of the computational cost of each
method given the serial computation approach used in this study. While the CO and
hybrid methods require a significantly higher numbers of simulations, a large number of
these simulations are independent, and could be performed in parallel across multiple
processors. For the CO algorithm, which also formed the first 100 iterations of the
hybrid method, this would reduce the time taken to complete FE simulations by a
factor of ten, as each generation contained ten independent chromosomes. The finite
difference approach to Jacobian calculation in the GD algorithm was also able to be
parallelized, but to a lesser extent given the limited number independent parameters
in this study. However, more advanced gradient calculation methods, such as the
adjoint method, have the potential to reduce or eliminate this cost scaling for the GD
algorithm as the number of parameters is increased [111]. The theoretical reduction in
the number of consecutive FE simulations required if performing reconstructions in a
parallel computing environment is also quantified in Table 5.4.
Table 5.4 The total number of FE simulations required for each reconstruction algorithm in
the two-parameter simulated case.
FE Solutions GD CO Hybrid
Serial system (actual) 60 1000 1060
Parallel system (theoretical) 20 100 120
Improvements in the sophistication of both the GD and CO algorithms should

lead to improved reconstruction success rates. Both algorithms were very simple im-
plementations of these techniques, and could possibly benefit from added features, such
as GD regularization, and more sophisticated CO evolution strategies. Efficiency im-
provements in the hybrid algorithm could be made by an investigation into the best
combination of constituent CO and GD algorithms, where the potential exists for the
crossover point between the two algorithms to be made dynamically as the reconstruc-
tion progresses.
The hybrid algorithm simulation study demonstrated that shape-based reconstruc-
tion was able to identify both a positional and stiffness contrast parameter for varying
physical geometries in the presence of noise. Of the three reconstruction algorithms

tested, a hybrid algorithm combining aspects of GD and CO was clearly the most suc-
cessful at avoiding local minima and reaching the global optimal solution. Significant
potential exists to refine and improve this form of algorithm when applying such a
method to experimental data sets, which are the next logical target for testing.
5.3 Hybrid Reconstruction Optimization
The promising results obtained from noisy simulated data in Section 5.2 indicated the
hybrid reconstruction method had significant potential for use in a DIET system. The
subsequent study described in this section focuses on optimizing the performance of
the hybrid algorithm as applied to the concentric silicone phantom data presented in
Section 4.3. A sensitivity analysis is performed in order to determine the influence of
different algorithm settings on the reconstruction performance, with regards both to
solution accuracy and computational efficiency. In addition, a statistical measure of re-
construction certainty is developed to allow the combination of multiple reconstructions
into a single result.
Estimated values for the material properties of the hard and soft silicone material
were provided by the surface error studies in Section 4.2. The two-parameter sweeps
performed on the heterogeneous silicone phantoms in Section 4.3.3 using both corrob-
orative and contradictive a priori stiffness assumptions indicated that surface error
metric values had high sensitivity to any assumed soft modulus value. This study
considered independent variation in three parameters for the concentric geometry, rep-
resenting soft modulus, Es0 , hard modulus, Eh0 , and interface position, Pc . The use of
three parameters provided a larger and more complicated error domain likely to contain
multiple local minima.
5.3.1 Sensitivity Analysis
The hybrid reconstruction algorithm introduced in Section 5.2 was modified for this
study to allow reconstruction of three independent parameters. A block diagram of
the three parameter hybrid algorithm is shown in Figure 5.12, where the optimization
involves the three example parameters A, B, and C. All reconstructions performed
during this sensitivity study used the experimentally measured concentric motion data
as the reference displacement set, and calculated the surface error metric Ω3 according
to Equation 3.17.
The hybrid reconstruction algorithm sensitivity analysis and optimization was per-
formed by varying several key settings controlling both CO and GD stages. As both
phases were performed separately in the hybrid method, this optimization was per-
formed on each component individually, before combining both parts. As this form of
5.3 HYBRID RECONSTRUCTION OPTIMIZATION 123
START START
G 10 = {A, B, C }10 , G 20 = {A, B, C }02, ... G N0 = {A, B, C }N0 {A, B, C }
FE Simulate at
Perform mating Perform mutation {A, B, C }
across all G on all A, B, C Calculate Ω
HYBRID ALG ORITHM

Stopping
STOP
critera
FE simulate at FE simulate at FE simulate at
{A, B, C }1 {A, B, C }2 {A, B, C }N
FINISH
Calculate Ω1 Calculate Ω2 Calculate ΩN CONTINUE
{A, B, C }+
FE Simulate at FE Simulate at FE Simulate at

Screen potential offspring
Find G+ leading to {A+ ∆A, B, C } {A, B+ ∆ B, C } {A, B, C+ ∆C }
using Simulated
minimum error, Ω+
Annealing (SA)
CONTINUE
Stopping FINISH Calculate J, H

STOP
critera G+ = {A, B, C }+ Update {A, B, C }
(a) Combinatorial Optimization algorithm (b) Gradient descent algorithm
Figure 5.12 A schematic of the hybrid reconstruction algorithm.
reconstruction algorithm is highly application specific, the optimized algorithm settings

chosen would represent the ideal case for the specific shape-based parameter reconstruc-
tion considered. To avoid initial bias in the CO sensitivity study, all reconstructions
started from the same initial population set.
5.3.1.1 Combinatorial Optimization
The investigation of algorithm settings for the CO optimization involved the results of
100 reconstructions, each based on a different initial population, l G0 , and generating
output l G+ , where 1 ≤ l ≤ 100. These results were normalized to give a measure of
reconstruction performance between 1.0 and 0.0, where 1.0 represented the mean error
at generation zero,
1 X ³l 0 ´
100
Ω̄0+ = Ω+ , (5.11)
100
l=1
and 0.0 indicated the minimum error value within the domain, Ω̂. To allow performance
characterization, the best-fit parameter values leading to Ω̂, termed {Â, B̂, Ĉ}, were
determined prior to the study.
The first phase of CO optimization varied the value of Πmate from 0.2 to 0.8 in
increments of 0.2. For these reconstructions, Πmutate was set to zero to eliminate the
influence of random mutation on the reconstruction results, and the number of genes
was fixed, NG = 10. The resulting CO algorithm performance across 100 generations is
shown in Figure 5.13(a), where each line represents the mean of 100 individual recon-
structions performed at a combination of settings. Optimal reconstruction performance
was seen when Πmate = 0.6, where the normalized algorithm performance value was
0.18 after 100 generations.
CO sensitivity to mutation was tested by fixing Πmate at the optimized value of
0.6, and varying Πmutate from 0.0 to 0.2. The resulting normalized CO algorithm
performance across 100 generations is shown in Figure 5.13(b). In this plot, the solid
line corresponds to the Πmate = 0.6 result in Figure 5.13(a). No significant improvement
in performance was noted at Πmutate values greater than 0.1, where the normalized
performance value was 0.05 after 100 generations.
Finally, the number of genes in the population, NG , was varied, with Πmate = 0.6,
and Πmutate = 0.1. CO reconstructions containing 5, 10, 20 and 40 genes were per-
formed, with results shown in Figure 5.13(c). In this plot the starting values for the
performance metric varied in each case, as the different values for NG led to a range
of values for Ω̄0+ . To normalize for the increased computational cost of updating ad-
ditional genes through each generation, the horizontal axis of this figure indicates the
total number of FE simulations performed. Regardless of this additional normaliza-
tion, adding genes to the population continued to improve algorithm performance. A
value of NG = 10 was selected as being optimal for this problem, given the physical
number of computer workstations available, and the absence of significant performance
improvement between 10 and 20 genes. The optimized CO algorithm settings used for
subsequent parameter reconstruction are summarized in Table 5.5.
Table 5.5 Optimal settings for the CO algorithm determined via sensitivity analysis.
Setting Value
Πmate 0.6
Πmutate 0.1
NG 10
5.3.1.2 Gradient Descent Optimization
Experiments with the simple GD algorithm introduced in Section 5.1 indicated that
when reconstructing the limited number of parameters in this study, regularization of
the approximate Hessian matrix was not required for numerical stability. This inherent
1.00
Normalized Performance
0.2
0.4
Πmate
0.6
0.8
0.18
0 25 50 75 100
Generation
(a) CO algorithm sensitivity to the probability of mating.
1.00
0.0
0.05
Πmutate
0.1
0.2
0.05
0 25 50 75 100
Generation
(b) CO algorithm sensitivity to the probability of mutation.
1.00
5
10
NG
20
40
0.01
0 100 200 300 400 500

Number of FE Simulations
(c) CO algorithm sensitivity to the number of genes in the popu-

lation.
Figure 5.13 Results from the sensitivity analysis performed on the CO reconstruction algo-
rithm.
1.00
0.66
0 5 10 15 20
Iteration
Figure 5.14 The performance of the GD reconstruction algorithm applied to the concentric
phantom experimental data.
stability limited the GD algorithm optimization study to the total number of iterations
performed. Figure 5.14 shows the normalized performance result from 100 GD recon-
structions starting at random parameter values within the domain. The performance
metric from all reconstructions was tracked over the total of 20 iterations performed,
with no improvement in performance observed beyond 10 iterations. The absence of
further improvement in the GD results beyond 10 iterations indicated that all instances
of the algorithm had converged prior to this point, where the normalized performance
metric was 0.66. The high minimum value of the performance metric in Figure 5.14
indicated that the majority of GD reconstructions converged to a parameter combina-
tion representing a local error minimum that was not close to the globally optimum
solution.
5.3.2 Heterogeneous Phantom Reconstruction Results
After optimizing the performance of the CO and GD components of the hybrid algo-
rithm, reconstructions were performed on the experimental phantom data using GD
and CO separately, followed by the hybrid algorithm that combined both techniques.
To evaluate each trial method performed on the experimental data, a total of 100
separate reconstructions were performed. Referring to Figure 5.12, the mean final re-
constructed value for each parameter over all 100 reconstructions, Ā+ , B̄+ , C̄+ , was
compared to the best-fit value for that parameter, Â, B̂, Ĉ, using two metrics. The Ψ2
metric acts as a measure of the mean normalized parameter error,
|Â − Ā+ | |B̂ − B̄+ | |Ĉ − C̄+ |

Ψ2 = + + . (5.12)
|Â| |B̂| |Ĉ|
Additionally, Σ2 was used as a measure of the normalized variance of the reconstructed

solutions, defined as
³ ´2 ³ ´2 ³ ´2
NX
=100 l Ā − Â l B̄ − B̂ l C̄ − Ĉ
+ + +
Σ2 = + + . (5.13)
l=1
(N − 1)Â2 (N − 1)B̂ 2 (N − 1)Ĉ 2
In the following results, the example parameter values {A, B, C} are represented by
the shape-based parameters relevant to this experiment, {Es0 , Eh0 , Pc }.
Figure 5.15 is the resulting stiffness solution distribution after performing 100 GD
reconstructions on the concentric data. Each reconstruction had random starting val-
ues and was run for a total of 10 iterations. Reconstruction error metrics Ψ2 and
Σ2 were 0.79 and 0.55 respectively. The finite-difference method used for the Jaco-
bian matrix calculation meant that a total of 40 FE simulations were required for
each reconstruction. The best-fit stiffness profile on the plot represents the parameter
values corresponding to Ω̂, while the measured line is based on the mechanical test-
ing performed in Section 3.2, and the measured interface positions previously given in
Table 4.5.
135 Ē 0 + σE 0
Measured
Ē 0
100
Best-fit
Ē 0 − σE 0
32
24
0 13 16 37
Radial Position (mm)
Figure 5.15 The reconstructed stiffness profile distribution from 100 GD reconstructions on
concentric phantom data, where the measured stiffness profile, and best-fit profile based on Ω̂,
are indicated.
Figure 5.16 shows the result of the optimized CO algorithm applied to the con-
centric phantom displacements, where results at a range of generations are presented.
Figure 5.17 shows results from the hybrid algorithm applied to the same data, where
optimized GD reconstruction was performed following CO, using a starting guess for
the parameters provided by the best result from the CO reconstructions, G+ . Ta-
ble 5.6 summarizes the performance metric results across the optimized reconstruction
algorithms considered, where the values of Ψ2 and Σ2 for the hybrid algorithm are
consistently less than the corresponding CO or GD algorithms.
Storage Modulus, E 0 (kPa) 135 Ē 0 + σE 0
Measured
Ē 0
100
Best-fit
Ē 0 − σE 0
32
24
0 13 16 37
(a) CO algorithm, 1 generation.
135
100
32
24
0 13 16 37
(b) CO algorithm, 10 generations.
135
100
32
24
0 13 16 37
(c) CO algorithm, 100 generations.
Figure 5.16 The stiffness profile distribution from 100 CO reconstructions on concentric
phantom data, where the measured stiffness profile, and best-fit profile based on Ω̂, are indi-
cated.
135 Ē 0 + σE 0

Measured
Ē 0
100
Best-fit
Ē 0 − σE 0
32
24
0 13 16 37
(a) Hybrid algorithm, 1 generation CO, 10 iterations GD.
135
100
32
24
0 13 16 37
(b) Hybrid algorithm, 10 generations CO, 10 iterations GD.
135
100
32
24
0 13 16 37
(c) Hybrid algorithm, 100 generations CO, 10 iterations GD.
Figure 5.17 The stiffness profile distribution from 100 hybrid reconstructions on concen-
tric phantom data, where the measured stiffness profile, and best-fit profile based on Ω̂, are
indicated.
To confirm the ability of the hybrid algorithm to return the correct stiffness dis-
tribution when given a motion data set from a homogenous phantom, the algorithm
was additionally performed 100 times on the soft homogeneous silicone phantom data
set previously studied in Section 4.2. The hybrid algorithm in this case comprised 10
CO generations followed by 10 GD iterations, a combination found to give the best
performance when considering computational cost, based on the results in Table 5.6.
Table 5.6 Results from the hybrid algorithm optimization.
Consecutive Solves
Iterations Metric Serial Parallel
Algorithm CO GD Ψ2 Σ2 (actual) (theory)
GD - 10 0.79 0.55 40 10
CO 1 - 0.64 0.19 10 1
10 - 0.41 0.09 100 10
100 - 0.25 0.04 1,000 100
Hybrid 1 10 0.39 0.07 50 11
10 10 0.25 0.03 140 20
100 10 0.23 0.03 1,040 110
The distribution of stiffness profiles from these 100 results is shown in Figure 5.18,
along with the corresponding measured, and minimum error profiles. Near constant
soft stiffness is clearly indicated in the result, with increased uncertainty towards the
center of the phantom as a result of the distance from the measured displacements on
the surface. The reconstruction metrics Ψ2 and Σ2 are not relevant in the homogenous
case, where the only independent parameter is Es0 . The best-fit storage modulus value
of 28 kPa in Figure 5.18, representing the value of Es0 corresponding to Ω̂ for this error
domain, is in agreement with the estimated value of E 0 = 28 kPa for the soft silicone
phantom in Section 4.2.
5.3.3 Discussion
The results shown in Figures 5.15–5.17 are the distribution of the spatially interpreted
stiffness solutions resulting from performing 100 of the respective reconstructions. As
such, they should not be interpreted as the ‘result’ of a reconstruction, but rather a
distribution drawn from several reconstruction results. Reduced values of Ψ and Σ2 are
desirable, as they indicate both reduced mean error and reduced uncertainty, which is
important when performing a limited number of reconstructions on a real data set.
The distribution of GD reconstruction results shown in Figure 5.15 do not give a
close match to the best-fit phantom stiffness distribution, with reconstruction metrics
135 Ē 0 + σE 0

Measured
Ē 0
100
Best-fit
Ē 0 − σE 0
32
28
0 13 16 37
Figure 5.18 The reconstructed stiffness profile distribution from 100 hybrid reconstructions
performed on the soft homogenous silicone phantom data, where the measured stiffness profile,
and best-fit profile based on Ω̂, are indicated.
of Ψ2 = 0.79, Σ2 = 0.55. The large variance visible in the plot at all radial positions
indicates that the GD reconstructions failed to converge consistently for the parameters
Es0 , Eh0 and Pc . Even using a damped material model, it is apparent that local minima
severely hamper the algorithm performance, a result consistent with previous studies
applying the GD method to reconstructions with fewer parameters in Sections 5.1
and 5.2.
The accuracy of reconstructed solutions showed a marked improvement when using

an optimized CO approach. Mean error and variance in the stiffness profile distributions
in Figure 5.16 are significantly lower than the GD case, particularly in the estimation
of the soft modulus value, Es0 . Improvement is also visible in the definition of the step
stiffness change between hard and soft regions. Performance metrics of Ψ = 0.25 and
Σ2 = 0.04 were achieved after 100 generations, where a total of 1000 FE simulations
were required for each reconstruction.
Adding GD reconstruction to CO to form a hybrid method led to further improve-

ments in the distribution of results in all cases shown in Figure 5.17. Significantly, the
total number of FE solves required to obtain the hybrid result in Figure 5.17(b) is only
14% of the number of solves required to achieve the CO result shown in Figure 5.16(c).
This result demonstrates that a hybrid approach can achieve equivalent performance
to CO performed alone, while requiring markedly fewer computational resources.
Table 5.6 summarizes the number of FE simulations required for each instance of
the three algorithms considered. The results from this study using a serial computation
approach show that performing fewer than 100 CO generations can significantly reduce
computational cost, while retaining a significant reconstruction performance advan-
tage in relation to the GD algorithm. When considering a parallel implementation of

the hybrid algorithm, where independent FE solves are performed concurrently across
multiple processors, the total number of consecutive FE solves reduces to 20 for an
algorithm comprising 10 generations of CO and 10 iterations of GD reconstruction.
Considering the time required to perform one FE simulation on a modern desktop
computer, a single instance of a hybrid reconstruction algorithm could be performed
in as little as fifteen minutes on a cluster of such machines. Even considering the
additional time required to build up the database of reconstructed solutions used to
generate the plots in Figures 5.16 and 5.17, this result is encouraging when considering
the exponential increase in the complexity of computer hardware originally predicted by
Moore, where such complexity and performance increases are continuing in the present
day [112–114].
A reasonable match was obtained between the reconstructed concentric phantom
stiffness profile using the hybrid algorithm, and the measured stiffness profile based on
static test results from Section 3.2. As the measured modulus values were from samples
produced at a different time, they were not expected to provide exact target values for
reconstruction, given the number of experimental uncertainties. The smaller inclusion
diameter indicated by the best-fit case may be a result of the stiffness interpolation
required when solving a FE problem with nodally-based material properties. The effects
of this property interpolation should be reduced with the finer nodal spacing in the FE
mesh used for simulations.
The hybrid algorithm optimization study described in this section demonstrated
the ability of a three parameter approach to reconstruct heterogeneous phantom stiff-
ness using experimentally measured displacements. Though the computational cost
was high, the CO method obtained significantly better reconstruction results than a
GD approach, confirming the findings of the previous simulation study. The optimized
hybrid reconstruction method demonstrated an ability to better the performance of CO
performed alone, with almost an order of magnitude reduction in computational cost.
When the promising results from the contradictive approach to parameter estimation
in Section 4.3 are considered, a hybrid approach to inclusion imaging has the potential
to solve problems with increased complexity. Here, more realistic internal geometries
can be described without increasing the total number of parameters required, by fixing
the modulus value of any predicted hard inclusion.
5.4 Spherical Phantom Reconstructions
The final elastographic reconstruction study performed considered a more complicated

and clinically-relevant heterogeneous phantom geometry. In this experiment, a com-
bination of the optimized hybrid algorithm from Section 5.3 and the contradictive a
5.4 SPHERICAL PHANTOM RECONSTRUCTIONS 133
priori stiffness assumption described in Section 4.3.3 are used to add an additional
shape parameter to the reconstruction problem. The new geometry used for this study
was a hard inclusion completely contained within a soft body, representing the closest
experimental approximation to a clinical scenario considered to this point.
5.4.1 Spherical Silicone Phantom
A ‘spherical’ heterogeneous silicone phantom was created using the same methods and
materials as the stacked and concentric silicone phantoms described in Chapter 3.
The phantom comprised a 40 mm diameter sphere of hard silicone suspended within
a soft silicone cylinder of the same external geometry as previous silicone phantoms.
Creating the phantom involved first casting the spherical inclusion from hard silicone,
then suspending this sphere within the cylindrical mold phantom mold while pouring
soft silicone to the desired level.
Experimental motion capture from the spherical phantom used the same hard-
ware and settings as described in Section 3.3.2, where mechanical actuation was at a
frequency of 100 Hz, with a 0.5 mm amplitude. Figure 5.19 contains an image of the
phantom during actuation, along with the tracked reference points on the phantom
surface across one steady state motion cycle. The motion fitting performed on this
experimental data used the same damped assumption as all previous silicone phantom
studies.
(c)
(a) Motion capture. (b) Tracked reference point locations. (c)
Figure 5.19 The experimentally observed motion of the spherical silicone phantom.
5.4.2 Parameter Reconstruction
The elastic reconstruction of the spherical phantom considered a total of three parame-
ters, allowing the hybrid algorithm applied in Section 5.3 to be used in this study with
minimal modification. The contradictive approach to internal stiffness estimation in-
troduced in Section 4.3.2 required any hard inclusion to have a constant stiffness value,
meaning the value of Eh0 was fixed, and hence no longer a reconstructed parameter. The
loss of this second stiffness parameter allowed the more complicated spherical phantom
geometry to be described using an additional positional value without increasing the to-
tal number of independent parameters. Here, the ratio of the inclusion diameter to the
overall phantom diameter was described by the parameter Pd . Physical measurements
of the phantom gave an actual value for this parameter of P̂d = 57%. The vertical po-
sition of the spherical inclusion, which was assumed to always lie along the central axis
of the cylinder, was described by the height parameter Ph , where P̂h = 50%. During all
reconstructions, the soft modulus value, Es0 , was expressed to the nearest 1 kPa, while
the values for the positional parameters Pd and Ph were calculated to the nearest 5 %.
5.4.2.1 Spherical Inclusion Phantom
A first attempt at three parameter reconstruction of the spherical phantom was made
using the hybrid algorithm settings obtained in Section 5.3, where Πmate = 0.6, Πmutate =
0.2, and the number of genes was set at NG = 10. Both CO and GD phases of the
hybrid algorithm were run for 10 iterations. Figure 5.20(a) shows the mean stiffness
distribution resulting from 100 hybrid reconstructions performed on the spherical ge-
ometry. The plane of the figure passes through the center of the cylindrical phantom,
and the stiffness value at each point within the image is calculated as the mean indi-
cated stiffness value over all reconstruction results. The measured position and size of
the actual phantom inclusion is indicated as a dashed circle. Two distinct regions of
high stiffness are visible, where an area of high stiffness near the top of the phantom is
indicated in addition to the expected centrally located hard region.
The inconclusive result in Figure 5.20(a) led to an adjustment in the hybrid al-
gorithm settings, based on the results from the optimization study performed in Sec-
tion 5.3. The previous study found that increasing the number of genes in the CO phase
of the hybrid algorithm had a significant performance benefit, even when considering
the increased computational cost associated with the larger population. With this re-
sult in mind, reconstructions were performed on the spherical geometry using greater
values for NG , where the stiffness distribution resulting from hybrid reconstructions
performed with NG = 50 is shown in Figure 5.20(b). Here, improvement can be seen
in the spatial modulus distribution, where the presence of the large inclusion near the
top of the phantom is somewhat diminished.
25 Modulus Mean, Ē 0 (kPa) 100 26 Modulus Mean, Ē 0 (kPa) 100
74 74
Height (mm)
Height (mm)
0 0
{37 Radial Position (mm) 37 {37 Radial Position (mm) 37
(a) Hybrid algorithm, NG = 10. (b) Hybrid algorithm, NG = 50.
Figure 5.20 A spatial representation of the spherical phantom modulus reconstructed using
the hybrid algorithm with varying numbers of genes.
Further testing led to satisfactory reconstruction performance with a value of

NG = 100. Figure 5.21(a) shows the resulting spatial modulus distribution at this
value. In this case, the result strongly indicated a preference for the central inclusion
location at the measured height, with a slightly smaller diameter. The reconstructed
soft modulus value was approximately Es0 = 26 kPa. The variance of the reconstructed
modulus within the phantom is shown in Figure 5.21(b). Figure 5.21(c) indicates the
reconstructed phantom stiffness variation with height, where the horizontal axis of the
plot corresponds to the dashed red transect in Figure 5.21(a). The dashed line in
Figure 5.21(c) represents the best-fit modulus profile that is the target of the recon-
struction in this case. Figures 5.21(b) and (c) indicate that the highest variance values
occurs near the edges of the central inclusion.
5.4.2.2 Homogeneous Phantom Reconstruction
The specificity of the hybrid algorithm applied to the spherical geometry was evaluated
by reconstructing internal stiffness using the spherical model and the experimental
motion data obtained from the soft homogeneous silicone phantom. In this case, the
expected result was an entirely soft distribution free from significant hard inclusions.
The actual reconstructed stiffness distribution obtained is shown in Figure 5.22, where
all reconstructions used the same algorithm settings as those used when generating
Figure 5.21.
26 Modulus Mean, Ē 0 (kPa) 100 0 Standard Deviation, σE0 (kPa) 37
74 74
Height (mm)
Height (mm)
0 0
(a) Mean reconstructed modulus. (b) Reconstructed modulus variance.
120
100 Ē 0 + σE 0
Ē 0
Ē 0 − σE 0
26
0
0 17 54 74
Height (mm)
(c) Modulus variation along the dashed vertical transect in (a), where the target
profile is indicated.
Figure 5.21 The spherical silicone phantom modulus distribution reconstructed using a con-
tradictive hybrid algorithm with 100 genes.
28 Modulus Mean, Ē 0 (kPa) 100 0 Standard Deviation, σE0 (kPa) 36
74 74
Height (mm)
Height (mm)
0 0
(a) Mean reconstructed modulus. (b) Reconstructed modulus variance.
120
Ē 0 + σE 0
Ē 0
Ē 0 − σE 0
28
0
0 74
Height (mm)
(c) Modulus variation along the dashed vertical transect in (a), where the target
profile is indicated.
Figure 5.22 The soft homogenous silicone phantom modulus distribution reconstructed using
a contradictive hybrid algorithm with 100 genes.
Two small areas of moderate stiffness are visible in the resulting plots at positions
near the top of the model. In both cases, the modulus variance in the vicinity of the
areas of increased stiffness is correspondingly high. The estimated value for the soft
modulus was Es0 = 28 kPa.
5.5 Discussion
The reconstructed results in Figure 5.20 indicate that the three dimensional error do-
main for the spherical reconstruction problem does not contain an obvious single min-
imum value. A bi-modal distribution is apparent, with Figure 5.20(a) indicating that
approximately half of all reconstructions predicted a large inclusion located near the
top of the phantom model that would be visible on the upper phantom surface. While
the use of CO in previous reconstructions allowed the hybrid algorithm to avoid such
local minima, the presence of two potential solutions with very similar surface error
metric values presents a significant challenge to the hybrid algorithm when using a
population comprising 10 genes.
The improvement in reconstructed modulus distribution when adding genes to the
CO phase of the hybrid reconstruction indicates that while the presence of the upper
inclusion was the result of a local error minimum, the global error minimum Ω̂ closely
corresponds to parameters describing the actual phantom geometry. Additional genes
increase the diversity of the CO population, meaning that mating and mutation have
a greater chance of creating an offspring gene closer to the globally optimal parameter
combination. The population diversity when NG = 100 reached a level at which the
globally optimum solution was preferred by almost all reconstructions.
Figure 5.23 shows a plane through the error domain for the spherical case at Es0 =
27 kPa, containing the global minimum error value Ω̂. This plot confirms that the
strongly bi-modal distribution observed during the initial reconstructions is the result
of a local error minimum within the parameter domain. A region of low motion error
is visible at inclusion diameter and height values greater than the globally optimal
solution. The large three-dimensional size of this region means that a large number
of solutions converge here, despite the region not containing Ω̂. However, the positive
result obtained following an increase in the size of the CO population to NG = 100
indicates that significant local minimum can be overcome with prudent modification of
the hybrid algorithm settings.
The existence of substantial local error minimum in the spherical case is likely
due to a lack of motion data from the upper surface of the phantom. The surface
error studies performed on the stacked phantom in Section 4.3 indicate that when
surface reference points are located across an external boundary between soft and hard
material, the existence and position of such an interface can be clearly determined.
5.5 DISCUSSION 139
100
Inclusion Height, Ph (%)
50
0
0 45 100
Inclusion Diameter, Pd (%)
Figure 5.23 A plane at Es0 = 27 kPa through the spherical phantom error domain, where the
location of global minimum error, Ω̂, is indicated (+).
Additional reference points on the top surface of the phantom should allow the motion
error metric to clearly identify the presence or absence of inclusions near or at the top
surface of the model based on their surface displacement values.
The slight underestimation of the size of the spherical inclusion in Figure 5.21
is likely a result of the nodal stiffness interpolation used in the FE model, and is in
line with the results from the concentric phantom in Section 5.3. Additionally, the
rounding of Pd during the reconstructions may have led to a small underestimation of
the predicted inclusion diameter.
The specificity analysis of the hybrid reconstruction algorithm using the homoge-
neous phantom experimental data indicated the presence of small, moderate stiffness
regions near the top of the phantom. While the soft modulus value reconstructed
across the majority of the phantom was in good agreement with the 28 kPa value ob-
tained from the soft homogeneous silicone phantom in Section 4.2, the small areas of
elevated stiffness are undesirable, as the ability of any screening system to minimize
the occurrence of false positive results is important. Consideration of the error domain
for the homogeneous phantom reconstruction, shown in Figure 5.24, reveals that the
presence of the small inclusions near the top of the model are due to inaccuracies in
the error domain. Here, the global error metric minimum Ω̂ occurs at Pd = 20%,
Ph = 75%, conflicting with the expected location of this minimum, along the left edge
of the figure at Pd = 0%. While the difference in motion error metric between these
two locations is very small, the increased performance of the hybrid algorithm with
100 genes meant that a large number of solutions converged to this globally optimal
combination of parameters. This result is further evidence that motion data at the
upper phantom surface may improve the ability of the surface error metric to correctly
eliminate geometries that predict inclusions near the top of the model.
100
75
Inclusion Height, Ph (%)
0
0 20 100
Inclusion Diameter, Pd (%)
Figure 5.24 A plane at Es0 = 26 kPa through the soft homogeneous phantom error domain,
where the location of minimum error within the plane is indicated (+).
Despite the challenges presented by both error domains, the presence of a hard
inclusion in the spherical case was more strongly suggested than the presence of any
hard area in the homogeneous phantom. All areas of elevated stiffness corresponded
to the highest observed values for modulus variance in the homogeneous phantom
reconstructions. This result differs from the spherical case, where variance values within
the interior of the hard inclusion were lower, and the highest variance values were
around the edge of the inclusion, as expected.
Increasing the number of genes in the CO stage of the hybrid reconstruction gives
a subsequent increase in computational cost, summarized in Table 5.7. Here, the hy-
brid reconstructions with NG = 100 require 10 times the number of FE solves as the
baseline case where NG = 10, when performed on a serial system. However, the in-
herent suitability of CO to parallel computation means that if performed on a large
parallel system of processors, an increase in genes would not increase the overall time
required for computation, as each gene would be simulated on a dedicated processor.
Additionally, it should be noted that all the hybrid reconstructions required signifi-
cantly less computational effort than an exhaustive search within the problem domain.
5.5 DISCUSSION 141
For the spherical geometry considered in this case, with a total of three independent
parameters, an exhaustive search would involve a total of 4,851 FE simulations.
Table 5.7 Computational cost considerations for the hybrid algorithm applied to the spherical
silicone phantom.
Consecutive Solves
Serial Parallel
Algorithm NG (actual) (theory)
Hybrid 10 140 20
50 540 20
100 1,040 20
The final reconstruction study indicated that the hybrid reconstruction method had
the ability to reconstruct the internal stiffness distribution of a more realistic geometry
containing a spherical inclusion. While the algorithm required additional tuning to
achieve a positive result in this case, the additional computational cost could be offset
with the use of parallel computation. The continued presence of local error minima in
this study highlighted the requirement of a DIET system to have accurate motion data
from as much of the exposed object surface as possible, if reconstruction results are to
return the expected stiffness distribution.
Chapter 6
Conclusions
The research presented in this thesis has shown the DIET approach to stiffness recon-
struction to be both technically and clinically realistic. The feasibility of the system
has been evaluated and proven with a range of analytical and experimental testing.
Promising results have been obtained that warrant further system development to-
wards the eventual goal of a clinical trial. This chapter presents the technical and
clinical conclusions that can be drawn from the research performed to this point.
6.1 Technical Outcomes
A number of strategies for fitting experimentally observed phantom displacements to a

well defined motion path were developed and implemented. Each approach was based
on the modelling assumptions used for the relevant FE simulation. These fitting algo-
rithms allowed experimentally measured motion to be directly compared with computer
model displacement output from FE simulation, and form an important link between
observed and simulated behavior.
The methods devised for experimental motion capture and processing were vali-
dated by a testing phase using tissue-approximating phantoms. The surface motion
error metrics developed allowed quantitative comparison of measured and simulated
motions. Implementation of these metrics in conjunction with the motion fitting al-
gorithms confirmed that experimentally observed motion can be accurately compared
with simulated displacements. These techniques were developed using simulated and
phantom data, however they are directly applicable to experimental data from any
source.
The motion studies completed on experimental data indicated that surface motion
measurements alone are sufficient to estimate an internal stiffness distribution using
a limited number of independent parameters. The Young’s Modulus values estimated
using surface error were a good match to independent mechanical test results when
144 CHAPTER 6 CONCLUSIONS
considering the experimental uncertainties present. This result is the first of its kind
in the field, and indicates a major milestone in the technical validation of the DIET
approach to stiffness reconstruction.
A number of techniques for mechanical property estimation and reconstruction

were developed and tested. The reconstruction methods included both traditional
deterministic approaches and novel methods involving stochastic components. The ini-
tial validation of each technique used simulated data, providing an environment where
noise levels and other uncertainties were controllable. Following validation, the estima-
tion and reconstruction methods were successfully applied to experimental data from
tissue-approximating phantoms, showing that identifying a minimum-error solution and
automating the search for it is feasible.
Metrics to measure both the accuracy of a reconstruction method and its compu-
tational cost were developed and implemented. These metrics allowed the evaluation
of each reconstruction algorithm to include performance aspects from both a technical
and practical standpoint. In addition, the computational cost metrics give a useful indi-
cation of the time required to perform the reconstruction on modern, parallel computer
systems.
A novel hybrid reconstruction algorithm was implemented and was able to suc-
cessfully reconstruct internal stiffness based on minimal surface motion measurements.
While this method was identified as having a greater computational cost than a gradient-
descent method performed in isolation, the performance of the hybrid approach far
outweighed the extra computation required. In addition, the success of the DIET re-
constructions with such a limited set of input data gives this method a distinct advan-
tage over alternative modalities such as MRE, where the full-volume data sets involved
require significant resources and computation at the data acquisition stage.
During both analytical and experimental stages of the research, sources of mod-
elling and experimental error were identified. The magnitude and consequences of these
errors were considered when evaluating the accuracy of the results obtained. When
necessary and practical, advances in experimental technique and modelling strategy
allowed these sources of error to be minimized or eliminated, giving greater confidence
in the subsequent outcomes.
When considered from a technical standpoint, the DIET approach to stiffness re-
construction provides an output that is comparable to similar elastographic techniques.
That these results are obtained with significantly less input data presents a significant
advantage when considering the practicalities of experimental data collection. In addi-
tion, the effective computational cost of DIET reconstruction is rapidly diminishing as
parallel computing becomes more mainstream.
6.2 CLINICAL OUTCOMES 145
6.2 Clinical Outcomes
Analytical pre-cursor studies were completed as a proof of concept for the DIET system.
The result of this work indicated that the concept of stiffness reconstruction using
DIET may have clinical applicability, and warranted further investigation. While the
geometry used in these studies was not an exact match to clinical reality, it was shown
that the observed surface motions from an object the same size as a human breast
could be used to differentiate between internal stiffness configurations.
The experimental results from gelatine and silicone phantoms indicated that the
DIET method can reconstruct elastic properties in an experimental situation. Here,
stiffness identification of homogeneous phantoms was performed with good agreement
to independent test results. Significantly, the difference in observed surface motion
between homogeneous phantoms with a stiffness contrast less than independently-
measured tumor contrast values was clearly evident.
Further experiments using silicone phantoms indicated that in addition to homo-
geneous phantoms, heterogeneous stiffness distributions could be identified using only
surface measurements. These experiments provided a further step towards clinical re-
ality, where the identification of areas of different stiffness within a single object is a
key objective. Results from the concentric and spherical silicone phantoms were ob-
tained with measurements made entirely on the soft region of the phantom, in a similar
manner to the envisaged experimental data collection from a human breast.
The contradictive approach to inclusion imaging was found to give more accurate
results than the traditional corroborative approach. The contradictive approach places
more emphasis on correctly identifying parameters that have the greatest affect on the
observed motion in an attempt to increase the efficiency of the reconstruction process.
In a clinical scenario, any loss of accuracy when assuming an inclusion stiffness value
would not present a significant disadvantage, as the DIET system is intended as a
screening test focused on the presence of inclusions, rather than their exact stiffness.
The final evaluation of the DIET system used a phantom with a spherical inclu-
sion, and showed that reconstruction of simple geometry resembling the configuration
of a breast tumor is possible. While the exact stiffness was not characterized, the
high stiffness of the inclusion was identified to within accuracy limits acceptable for
a screening test. In addition, the specificity of the system was demonstrated, along
with clear indications of how this metric could be improved. These encouraging results
suggest that further system development should be pursued, with the medium-term
aim of obtaining experimental data from a human subject.
Chapter 7
Future Work
The work presented in this thesis developed a number of aspects of the DIET system
from a theoretical proof of concept experiment through to a successful elastic property
reconstruction on heterogeneous, tissue-approximating phantoms. Although the more
complicated geometries that will be present in a clinical scenario may require a more
sophisticated approach to the description of the shape of an inclusion, the framework
developed during this research should allow advances in actuation, data capture, and
model simulation to be incorporated without significant effort. This chapter includes
thoughts on the development required to improve the DIET system towards its appli-
cation in a clinical scenario.
7.1 Short-Term Development Goals
Accurate stiffness reconstruction based on surface motion requires a comprehensive and

exact experimental measurement of surface displacements. Issues with surface motion
error metric identified during the reconstruction of the spherical phantom confirmed
the importance of having complete surface motion data available for the reconstruction
problem. In this case, additional reference points across the top of the phantom should
improve the reconstruction solution distribution and eliminate the spurious areas of
high stiffness visible in the results. Confirming this improvement would require the
collection of an additional data set with an increased reference point density, including
the continuation of the vertical line of reference points across the top of the phantom.
While the range of reconstructions performed were able to clearly indicate the pres-
ence of an inclusion, the limits of this reconstruction have not been explored. Such a
study would involve creating a range of phantoms with spherical inclusions of vary-
ing size. In conjunction with an increased reference point coverage, this experiment
should provide an indication of the minimum size inclusion the DIET system can ac-
148 CHAPTER 7 FUTURE WORK
curately detect. In addition to inclusion size, variations in inclusion stiffness provide

an additional avenue for investigation.
Reducing the amplitude of actuation for subsequent phantom testing may help to
reduce the observed under-estimation of experimental displacement by the linear FE
model. Recent results from the motion tracking system have indicated that reduced
point motion amplitude would not present a significant problem when tracking points,
removing the requirement for the larger amplitudes used for the experimental work in
this thesis. While the addition of material non-linearity to the FE model provides an
alternative approach to improving the displacement match, such a change would add
significant complexity to any reconstruction problem that is undesirable at this stage.
Further increases in computational capacity will allow more refined meshes to be

used for FE simulation. Mesh refinement will help to reduce the diffusive effect of
nodally-distributed material properties on the interface between hard and soft material
in the FE model. It is hoped that greater definition between adjacent material regions
of different properties will improve the slight under-estimation of inclusion size observed
in both the concentric and spherical reconstructed results.
Varying the frequency of actuation has the potential to provide additional exper-
imental information useful for the reconstruction problem. Here, a range of data sets
taken across different frequencies would allow the phantom motion at a range of mode
shapes to be characterized. The presence of an inclusion may be more obvious in a
specific mode of vibration, allowing a reconstruction performed using data at more
than one actuation frequency to identify inclusions not visible at a single frequency
alone. Clinically, such an approach should not present a significant challenge, allowing
motion at a range of frequencies to be obtained from a patient in a matter of minutes.
Shape functions, such as radial-basis functions, may provide an alternative ap-

proach to describing the geometry of an inclusion. Such functions have the ability to
define inclusions of a more complicated nature that cannot be accurately approximated
by the spherical model currently used. In addition, such shape functions may be differ-
entiable with respect to surface displacements, allowing an analytic calculation of the
Jacobian matrix for gradient-descent reconstructions, which would be advantageous as
the number of shape parameters is increased.
The scenario of multiple inclusions has not been addressed in research to this
point. This situation presents an additional experimental scenario relevant to DIET
system development, particularly as the case of multiple tumors within a single breast is
clinically relevant. The description of internal geometry in such cases may benefit from
the use of the shape functions previously described, and is likely to require an increase in
the number of reconstructed parameters. The current phantom manufacturing process
would not require significant modification in order to create phantoms with multiple
7.2 LONG-TERM DIET SYSTEM DEVELOPMENT 149
inclusions, and reconstruction of internal geometry will prove a significant challenge to

an advanced version of the current reconstruction algorithm.
Further studies using silicone for phantom manufacture should include an in-depth
study of the silicone material properties. While the static tests performed in the current
work provided a useful benchmark for reconstruction performance, there was an obvi-
ous variation in silicone properties across the different batches used for homogeneous
phantoms, heterogeneous phantoms, and material testing. Material property varia-
tions with temperature should be included in this work, as the notable warming of
the silicone phantoms during actuation may lead to variations in the observed stiffness
value.
Significant potential was identified for a reduction in computational time through
the use of parallel methods for reconstruction algorithm FE simulations. This paral-
lelization should be straightforward through the use of protocols such as the Message
Passing Interface (MPI) libraries. Parallelization using MPI has not been required up
to this point due to the small number of parameters reconstructed, but will become
inevitable as the complexity of the reconstruction problem increases.
7.2 Long-Term DIET System Development
Development of a DIET system capable of performing a clinical trial involves a number

of important milestones outside the scope of work presented in this thesis. Concurrent
research is investigating a technique where colored glitter can be used to significantly
improve the reference point density of the surface of the phantoms [109]. This technique
should provide a large increase in the amount of surface motion data available to the
reconstruction problem, while presenting a clinically attractive method for obtaining a
high point density on the breast surface.
An alternative to manually applied surface features is the use of the natural varia-
tion in skin appearance to form reference points for motion tracking. Such points can
be identified from image cross-correlation and speckle patterns, and have the ability to
provide a high point density from moving skin [95]. This technique may be useful for
motion tracking in a clinical scenario, as it would avoid any requirement for artificial
reference points that may be inconvenient or impractical to apply to the breast of a
patient.
Adding more cameras to the current imaging setup will allow displacements to be
tracked over the entire surface of a phantom, an improvement on the single line of dots
currently used. This improved surface coverage would release the current constraint
that requires all phantoms to be symmetric. Asymmetry would allow more realistic
phantoms to be developed, where the inclusion position can vary radially as well as
vertically. In addition, phantoms with asymmetric external geometry that closely re-
150 CHAPTER 7 FUTURE WORK
semble breast geometry can be tested, providing more clinically relevant data for the
reconstruction problem.
Boundary Element Methods (BEM) are being investigated as an alternative to the
current FE method used for simulating model displacements using a computer [115].
By formulating the forward problem in terms of boundary integral equations, surface
displacements can be obtained using a mesh that only describes the external and inter-
nal boundaries between discrete regions. While currently limited to a two-dimensional
case, BEM is showing promise for incorporation into the DIET system, allowing model
simulation using only the relevant surface geometry information.
A major goal for the DIET system is a clinical test that would provide a set
of surface displacements from a real breast. Such a test requires the development
of a patient positioning device, and refinement of the actuation and motion capture
system. While these developments will present a range of significant challenges, having
experimental data from a real breast will allow reconstruction algorithms to be refined
for use with clinically relevant data.
Appendix A
The Finite Element Method
The finite element (FE) method is a powerful tool that allows computers to simulate
a large range of physical scenarios. When used to solve problems in computational
solid mechanics, FE methods can generate transient and steady-state solutions in both
static and time-harmonic scenarios. The material in this appendix is intended to give
a brief overview of the FE approach used to solve steady-state, time-harmonic solid
mechanics problems for three-dimensional models.
Weighted Residuals
The equilibrium conditions for a solid body in a static system can be expressed as a
null divergence of the stress tensor,
σij,j = 0. (A.1)
When considering the inertial forces present in an undamped, steady-state, time-

harmonic system, Equation A.1 becomes
σ̄ij,j + ρω 2 ūi = 0, (A.2)
where ρ represents the material density, ω the system frequency, and ūi motion am-
plitude, and σ̄ the corresponding time-harmonic stress tensor. In the case where the
exact displacement solution within the body is not known, and approximated as û,
Equation A.2 will have a non-zero residual term,
R = σ̂ij,j + ρω 2 ûi , (A.3)

152 APPENDIX A: THE FINITE ELEMENT METHOD
where σ̂ are the stresses given by the approximate displacement solution û. In the
ideal case this residual term will be zero over the entire domain, meaning any weighted
integration of R should also be zero,
D ¡ ¢E
W k σ̂ij,j + ρω 2 ûi = 0, (A.4)
where the angled brackets indicate integration over the three-dimensional domain, and
W k is an arbitrarily-chosen weighting function. As the number of weighting functions,
Nk , increases, the match between the approximate and exact stress solutions improves
towards the limit,
lim û = ū. (A.5)

Nk →∞
Interpolating Polynomials
The FE method makes use of interpolating functions, φk , to describe the value of field
variables at any position within the domain. It is the local influence of these interpolat-
ing polynomials that leads to the quantification ‘finite’ in ‘finite element methods’. The
shape functions used for field interpolation can take a variety of forms, including simple
linear functions, and any number of higher-order descriptions. While higher-order func-
tions, such as cubic polynomials, allow a more accurate description of field variables
within a discretized element, the integration performed on these elements has added
complexity. Given limited computational resources, a choice must be made between
using high-order polynomials on larger elements, or smaller elements with simpler in-
terpolating functions. Linear interpolating functions were used in this work, allowing a
fine discretization of the domain while reducing integration times. An example of the
linear tetrahedral element used in all three-dimensional FE code is shown in Figure A.1,
where the four nodes are numbered according to a right-handed convention.
When considering the case where the interpolating polynomial is linear, field vari-
able interpolation within a four-noded tetrahedral element takes the form
4
X
ûe = ûk φk (x, y, z), (A.6)
k=1
where in this example ûe indicates an approximation to the displacement solution ū at

the position (x, y, z).
APPENDIX A: THE FINITE ELEMENT METHOD 153
Node 4
Node 2
Node 3
Node 1
Figure A.1 A three-dimensional linear tetrahedral element with nodes numbered according
to a right-handed convention.
The Galerkin Method
The Galerkin method refers to a technique where the weighting functions described in
relation to the method of weighted residuals additionally form the basis functions for
the interpolating polynomials,
W k = φk (A.7)
This approach is widely popular in FE techniques due to its simplicity and effectiveness.
It is standard practice in FE methods to perform the integration of the weighted
residual statement in Equation A.4 using Green’s theorem,
D ¡ ¢E
φki σ̂ij,j + ρω 2 ûi =0
D E ZZ
− φki,j σ̂ij + φki ρω 2 ûi + ° φki σ̂ij nj dS = 0, (A.8)
where nj is a surface outward normal unit vector. Equation A.8 is termed the ‘weak’
formulation of the weighted residual statement, as stress boundary conditions are only
required to match an integration over the surface upon which they are applied.
Stiffness Matrix Terms
Expressing the differential system provided by Equation A.2 in terms of displacements

requires a conversion from stress to strain terms. The relationship between displace-
ments, u, and strains, ², for a linear elastic material can be expressed in tensor form
as
1
²̄ij = 2 (ūi,j + ūj,i ) , (A.9)
while the corresponding relationship between strains and stresses can be described as
σ̄ij = λδij ²̄kk + 2µ²̄ij , (A.10)
where λ and µ are Lamé’s constants, and δij is the Kronecker delta function. Substi-
tuting the relationships defined in Equations A.9 and A.10 into the weak form, given
in Equation A.8 for the time-harmonic case, and writing out terms in full for the
x-direction gives
D ∂ û µ ¶
x ∂ ûx ∂ ûy ∂ ûz ∂φi E
2µ +λ + + + ...
∂x ∂x ∂y ∂z ∂x
D µ ∂ û ¶
∂ ûy ∂φi E D
µ ¶
∂ ûx ∂ ûy ∂φi E
x
µ + + µ + = −ρω 2 ûx . (A.11)
∂y ∂x ∂x ∂y ∂x ∂x
In order to solve differential equations using FE methods, it is desirable to form a

system of equations
£ ¤
k {u} = {b} , (A.12)
where the stiffness matrix [k] contains the material property terms, and {b} represents
forcing terms arising from inertial forces or externally-applied constraints. For the
three-dimensional case, each node has three degrees of freedom representing orthogonal
displacement components, and local stiffness matrix terms are defined
* +
∂φi ∂φj ∂φi ∂φj ∂φi ∂φj
k11 = (2µ + λ) +µ +µ
∂x ∂x ∂y ∂y ∂z ∂z
* +
∂φi ∂φj ∂φi ∂φj
k12 = λ +µ
∂x ∂y ∂y ∂x
* +
k13 = λ +µ
∂x ∂z ∂z ∂x
* +
k21 = λ +µ
∂y ∂x ∂x ∂y
* +
k22 = µ + (2µ + λ) +µ (A.13)
∂x ∂x ∂y ∂y ∂z ∂z
* +
k23 = λ +µ
∂y ∂z ∂z ∂y
* +
k31 = λ +µ
∂z ∂x ∂x ∂z
* +
k32 = λ +µ
∂z ∂y ∂y ∂z
* +
k33 = µ +µ + (2µ + λ) .
∂x ∂x ∂y ∂y ∂z ∂z
Following the assembly of these local element matrices into a global stiffness matrix,
which is sparse, symmetric, and positive definite, the global displacement field û can
be calculated via matrix inversion.
Viscoelasticity
The response of viscoelastic materials to time-harmonic loading can be simplified by

representing the material as a combination of connected springs and dashpots that
model stiffness and damping behavior. While a number of complicated rheological
models have been proposed to approximate the observed response of different viscoelas-
tic materials, this section will limit consideration to the commonly-used Voigt element
shown in Figure A.2 for simplicity. Here, a linear spring with stiffness E, and a dashpot
with damping parameter η, are configured in parallel.
The parallel spring and dashpot within the Voigt element require that strains are
equivalent in both components,
² = ²1 = ²2 . (A.14)
Figure A.2 A Voigt element, where a spring and dashpot are connected in parallel.
In addition, the total stress within the element can be expressed as the sum of the
individual component stresses,
σ = σ1 + σ2 (A.15)
Utilizing the stress-strain relationships for a linear spring and a dashpot, this combined
element stress can be expressed as
σ = E² + η ²̇, (A.16)
where E has units of Nm−2 , η has units of Nm−2 s, and ²̇ indicates the time-derivative
of strain. Substituting a time-harmonic stress and strain variation into the stress
definition provided by Equation A.16 and simplifying,
σ̄ = (E + iωη) ²̄, (A.17)
where the overbar indicates the amplitude of the stress and strain terms. In this case,
the equivalent elastic modulus of the element is a complex term with both real and
imaginary components,
E ∗ = E 0 + iE 00 , (A.18)
where E 0 = E, and E 00 = ωη.
While the modulus components derived here are accurate only for a Voigt element,
a complex-valued elastic modulus can be developed from many combinations of springs
and dashpots. When simulating damped model displacements using FE methods in this
work, the generalized description of this complex modulus provided by Equation A.18
was used, and no direct comparison to a specific rheological model was made.
Model Performance Near the Incompressible Limit
A simple experiment was designed to characterize the accuracy of compressible elas-

ticity FE simulations as the value of Poisson’s ratio, ν, approaches the incompressible
limit. A simple three-dimensional model was created by extruding a square cross-
section with side lengths of 38 mm to a height of 200 mm. Boundary conditions con-
strained the lower face of the block in all three orthogonal directions, with a 0.5 mm,
100 Hz sinusoidal displacement prescribed in one horizontal plane. This actuation in-
duced regular mechanical shear waves throughout the model volume. A damped mate-
rial model was used for all simulations, with the material storage modulus E 0 = 20 kPa,
and damping ratio ζ = 0.1. Material density was fixed at ρ = 1000 kg/m3 , and Poisson’s
ratio was varied from 0.45 to 0.499 in six increments.
The reference solution for this situation was obtained via FE simulation using an
incompressible material model. For the incompressible simulations, the model was
meshed using 27-noded quadratic elements, with an average element edge length of
4.7 mm, giving a total of 11,000 nodes, and 1,200 elements. The observed displacement
fields for this model over the range of values for Poisson’s ratio are shown in Figure A.3.
(a) ν = 0.45. (b) ν = 0.47. (c) ν = 0.48. (d) ν = 0.49. (e) ν = 0.495. (f) ν = 0.499.
Figure A.3 Incompressible FE model displacement fields at a range of Poisson’s ratio values.
As expected, the observed displacement fields for the incompressible simulations

remain consistent as Poisson’s ratio approaches the incompressible limit at ν = 0.5.
Minor changes in the displacement field are due to the influence of the changing Pois-
son’s ratio value on the material shear modulus. The effect of material damping can
be seen in the reduction in shear wave amplitude towards the top of the model.
The FE mesh for compressible elasticity simulations was composed of linear tetra-
hedral elements. The average element edge length was approximately 2.5 mm, giving
a total of 27,000 nodes, and 146,000 elements. Displacement fields from FE simulation
using the compressible FE model are shown in Figure A.4 for the range of Poisson’s
ratio values considered.
(a) ν = 0.45. (b) ν = 0.47. (c) ν = 0.48. (d) ν = 0.49. (e) ν = 0.495. (f) ν = 0.499.
Figure A.4 Compressible FE model displacement fields at a range of Poisson’s ratio values.
The displacement fields from the compressible simulations show good agreement
with the reference solution at values of Poisson’s ratio up to and including ν = 0.49.
Below this value, the observed wavelength and motion field are very similar to the
reference case in Figure A.3. At ν = 0.49, small inconsistencies are visible, but the
overall motion field is still regular, and generally consistent with the reference solution.
Simulated displacement fields in Figure A.4 at Poisson’s ratio values above ν = 0.49
indicate that FE simulations in this region begin to lose accuracy when compared to
the reference results. At ν = 0.495, clear variation is apparent, and at ν = 0.499, the
standing wave pattern has been significantly disrupted, and can not be relied on as an
accurate approximation to real-world behavior.
The simulations performed using a compressible elasticity FE model show good
agreement to a reference solution generated using an incompressible material model
for values of Poisson’s ratio up to and including ν = 0.49. The appearance of slight
deviation from the reference solution at ν = 0.49 means that this value should represent
the upper limit for simulations if the compressible FE model is expected to be free from
significant numerical inaccuracies.
Appendix B
Silicone Material Data
This appendix contains Material Safety Data Sheets (MSDS) for the two types of
silicone used during phantom manufacture.
P.O. Box 1339 P.O. Box 1339
160
Lakeside, AZ 85929 Lakeside, AZ 85929
Factor II, Factor II,
Phone:(928) 537-8387 Phone:(928) 537-8387
Incorporated Fax: (928) 537-0893 Incorporated Fax: (928) 537-0893
MATERIAL SAFETY DATA SHEET

A-341 Part A
5. FIRE FIGHTING MEASURES
Soft Gel FLASH POINT: (TCC) N/A
In Case of Emergency Contact Factor II, Inc. 928-537-8387
FLAMMABLE LIMITS IN AIR (by volume): Not Applicable N/A
Factor II Technology urges each customer or recipient of this MSDS to study it carefully to become aware of and
understand the hazards associated with the product. The reader should consider consulting reference works or EXTINGUISHING MEDIA: Dry Chemical ,Foam or Carbon Dioxide
individuals who are experts in ventilation, toxicology, and fire prevention, as necessary or appropriate to the use and
understanding of the data contained in this MSDS. SPECIAL FIREFIGHTING PROCEDURES: None known.
To promote safe handling each customer or recipient should: (1) notify and furnish its employees, agents, contractors, UNUSUAL FIRE AND EXPLOXION HAZARDS: None Known
customers, and others whom it knows or believes will use this material of the information regarding hazards or safety; (2)
request its customers to notify their employees, customers and other users of the product of this information. 6. ACCIDENTAL RELEASE MEASURES
STEPS TO BE TAKEN IF MATERIAL IS RELEASED OR SPILLED:
1. CHEMICAL PRODUCT Contain the spill or leak, scrape up with cardboard or a ra and place in a container.
PRODUCT NAME: RTV Heat Cure Silicone WASTE DISPOSAL METHOD: Dispose of in accordance with all Federal, State, and local regulations.
CHEMICAL NAME: Organopolysiloxane Mixture
7. HANDLING AND STORAGE
2. PRODUCT COMPOSITION PRECAUTIONS TO BE TAKEN IN HANDLING AND STORAGE:
MATERIAL CAS# Keep container tightly closed. Store in cool place and keep away from Heat or Flame.
POLYMER-ORGANOPOLYSILOXANE 68083192 <60 Don’t lay container on its side.
MIXTURE 13983170/131 <20
MINERAL FILLERS 4234 <20 8. PHYSICAL AND CHEMICAL PROPERTIES (based on typical material)
SILICA FILLER 67762907 APPEARANCE & ODOR: Clear odorless paste
BOILING POINT: 500 F
VAPOR PRESSURE at 77 Fmm Hg
3. HAZARDS and PROTECTION IDENTIFICATION
VAPOR DENSITY (air=1) N/A
Personal Protection Recommended: Use local exhaust. Ventilation is required. Use a NIOSH approved respirator to SOLUBILITY IN WATER (By wt): Insoluble
prevent overexposure. Reference 29CFR 1910.134 for Federal standards concerning respiratory protection. Wear SPECIFIC GRAVITY (water=1) 1.23
impervious gloves and protective clothing as required to prevent skin contact. Wear protective goggles to prevent eye
contact. An eyewash and safety shower should be nearby and ready for use. Note: The above information is not intended for use in preparing product specifications.
4. FIRST AID MEASURES 9. STABILITY AND REACTIVITY DATA
EMERGENCY AND FIRST AID PROCEDURES: STABILITY: Stable. Hazardous polymerization will not occur.
SWALLOWING: INCOMPATIBILITY: None known
Induce Vomiting HAZARDOUS COMBUSTION OR DECOMPOSITION PRODUCTS: Carbon Monoxide, Carbon Dioxide, Silicon
SKIN: Dioxide
Wash with soap and water.
INHALATION: 10. TOXICOLOGICAL INFORMATION
Remove to fresh air.. Not available
EYES:
Immediately flush eyes with water and continue washing for at least 15 minutes. If irritation persists seek medical 11. ECOLOGICAL INFORMATION
attention. ECOTOXICOLOGICAL INFORMATION: Keep out of sewage system.
NOTES TO PHYSICIAN:
There is no specific antidote. Treatment of overexposure should be directed at the control of symptoms and the clinical 12. DISPOSAL CONSIDERATIONS
condition of the patient. Use proper landfill disposal or incineration.
QA-MSDS-07 QA-MSDS-07
APPENDIX B: SILICONE MATERIAL DATA
Revised 3/27/2007 Revised 3/27/2007

P.O. Box 1339 P.O. Box 1339
Phone:(928) 537-8387 Phone:(928) 537-8387
Dispose of in accordance with all Federal, State, and local regulations.

MATERIAL SAFETY DATA SHEET
C H I P REGULATIONS A-341 Part B
Chemicals (Hazards Information and Packaging) Regulations 1993 requires physico-chemical and health hazard Soft Gel
determination of all substances and preparations manufactured, transported, stored, modified, or consumed within the
EEC. Components present in this product at a level which could require reporting under the statute are:
****NONE **** Factor II Technology urges each customer or recipient of this MSDS to study it carefully to become aware of and
understand the hazards associated with the product. The reader should consider consulting reference works or
FEDERAL EPA individuals who are experts in ventilation, toxicology, and fire prevention, as necessary or appropriate to the use and
understanding of the data contained in this MSDS.
Comprehensive Environniental Response Compensation and Liability Act of 1980 (CERCLA) requires notification of the
National Response Center of release of quantities of Hazardous Substances equal to or greater than the reportable To promote safe handling each customer or recipient should: (1) notify and furnish its employees, agents, contractors,
quantities (RQ's) in 40 CFR 302.4. Components present in this product at a level which could require reporting under the customers, and others whom it knows or believes will use this material of the information regarding hazards or safety; (2)
statute are: ****NONE **** request its customers to notify their employees, customers and other users of the product of this information.
Superfund Amendments and Reauthorization Act of 1986 (SARA) Title III requires emergency planning based on 1. CHEMICAL PRODUCT
Threshold Planning Quantities (TPQ's) and release reporting based on Reportable Quantities (RQ's) in 40 CFR 355 PRODUCT NAME: Catalyst for RTV Heat Cure Silicone
(used for SARA 302, 304, 3 1, and 312). Components present in this product at a level which could require CHEMICAL NAME: Organopolysiloxane Mixture
reporting under the statute are: **** NONE ****
2. PRODUCT COMPOSITION
Superfund Amendments and Reauthorization Act of 1986 (SARA) Title III requires submission of annual reports of
release of toxic chemicals that appear in 40 CFR 372 (for SARA 3 ~ 3). This information must be included in all MSD S's MATERIAL CAS#
that are copied and distributed for this material. Components present in this product at a level which could require PROPRITARY-ORGANOPOLYSILOXANE 63148607/471 >90
reporting under this statute are: *** NONE**** MIXTURE 341 <10
INVENTORY STATUS
METHYL HYDROGEN SILOXANE 68037592
The ingredients of this product are listed on, or are exempt from listing on, the TSCA inventory.
3. HAZARDS and PROTECTION IDENTIFICATION
13. OTHER INFORMATION
Personal Protection Recommended: Use local exhaust. Ventilation is required. Use a NIOSH approved respirator to
HMIS FORMAT:
prevent overexposure. Reference 29CFR 1910.134 for Federal standards concerning respiratory protection. Wear
Health: 0 Flammability: 0 Reactivity: 0
impervious gloves and protective clothing as required to prevent skin contact. Wear protective goggles to prevent eye
We believe that the information contained herein is current as of the date of this Material Safety Data Sheet, and is contact. An eyewash and safety shower should be nearby and ready for use.
offered in good faith. Since the use of this information and of these opinions and the conditions of the use of the product
are not within the control of Factor II Technology, it is the user's obligation to determine the conditions of safe use of the 4. FIRST AID MEASURES
product. EMERGENCY AND FIRST AID PROCEDURES:
Factor II Technology Regulatory Compliance Department SKIN:
Wash with soap and water.
INHALATION:
Remove to fresh air.
EYES:
Immediately flush eyes with water and continue washing for at least 15 minutes. If irritation persists seek medical
attention.
NOTES TO PHYSICIAN:
There is no specific antidote. Treatment of overexposure should be directed at the control of symptoms and the clinical
condition of the patient.
Revised 3/27/2007
161
Revised 3/27/2007
P.O. Box 1339 P.O. Box 1339
162
Phone:(928) 537-8387 Phone:(928) 537-8387
5. FIRE FIGHTING MEASURES

FLASH POINT: (TCC) N/A C H I P REGULATIONS
FLAMMABLE LIMITS IN AIR (by volume): Not Applicable N/A Chemicals (Hazards Information and Packaging) Regulations 1993 requires physico-chemical and health hazard
determination of all substances and preparations manufactured, transported, stored, modified, or consumed within the
EXTINGUISHING MEDIA: Dry Chemical ,Foam or Carbon Dioxide EEC. Components present in this product at a level which could require reporting under the statute are:
****NONE ****
SPECIAL FIREFIGHTING PROCEDURES: None known.
FEDERAL EPA
UNUSUAL FIRE AND EXPLOXION HAZARDS: None Known
Comprehensive Environniental Response Compensation and Liability Act of 1980 (CERCLA) requires notification of the
6. ACCIDENTAL RELEASE MEASURES National Response Center of release of quantities of Hazardous Substances equal to or greater than the reportable
STEPS TO BE TAKEN IF MATERIAL IS RELEASED OR SPILLED: quantities (RQ's) in 40 CFR 302.4. Components present in this product at a level which could require reporting under the
Contain the spill or leak, scrape up with cardboard or a ra and place in a container. statute are: ****NONE ****
WASTE DISPOSAL METHOD: Dispose of in accordance with all Federal, State, and local regulations.
Superfund Amendments and Reauthorization Act of 1986 (SARA) Title III requires emergency planning based on
7. HANDLING AND STORAGE Threshold Planning Quantities (TPQ's) and release reporting based on Reportable Quantities (RQ's) in 40 CFR 355
PRECAUTIONS TO BE TAKEN IN HANDLING AND STORAGE: (used for SARA 302, 304, 3 1, and 312). Components present in this product at a level which could require
Keep container tightly closed. Store in cool place and keep away from Heat or Flame. reporting under the statute are: **** NONE ****
8. PHYSICAL AND CHEMICAL PROPERTIES (based on typical material) Superfund Amendments and Reauthorization Act of 1986 (SARA) Title III requires submission of annual reports of
APPEARANCE & ODOR: Slight solvent odor release of toxic chemicals that appear in 40 CFR 372 (for SARA 3 ~ 3). This information must be included in all MSD S's
BOILING POINT: 500 F that are copied and distributed for this material. Components present in this product at a level which could require
VAPOR PRESSURE at 77 Fmm Hg reporting under this statute are: *** NONE****
VAPOR DENSITY (air=1) N/A INVENTORY STATUS
SOLUBILITY IN WATER (By wt): Insoluble The ingredients of this product are listed on, or are exempt from listing on, the TSCA inventory.
SPECIFIC GRAVITY (water=1) 1.00
Note: The above information is not intended for use in preparing product specifications. HMIS FORMAT:
Health: 0 Flammability: 0 Reactivity: 0
9. STABILITY AND REACTIVITY DATASTABILITY: Stable. Hazardous polymerization will not occur. We believe that the information contained herein is current as of the date of this Material Safety Data Sheet, and is
INCOMPATIBILITY: None known offered in good faith. Since the use of this information and of these opinions and the conditions of the use of the product
HAZARDOUS COMBUSTION OR DECOMPOSITION PRODUCTS: Carbon Monoxide, Carbon Dioxide, Silicon are not within the control of Factor II Technology, it is the user's obligation to determine the conditions of safe use of the
Dioxide, hYDROGEN product.
INCOMPATIBILITY: avoid contact with acidic bases and oxidizing agents Factor II Technology Regulatory Compliance Department
CONDITIONS TO AVOID: product generates flammable gas on contact with acidic, basic and oxidizing material.
10. TOXICOLOGICAL INFORMATION

Not available
11. ECOLOGICAL INFORMATION

ECOTOXICOLOGICAL INFORMATION: Keep out of sewage system.
12. DISPOSAL CONSIDERATIONS

Use proper landfill disposal or incineration.
Dispose of in accordance with all Federal, State, and local regulations.
Revised 3/27/2007 Revised 3/27/2007

Factor II, Incorporated Factor II, Incorporated
Inventing and Innovating… Inventing and Innovating…
(Information: 1.928.537.8387) (Information: 1.928.537.8387)
In case of emergency call number above In case of emergency call number above
MATERIAL SAFETY DATA SHEET Chronic Effects: This product does not contain any ingredient designated by IARC, NTP, ACGIH or OSHA as probable or
suspected human carcinogens.
LSR 05,40,50,60,70 PART A

4. FIRST AID MEASURES
RHODIA’S V50054, 50007,50008,50009,50010
FIRST AID MEASURES FOR ACCIDENTAL:
1. PRODUCT AND COMPANY DESCRIPTION Eye Exposure:
In case of contact, immediately absorb excess with clean absorbent cloth or cotton. Then, hold eyelids open and flush with
a steady, gentle stream of water for at least 15 minutes. Seek medical attention if irritation develops or persists or if visual
RHODIA INC.
changes occur.
SILICONES
320 West Stanley Avenue
Skin Exposure:
Ventura CA 93001
Immediately wipe excess material off skin with a dry cloth; then wash skin with plenty of soap and water. Seek medical
attention if irritation develops or persists.
Emergency Phone Numbers:
FOR EMERGENCIES INVOLVING A SPILL, LEAK, FIRE, EXPOSURE OR ACCIDENT CONTACT: CHEMTREC (800-
Inhalation:
424-9300 within the United States or 703-527-3887 for international collect calls) or Rhodia CAERS (Communication and
Inhalation is not an expected route of exposure. If respiratory irritation or distress occurs remove victim to fresh air. Seek
Emergency Response System) at 800-916-3232.

medical attention if respiratory irritation or distress continues.
For Product Information: (805) 653-5638
Ingestion:
If victim is conscious and alert, give 1-2 glasses of water to drink. Do not give anything by mouth to an unconscious
Product Status: FDA regulated use only. person. Seek medical attention. Do not leave victim unattended.
Chemical Name or Synonym: REINFORCED DIMETHYL METHYLVINYL SILOXANES MEDICAL CONDITIONS POSSIBLY AGGRAVATED BY EXPOSURE:
No specific information found.
2. COMPOSITION/INFORMATION ON INGREDIENTS NOTES TO PHYSICIAN:
All treatments should be based on observed signs and symptoms of distress in the patient. Consideration should be given
Component CAS Reg Number OSHA Percentage to the possibility that overexposure to materials other than this product may have occurred.
Hazard
DIMETHYL SILICONE ELASTOMER BASE ************ N 100 Treat symptomatically. No specific antidote available.
3. HAZARDS IDENTIFICATION 5. FIRE FIGHTING MEASURES
A. EMERGENCY OVERVIEW:
Physical Appearance and Odor: clear paste-like liquid, odorless. FIRE HAZARD DATA:
Warning Statements: Flash Point: 190 C (374 F). Flammability Class: WILL BURN.
Based on currently available data, this product does not meet the regulatory definition of a hazardous substance. However,
good industrial hygiene practices should be used in handling it. Method Used: Cleveland Open Cup
B. POTENTIAL HEALTH EFFECTS:

Acute Eye: Slightly irritating. May cause redness, irritation.
Acute Skin: Non-irritating. Flammability Limits (vol/vol%): Lower: Upper:
Acute Inhalation: Inhalation not likely. 4 75
Acute Ingestion: Practically non-toxic.
Revised 12/14/2006 Revised 12/14/2006
163
164

Store in tightly closed containers. Store in an area that is clean, well-ventilated, away from ignition sources, away from
Extinguishing Media: Recommended: dry chemical, foam, carbon dioxide, water fog. incompatible materials (see Section 10. Stability and Reactivity).
Special Fire Fighting Procedures: 8. EXPOSURE CONTROLS/PERSONAL PROTECTION

Firefighters should wear NIOSH/MSHA approved self-contained breathing apparatus and full protective clothing. Cool
containers exposed to fire with water.
Unusual Fire and Explosion Hazards: Introductory Remarks:

Product will burn under fire conditions. These recommendations provide general guidance for handling this product. Because specific work environments and
material handling practices vary, safety procedures should be developed for each intended application. While developing
Hazardous Decomposition Materials (Under Fire Conditions): safe handling procedures, do not overlook the need to clean equipment and piping systems for maintenance and repairs.
formaldehyde Waste resulting from these procedures should be handled in accordance with Section 13: Disposal Considerations.
oxides of carbon Assistance with selection, use and maintenance of worker protection equipment is generally available from equipment
silica (crystalline) manufacturers.
This product can form formaldehyde vapors when heated to temperatures above 150 degrees C in the presence of air.
Autoignition Temperature: > 300 C (572 F) Formaldehyde is a potential cancer hazard, a known skin and respiratory sensitizer, and an irritant to the eyes, nose,
throat, skin, and digestive system. Safe handling conditions may be maintained by keeping vapor concentrations within the
OSHA Permissible Exposure Limit for formaldehyde.
6. ACCIDENTAL RELEASE MEASURES
Exposure Guidelines:
No exposure limits were found for this product or any of its ingredients.
Evacuation Procedures and Safety:
Engineering Controls:
Wear appropriate protective gear for the situation. See Personal Protection information in Section 8. CAUTION: Spilled Where engineering controls are indicated by use conditions or a potential for excessive exposure exists, the following
material may make the floor slippery. Do not leave traces of product on floors, ladders, etc., as this may present a slipping traditional exposure control techniques may be used to effectively minimize employee exposures: general area
hazard. dilution/exhaust ventilation.
Containment of Spill: Respiratory Protection:
Follow procedure described below under Cleanup and Disposal of Spill. When respirators are required, select NIOSH/MSHA approved equipment based on actual or potential airborne
concentrations and in accordance with the appropriate regulatory standards and/or industrial recommendations.
Cleanup and Disposal of Spill: For reasonably foreseeable industrial end uses of this material, respiratory protection should not be necessary.
Absorb with an inert absorbent. Scrape up and place in appropriate closed container (see Section 7: Handling and
Storage). Clean up residual material with an appropriate solvent like paint thinner or mineral spirits, provided that there is Eye/Face Protection:
good ventilation and no sources of ignition. Eye and face protection requirements will vary dependent upon work environment conditions and material handling
practices. Appropriate ANSI Z87 approved equipment should be selected for the particular use intended for this material.
Environmental and Regulatory Reporting: It is generally regarded as good practice to wear a minimum of safety glasses with side shields when working in industrial
Do not flush to drain. environments.
7. HANDLING AND STORAGE Skin Protection:

Skin contact should be minimized through use of gloves and suitable long-sleeved clothing (i.e., shirts and pants).
Consideration must be given both to durability as well as permeation resistance.
Minimum/Maximum Storage Temperatures: Work Practice Controls:

< 0 C (32 F) Personal hygiene is an important work practice exposure control measure and the following general measures should be
taken when working with or handling this material:
(1) Do not store, use, and/or consume foods, beverages, tobacco products, or cosmetics in areas where this
material is stored.
Handling: (2) Wash hands and face carefully before eating, drinking, using tobacco, applying cosmetics, or using the toilet.
Avoid breathing vapors and mists. Avoid direct or prolonged contact with skin and eyes. (3) Wash exposed skin promptly to remove accidental splashes or contact with this material.
Storage:
Revised 12/14/2006 Revised 12/14/2006
9. PHYSICAL AND CHEMICAL PROPERTIES Avoid The Following To Inhibit Hazardous Polymerization:
not applicable
Physical and Chemical properties here represent typical properties of this product. Contact the business area using the
Product Information phone number in Section 1 for its exact specifications.
Physical Appearance: clear paste-like liquid. Acute Eye Irritation:

Odor: odorless. No test data found for product.
pH: Not Applicable
Specific Gravity: 1.1 at 25 C (77 F). Acute Skin Irritation:
Water Solubility: insoluble No test data found for product.
Melting Point Range: Not Available
Freezing Point Range: < 0 C (32 F) Acute Dermal Toxicity:
Boiling Point Range: > 260 C (500 F) at 760 mmHg The following data is for the specified ingredients.
Vapor Pressure: < 0.1 mmHg at 25 C (77 F)
Vapor Density: Not Available Acute Respiratory Irritation:
Viscosity: viscosity (centipoises) : 3000 to 4000 cps at 25 C (77 F). No test data found for product.
Acute Inhalation Toxicity:

10. STABILITY AND REACTIVITY No test data found for product.
Acute Oral Toxicity:

The following data is for the specified ingredients.
Chemical Stability:
This material is stable under normal handling and storage conditions described in Section 7. Chronic Toxicity:
This product does not contain any substances that are considered by OSHA, NTP, IARC or ACGIH to be "probable" or
Conditions To Be Avoided: "suspected" human carcinogens.
heat
open flame
spark 12. ECOLOGICAL INFORMATION
Materials/Chemicals To Be Avoided:
strong bases
strong acids Ecotoxicological Information: No data found for product.
strong oxidizing agents Chemical Fate Information: No data found for product.
The Following Hazardous Decomposition Products Might Be Expected:

Decomposition Type: thermal

dimethylcyclosiloxanes Waste Disposal Method:
methylphenylcyclosiloxanes Chemical additions, processing or otherwise altering this material may make the waste management information
presented in this MSDS incomplete, inaccurate or otherwise inappropriate. Please be advised that state and local
Decomposition Type: oxidative/thermal formaldehyde requirements for waste disposal may be more restrictive or otherwise different from federal laws and regulations. Consult
state and local regulations regarding the proper disposal of this material.
Hazardous Polymerization Will Not Occur.
Container Handling and Disposal:
Any containers or equipment used should be decontaminated immediately after use.
Revised 12/14/2006 Revised 12/14/2006
165
166

EPA Hazardous Waste - NO National Fire Protection Association Hazard Ratings--NFPA(R):

1 Health Hazard Rating--Slight
1 Flammability Rating--Slight
14. TRANSPORTATION INFORMATION 0 Instability Rating--Minimal
National Paint & Coating Hazardous Materials Identification System--HMIS(R):

Transportation Status: IMPORTANT! Statements below provide additional data on listed DOT classification. 1 Health Hazard Rating--Slight
The listed Transportation Classification does not address regulatory variations due to changes in package size, mode of 1 Flammability Rating--Slight
shipment or other regulatory descriptors. 0 Reactivity Rating--Minimal
Reason for Revisions:
US Department of Transportation Change and/or addition made to Section 3, Section 8.
Shipping Name: NOT REGULATED
Key Legend Information:
15. REGULATORY INFORMATION ACGIH - American Conference of Governmental Industrial Hygienists
OSHA - Occupational Safety and Health Administration
TLV - Threshold Limit Value
PEL - Permissable Exposure Limit
Inventory Status TWA - Time Weighted Average
Inventory Status STEL - Short Term Exposure Limit
UNITED STATES (TSCA) Y NTP - National Toxicology Program
CANADA (DSL) Y IARC - International Agency for Research on Cancer
EUROPE (EINECS/ELINCS) P ND - Not determined
AUSTRALIA (AICS) Y RPI - Rhodia Established Exposure Limits
JAPAN (MITI) N
SOUTH KOREA (KECL) Y Disclaimer: The information herein is given in good faith but no warranty, expressed or implied, is made.
Y = All ingredients are on the inventory.

E = All ingredients are on the inventory or exempt from listing.
P = One or more ingredients fall under the polymer exemption or are on the no longer polymer list. All other ingredients are
on the inventory or exempt from listing.
N = Not determined or one or more ingredients are not on the inventory and are not exempt from listing.
FEDERAL REGULATIONS
Inventory Issues:
All functional components of this product are listed on the TSCA Inventory.
SARA Title III Hazard Classes:

Fire Hazard - NO
Reactive Hazard - NO
Release of Pressure - NO
Acute Health Hazard - NO
Chronic Health Hazard - NO
STATE REGULATIONS:
This product does not contain any components that are regulated under California Proposition 65.
Revised 12/14/2006 Revised 12/14/2006
MATERIAL SAFETY DATA SHEET Chronic Effects: This product does not contain any ingredient designated by IARC, NTP, ACGIH or OSHA as probable or
suspected human carcinogens.
LSR 05,40,50,& 60,70 PART B

4. FIRST AID MEASURES
RHODIA’S V50054, 50007,50008,50009,50010
FIRST AID MEASURES FOR ACCIDENTAL:
1. PRODUCT AND COMPANY DESCRIPTION Eye Exposure:
In case of contact, immediately absorb excess with clean absorbent cloth or cotton. Then, hold eyelids open and flush with
a steady, gentle stream of water for at least 15 minutes. Seek medical attention if irritation develops or persists or if visual
RHODIA INC.
changes occur.
SILICONES
320 West Stanley Avenue
Skin Exposure:
Ventura CA 93001
Immediately wipe excess material off skin with a dry cloth; then wash skin with plenty of soap and water. Seek medical
attention if irritation develops or persists.
Emergency Phone Numbers:
FOR EMERGENCIES INVOLVING A SPILL, LEAK, FIRE, EXPOSURE OR ACCIDENT CONTACT: CHEMTREC (800-
Inhalation:
424-9300 within the United States or 703-527-3887 for international collect calls) or Rhodia CAERS (Communication and
Inhalation is not an expected route of exposure. If respiratory irritation or distress occurs remove victim to fresh air. Seek
Emergency Response System) at 800-916-3232.

medical attention if respiratory irritation or distress continues.
For Product Information: (805) 653-5638
Ingestion:
If victim is conscious and alert, give 1-2 glasses of water to drink. Do not give anything by mouth to an unconscious
Product Status: FDA regulated use only. person. Seek medical attention. Do not leave victim unattended.
Chemical Name or Synonym: REINFORCED DIMETHYL METHYLVINYL SILOXANES MEDICAL CONDITIONS POSSIBLY AGGRAVATED BY EXPOSURE:
No specific information found.
2. COMPOSITION/INFORMATION ON INGREDIENTS NOTES TO PHYSICIAN:
All treatments should be based on observed signs and symptoms of distress in the patient. Consideration should be given
Component CAS Reg Number OSHA Percentage to the possibility that overexposure to materials other than this product may have occurred.
Hazard
DIMETHYL SILICONE ELASTOMER BASE ************ N 100 Treat symptomatically. No specific antidote available.
3. HAZARDS IDENTIFICATION 5. FIRE FIGHTING MEASURES
A. EMERGENCY OVERVIEW:
Physical Appearance and Odor: clear paste-like liquid, odorless. FIRE HAZARD DATA:
Warning Statements: Flash Point: 190 C (374 F). Flammability Class: WILL BURN.
Based on currently available data, this product does not meet the regulatory definition of a hazardous substance. However,
good industrial hygiene practices should be used in handling it. Method Used: Cleveland Open Cup
B. POTENTIAL HEALTH EFFECTS:

Acute Eye: Slightly irritating. May cause redness, irritation.
Acute Skin: Non-irritating. Flammability Limits (vol/vol%): Lower: Upper:
Acute Inhalation: Inhalation not likely. 4 75
Acute Ingestion: Practically non-toxic.
Revised 12/14/2006 Revised 12/14/2006
167
168

Store in tightly closed containers. Store in an area that is clean, well-ventilated, away from ignition sources, away from
Extinguishing Media: Recommended: dry chemical, foam, carbon dioxide, water fog. incompatible materials (see Section 10. Stability and Reactivity).
Special Fire Fighting Procedures: 8. EXPOSURE CONTROLS/PERSONAL PROTECTION

Firefighters should wear NIOSH/MSHA approved self-contained breathing apparatus and full protective clothing. Cool
containers exposed to fire with water.
Unusual Fire and Explosion Hazards: Introductory Remarks:

Product will burn under fire conditions. These recommendations provide general guidance for handling this product. Because specific work environments and
material handling practices vary, safety procedures should be developed for each intended application. While developing
Hazardous Decomposition Materials (Under Fire Conditions): safe handling procedures, do not overlook the need to clean equipment and piping systems for maintenance and repairs.
formaldehyde Waste resulting from these procedures should be handled in accordance with Section 13: Disposal Considerations.
oxides of carbon Assistance with selection, use and maintenance of worker protection equipment is generally available from equipment
silica (crystalline) manufacturers.
This product can form formaldehyde vapors when heated to temperatures above 150 degrees C in the presence of air.
Autoignition Temperature: > 300 C (572 F) Formaldehyde is a potential cancer hazard, a known skin and respiratory sensitizer, and an irritant to the eyes, nose,
throat, skin, and digestive system. Safe handling conditions may be maintained by keeping vapor concentrations within the
OSHA Permissible Exposure Limit for formaldehyde.
6. ACCIDENTAL RELEASE MEASURES
Exposure Guidelines:
No exposure limits were found for this product or any of its ingredients.
Evacuation Procedures and Safety:
Engineering Controls:
Wear appropriate protective gear for the situation. See Personal Protection information in Section 8. CAUTION: Spilled Where engineering controls are indicated by use conditions or a potential for excessive exposure exists, the following
material may make the floor slippery. Do not leave traces of product on floors, ladders, etc., as this may present a slipping traditional exposure control techniques may be used to effectively minimize employee exposures: general area
hazard. dilution/exhaust ventilation.
Containment of Spill: Respiratory Protection:
Follow procedure described below under Cleanup and Disposal of Spill. When respirators are required, select NIOSH/MSHA approved equipment based on actual or potential airborne
concentrations and in accordance with the appropriate regulatory standards and/or industrial recommendations.
Cleanup and Disposal of Spill: For reasonably foreseeable industrial end uses of this material, respiratory protection should not be necessary.
Absorb with an inert absorbent. Scrape up and place in appropriate closed container (see Section 7: Handling and
Storage). Clean up residual material with an appropriate solvent like paint thinner or mineral spirits, provided that there is Eye/Face Protection:
good ventilation and no sources of ignition. Eye and face protection requirements will vary dependent upon work environment conditions and material handling
practices. Appropriate ANSI Z87 approved equipment should be selected for the particular use intended for this material.
Environmental and Regulatory Reporting: It is generally regarded as good practice to wear a minimum of safety glasses with side shields when working in industrial
Do not flush to drain. environments.
7. HANDLING AND STORAGE Skin Protection:

Skin contact should be minimized through use of gloves and suitable long-sleeved clothing (i.e., shirts and pants).
Consideration must be given both to durability as well as permeation resistance.
Minimum/Maximum Storage Temperatures: Work Practice Controls:

< 0 C (32 F) Personal hygiene is an important work practice exposure control measure and the following general measures should be
taken when working with or handling this material:
(1) Do not store, use, and/or consume foods, beverages, tobacco products, or cosmetics in areas where this
material is stored.
Handling: (2) Wash hands and face carefully before eating, drinking, using tobacco, applying cosmetics, or using the toilet.
Avoid breathing vapors and mists. Avoid direct or prolonged contact with skin and eyes. (3) Wash exposed skin promptly to remove accidental splashes or contact with this material.
Storage:
Revised 12/14/2006 Revised 12/14/2006
9. PHYSICAL AND CHEMICAL PROPERTIES Avoid The Following To Inhibit Hazardous Polymerization:
not applicable
Physical and Chemical properties here represent typical properties of this product. Contact the business area using the
Product Information phone number in Section 1 for its exact specifications.
Physical Appearance: clear paste-like liquid. Acute Eye Irritation:

Odor: odorless. No test data found for product.
pH: Not Applicable
Specific Gravity: 1.1 at 25 C (77 F). Acute Skin Irritation:
Water Solubility: insoluble No test data found for product.
Melting Point Range: Not Available
Freezing Point Range: < 0 C (32 F) Acute Dermal Toxicity:
Boiling Point Range: > 260 C (500 F) at 760 mmHg The following data is for the specified ingredients.
Vapor Pressure: < 0.1 mmHg at 25 C (77 F)
Vapor Density: Not Available Acute Respiratory Irritation:
Viscosity: viscosity (centipoises) : 3000 to 4000 cps at 25 C (77 F). No test data found for product.
Acute Inhalation Toxicity:

10. STABILITY AND REACTIVITY No test data found for product.
Acute Oral Toxicity:

The following data is for the specified ingredients.
Chemical Stability:
This material is stable under normal handling and storage conditions described in Section 7. Chronic Toxicity:
This product does not contain any substances that are considered by OSHA, NTP, IARC or ACGIH to be "probable" or
Conditions To Be Avoided: "suspected" human carcinogens.
heat
open flame
spark 12. ECOLOGICAL INFORMATION
Materials/Chemicals To Be Avoided:
strong bases
strong acids Ecotoxicological Information: No data found for product.
strong oxidizing agents Chemical Fate Information: No data found for product.
The Following Hazardous Decomposition Products Might Be Expected:

Decomposition Type: thermal

dimethylcyclosiloxanes Waste Disposal Method:
methylphenylcyclosiloxanes Chemical additions, processing or otherwise altering this material may make the waste management information
presented in this MSDS incomplete, inaccurate or otherwise inappropriate. Please be advised that state and local
Decomposition Type: oxidative/thermal formaldehyde requirements for waste disposal may be more restrictive or otherwise different from federal laws and regulations. Consult
state and local regulations regarding the proper disposal of this material.
Hazardous Polymerization Will Not Occur.
Container Handling and Disposal:
Any containers or equipment used should be decontaminated immediately after use.
Revised 12/14/2006 Revised 12/14/2006
169
170

EPA Hazardous Waste - NO National Fire Protection Association Hazard Ratings--NFPA(R):

1 Health Hazard Rating--Slight
1 Flammability Rating--Slight
14. TRANSPORTATION INFORMATION 0 Instability Rating--Minimal
National Paint & Coating Hazardous Materials Identification System--HMIS(R):

Transportation Status: IMPORTANT! Statements below provide additional data on listed DOT classification. 1 Health Hazard Rating--Slight
The listed Transportation Classification does not address regulatory variations due to changes in package size, mode of 1 Flammability Rating--Slight
shipment or other regulatory descriptors. 0 Reactivity Rating--Minimal
Reason for Revisions:
US Department of Transportation Change and/or addition made to Section 3, Section 8.
Shipping Name: NOT REGULATED
Key Legend Information:
15. REGULATORY INFORMATION ACGIH - American Conference of Governmental Industrial Hygienists
OSHA - Occupational Safety and Health Administration
TLV - Threshold Limit Value
PEL - Permissable Exposure Limit
Inventory Status TWA - Time Weighted Average
Inventory Status STEL - Short Term Exposure Limit
UNITED STATES (TSCA) Y NTP - National Toxicology Program
CANADA (DSL) Y IARC - International Agency for Research on Cancer
EUROPE (EINECS/ELINCS) P ND - Not determined
AUSTRALIA (AICS) Y RPI - Rhodia Established Exposure Limits
JAPAN (MITI) N
SOUTH KOREA (KECL) Y Disclaimer: The information herein is given in good faith but no warranty, expressed or implied, is made.
Y = All ingredients are on the inventory.

E = All ingredients are on the inventory or exempt from listing.
P = One or more ingredients fall under the polymer exemption or are on the no longer polymer list. All other ingredients are
on the inventory or exempt from listing.
N = Not determined or one or more ingredients are not on the inventory and are not exempt from listing.
FEDERAL REGULATIONS
Inventory Issues:
All functional components of this product are listed on the TSCA Inventory.
SARA Title III Hazard Classes:

Fire Hazard - NO
Reactive Hazard - NO
Release of Pressure - NO
Acute Health Hazard - NO
Chronic Health Hazard - NO
STATE REGULATIONS:
This product does not contain any components that are regulated under California Proposition 65.
Revised 12/14/2006 Revised 12/14/2006
References
[1] A. Jemal, R. Seigel, E. Ward, T. Murray, J. Xu, and M.J. Thun. Cancer statistics
2007. CA: A Cancer Journal for Clinicians, 57:43–66, 2007.
[2] New Zealand Health Information Service. Cancer: New Registrations and Deaths
1999. New Zealand Ministry of Health, 2002.
[3] New Zealand Health Information Service. Cancer Patient Survival Covering the
Period 1994 to 2003. New Zealand Ministry of Health, 2006.
[4] D. Hanahan and R.A. Weinberg. The hallmarks of cancer. Cell, 100:57–70, 2000.
[5] D.B. Kopans. Breast Imaging. Lippincott-Ravens, 1998.
[6] R.R. Seeley, T.D. Stephens, and P. Tate. Anatomy & Physiology. McGraw-Hill,
2003.
[7] S.Z. Ali. An atlas of breast images. www.hopkinsbreastcenter.org/library/

educational_information/mal.shtml, 2007.
[8] V.L. Ernster and J. Barclay. Increases in ductal carcinoma in situ (DCIS) of the
breast in relation to mammography: a dilemma. Journal of the National Cancer
Institute, 22:151–156, 1997.
[9] F.L. Greene, D.L. Page, I.D Fleming, A. Fritz, C.M. Balch, D.G. Haller, and
M. Morrow, editors. AJCC Cancer Staging Manual. Lippincott Raven, 1997.
[10] National Comprehensive Cancer Network. Breast Cancer: Treatment Guidelines

for Patients. 2005.
[11] Canadian Cancer Society. Breast self-examination: What you can do. 2005.
[12] M.A.J. Chaplain, L. Graziano, and L. Preziosi. Mathematical modelling of the

loss of tissue compression responsiveness and its role in solid tumour development.
Mathematical Medicine and Biology, 23:197–229, 2006.
172 REFERENCES
[13] P.A. Newcomb, N.S. Weiss, B.E. Storer, D. Scholes, B.E. Young, and L.F. Voigt.
Breast self-examination in relation to the occurrence of advanced breast cancer.
Journal of the National Cancer Institute, 83:260–265, 1991.
[14] L. Holmberg, A. Ekbom, E. Calle, A. Mokdad, and T. Byers. Breast cancer

mortality in relation to self-reported use of breast self-examination: A cohort
study of 450,000 women. Breast Cancer Research and Treatment, 43:137–140,
1997.
[15] D.B. Thomas, D.L. Gao, S.G. Self, C.J. Allison, Y. Tao, J. Mahloch, R. Ray,
Q. Qin, R. Presley, and P. Porter. Randomized trial of breast self-examination in
Shanghai: methodology and preliminary results. Journal of the National Cancer
Institute, 89:355–365, 1997.
[16] R. Herdman and L. Norton, editors. Saving Women’s Lives: Strategies for Im-
proving Breast Cancer Detection and Diagnosis, Washington DC, 2005. The Na-
tional Academies Press.
[17] V.C. Cokkinides, A. Samuels, E.M. Ward, and M.J. Thun. Cancer Prevention &
Early Detection Facts & Figures. American Cancer Society, 2004.
[18] National Screening Unit. BreastScreen Aotearoa National Policy and Quality
Standards. New Zealand Ministry of Health, 2004.
[19] BreastScreen Aoteoroa. An update on the free national screening program. New
Zealand Ministry of Health, 2004.
[20] L. Tabar, M. Yen, B. Vitak, H.T. Chen, R.A. Smith, and S.W. Duffy. Mammog-
raphy service screening and mortality in breast cancer patients: 20-year follow-up
before and after introduction of screening. The Lancet, 361:1405–1410, 2003.
[21] O. Olsen and P.C. Gtzsche. Cochrane review on screening for breast cancer with
mammography. The Lancet, 358:1340–1342, 2001. Research letter.
[22] S.K. Moore. Better breast cancer detection. IEEE Spectrum, 38:50–54, 2001.
[23] C.T. Derk. Systemic sclerosis and cancer. www.scleroderma.org/medical/

other_articles/derk_2004_3.shtm, 2004.
[24] Partnership for Breast Care. Mammograms. www.partnershipforbreastcare.

org/PBC/problems/ways/mammogram.aspx, 2002.
[25] M.J. Yaffe. What should the burden of proof be for acceptance of a new breast-
cancer screening technique? The Lancet, 364:1111–1112, 2004. Comment.
REFERENCES 173
[26] M. Kriege, C.T.M. Brekelmans, C. Boetes, P.E. Besnard, H.M. Zonderland, I.M.
Obdeijn, R.A. Manoliu, T. Kok, H. Peterse, M.M.A. Tilanus-Linthorst, S.H.
Muller, S. Meijer, J.C. Oosterwijk, L.V.A.M. Beex, R.A.E.M. Tollenaar, H.J.
de Koning, E.J.T. Rutgers, and J.G.M Klijn. Effacy of MRI and mammography
for breast cancer screening in women with familial or genetic disposition. New
England Journal of Medicine, 351:427–437, 2004.
[27] C.A. Beam, P.M. Layde, and D.C. Sullivan. Variability in the interpretation
of screening mammograms by US radiologists. Archives of Internal Medicine,
156:209–213, 1996.
[28] S. Ciatto, M.R. Del Turco, G. Risso, S. Catarzi, R. Bonardi, V. Viterbo,

P. Gnutti, B. Guglielmoni, L. Pinelli, A. Pandiscia, F. Navarra, A. Lauria,
R. Palmiero, and P.L. Indovina. Comparison of standard reading and computer
aided detection (CAD) on a national proficiency test of screening mammography.
European Journal of Radiology, 45:135–138, 2003.
[29] H. Jonsson, L-G. Larsson, and P. Lenner. Detection of breast cancer with mam-
mography in the first screening round in relation to expected incidence in different
age groups. Acta Oncologica, 42:22–29, 2003.
[30] P.R. Snoeren and N. Karssemeijer. Thickness correction of mammographic images

by means of a global parameter model of the compressed breast. IEEE Trans.
on Medical Imaging, 23:799–806, 2004.
[31] R. Sapir, M. Patlas, S.D. Strano, I. Hadas-Halpern, and N.I. Cherny. Does
mammography hurt? Journal of Pain and Symptom Management, 25:53–63,
2003.
[32] A. Asghari and M.K. Nicholas. Pain during mammography: the role of coping
strategies. Pain, 108:170–179, 2004.
[33] S. Kashikar-Zuck, F.J. Keefe, P. Kornguth, P. Beaupre, A. Holzberg, and D. De-

long. Pain coping and the pain experience during mammography: a preliminary
study. Pain, 73:165–172, 1997.
[34] P.J. Kornguth, F.J. Keefe, and M.R. Conaway. Pain during mammography:
characteristics and relationship to demographic and medical variables. Pain,
66:187–194, 1996.
[35] J.P. Hornak. The basics of MRI. http://www.cis.rit.edu/htbooks/mri, 1996.
[36] J. Safir, J.L Zito, M.E. Gershwind, D. Faegenburg, C.E. Tobin, P.D. Cayea, W.J.
Wortman, L.M. Sclafani, and V.E. Maurer. Contrast-enhanced breast MRI for
cancer detection using a commercially available system - a perspective. Clinical
Imaging, 22:162–179, 1998.
174 REFERENCES
[37] Aurora Imaging Technology. MRI Products. www.auroramri.com/mri/product/

biopsy.shtml, 2002.
[38] BOCA Radiology Group. Our experience with breast MRI. www.bocaradiology.
com/Procedures/breast_MRI/index.html, 2007.
[39] E. Warner, D.B. Plewes, R.S. Shumak, G.C. Catzavelos, L.S. Di Prospero, M.J.
Yaffe, V. Goel, P.L. Chart E. Ramsay, D.E.C. Cole, G.A. Taylor, M. Cutrara,
T.H. Samuels, J.P. Murphy, J.M. Murphy, and S.A. Narod. Comparison of breast
magnetic resonance imaging, mammography, and ultrasound for surveillance of
women at high risk for hereditary breast cancer. Journal of Clinical Oncology,
19:3524–3531, 2001.
[40] L. Liberman. Breast cancer screening with MRI - what are the data for patients
at hight risk? New England Journal of Medicine, 351:497–500, 2004. Editorial.
[41] Siemens Medical Imaging. Ultrasound systems. www.siemens.com/medical,

2006.
[42] D. Lister, A.J. Evans, H.C. Burrell, R.W. Blamey, A.R. Wilson, S.E. Pinder,
I.O. Ellis, C.W. Elston, and J. Kollias. The accuracy of breast ultrasound in
the evaluation of clinically benign discrete, symptomatic breast lumps. Clinical
Radiology, 53:490–492, 1998.
[43] R. Lawson. Implications of surface temperatures in the diagnosis of breast cancer.

Canadian Medical Association Journal, 75:309–311, 1956.
[44] W.C. Amalu, W.B. Hobbins, J.F Head, and R.L Elliot. Medical Devices and
Systems, chapter 25. Taylor & Francis, 2006.
[45] Y. Zou and Z. Guo. A review of electrical impedance techniques for breast cancer
detection. Medical Engineering and Physics, 25:79–90, 2003.
[46] V. Cherepenin, A. Karpov, A. Korjenevsky, V. Kornienko, A. Mazaletskaya,

D. Mazourov, and D. Meister. A 3D electrical impedance tomography (EIT)
system for breast cancer detection. Physiological Measurement, 22:9–18, 2001.
[47] N.K. Soni, A. Hartov, C. Kogel, S.P. Poplack, and K.D. Paulsen. Multi-frequency
electrical impedance tomography of the breast: new clinical results. Physiological
Measurement, 25:301–314, 2004.
[48] K.D. Paulsen, P.M. Meaney, and L.C. Gilman, editors. Alternative Breast Imag-
ing: Four Model-Based Approaches. Springer, 2005.
[49] G.G. Senaratne, R.B. Keam, W.L. Sweatman, and G.C. Wake. Inverse methods
for detection of internal objects using microwave technology: With potential for
REFERENCES 175
breast screening. In Proceedings of the 5th International Conference on Inverse

Problems in Engineering: Theory and Practice, pages S1–10. Leeds University
Press, 2005.
[50] J. Ophir, E.I. Céspedes, H. Ponnekanti, Y. Yazdi, and X. Li. Elastography: a

quantitative method for imaging the elasticity of biological tissues. Ultrasonic
Imaging, 13:111–134, 1991.
[51] L. Gao, K.J. Parker, R.M. Lerner, and S.F. Levinson. Imaging the elastic proper-
ties of tissue – a review. Ultrasound in Medicine and Biology, 22:959–977, 1996.
[52] A.P. Sarvazyan, A.R Skovoroda, S.Y. Emelianov, J.B. Fowlkes, J.G. Pipe, R.S
Adler, R.B. Boxton, and P.L Carson. Biophysical bases of elasticity imaging.
Acoustical Imaging, 21:223–240, 1995.
[53] A. Sarvazyan. Mechanical imaging: A new technology for medical diagnostics.

International Journal of Medical Informatics, 49:195–216, 1998.
[54] T.A. Krouskop, T.M. Wheeler, F. Kallel, B.S. Garra, and T. Hall. Elastic moduli
of breast and prostate tissues under compression. Ultrasonic Imaging, 20:260–274,
1998.
[55] P.S. Wellman, R.D. Howe, E. Dalton, and K.A. Kern. Breast tissue stiffness
in compression is correlated to histological diagnosis. Technical report, Harvard
BioRobotics Laboratory, 1999.
[56] A. Samani, J. Bishop, C. Luginbuhl, and D.B. Plewes. Measuring the elas-
tic modulus of ex vivo small tissue samples. Physics in Medicine and Biology,
48:2183–2198, 2003.
[57] A. Samani and D. Plewes. An inverse problem solution for measuring the elastic
modulus of intact ex vivo breast tissue tumours. Physics in Medicine and Biology,
52:1247–1260, 2007.
[58] A. Samani, J. Zubovits, and D. Plewes. Elastic moduli of normal and pathological
human breast tissues: An inversion technique based investigation of 169 samples.
Physics in Medicine and Biology, 52:1565–1576, 2007.
[59] R.M. Lerner, S.R Huang, and K.J. Parker. Sonoelasticity images derived from
ultrasound signals in mechanically vibrated tissues. Ultrasound in Medicine and
Biology, 16:231–239, 1990.
[60] K.J. Parker, S.R. Huang, R.A. Musulin, and R.M. Lerner. Tissue response to
mechanical vibrations for sonoelasticity imaging. Ultrasound in Medicine and
Biology, 16:241–246, 1990.
176 REFERENCES
[61] F. Lee, J.P. Bronson, R.M. Lerner, K.J. Parker, S-R. Huang, and J.B. Roach.
Sonoelasticity imaging: Results in in vitro tissue specimens. Radiology, 181:237–
239, 1991.
[62] J. Ophir, F. Kallel, T. Varghese, E. Konofagou, S.K. Alam, T. Krouskop,

B. Garra, and R. Righetti. Elastography. Comptes Rendus de l’Academie des
Sciences Series IV Physics, 2:1193–1212, 2001.
[63] R.G. Barr. Clinical applications of a real-time elastography technique in breast

imaging. In Proceedings of the 5th International Conference on the Ultrasonic
Measurement and Imaging of Tissue Elasticity, page 51, 2006.
[64] P.E. Barbone and J.C. Bamber. Quantitative elasticity imaging: what can and
cannot be inferred from strain images. Physics in Medicine and Biology, 47:2147–
2164, 2002.
[65] P.E. Barbone and N. Gokhale. Elastic modulus imaging: on the uniqueness and
nonuniqueness of the elastography inverse problem in two dimensions. Inverse
Problems, 20:283–296, 2004.
[66] P.E. Barbone and A.A. Oberai. Elastic modulus imaging: some exact solutions of
the compressible elastography inverse problem. Physics in Medicine and Biology,
52:1577–1593, 2007.
[67] F. Kallel and M. Bertrand. Tissue elasticity reconstruction using linear pertur-
bation method. IEEE Trans. on Medical Imaging, 15:299–313, 1996.
[68] M.M. Doyley, S. Srinivasan, S.A. Pendergrass, Z. Wu, and J. Ophir. Comparative
evaluation of strain-based and model-based modulus elastography. Ultrasound in
Medicine and Biology, 31:787–802, 2005.
[69] J. Bercoff, S. Chaffai, M. Tanter, L. Sandrin, S. Catheline, M. Fink, J.L. Gennis-

son, and M. Meunier. In vivo breast cancer detection using transient elastography.
Ultrasound in Medicine and Biology, 29:1387–1396, 2003.
[70] D.B. Plewes, J. Bishop, A. Samani, and J. Sciarretta. Visualization and quantifi-
cation of breast cancer biomechanical properties with magnetic resonance elas-
tography. Physics in Medicine and Biology, 45:1591–1610, 2000.
[71] R. Muthupillai, P. J. Rossman, D. J. Lomas, J. F. Greenleaf, S. J. Riederer,

and R. L. Ehman. Magnetic resonance elastography by direct visualization of
propagating acoustic strain waves. Science, 269:1854–1857, 1995.
[72] A. Manduca, T.E. Oliphant, M.A. Dresner, D.S. Lake, J.F. Greenleaf, and R.L.
Ehman. Comparitive evaluation of inversion algorithms for magnetic resonance
REFERENCES 177
elastography. In IEEE International Symposium on Biomedical Imaging, pages

997–1000. IEEE, 2002.
[73] E.E.W. Van Houten, M.M. Doyley, F.E. Kennedy, J.B. Weaver, and K.D.
Paulsen. Initial in vivo experience with steady-state subzone-based MR elastog-
raphy of the human breast. Journal of Magnetic Resonance Imaging, 17:72–85,
2003.
[74] E.E.W. Van Houten. Mechanical Property Reconstruction from MR Detected

Harmonic Displacement Data. PhD thesis, Thayer School of Engineering at Dart-
mouth College, 2001.
[75] J.B. Weaver, E.E.W. Van Houten, M.I. Miga, F.E. Kennedy, and K.D. Paulsen.
Magnetic resonance elastography using 3D gradient echo measurements of steady
state motion. Medical Physics, 28:1620–1628, 2001.
[76] M.M. Doyley, J.B. Weaver, E.E.W. Van Houten, F.E. Kennedy, and K.D.
Paulsen. Thresholds for detecting and characterizing focal lesions using steady-
state MR elastography. Medical Physics, 30:495–504, 2003.
[77] S.J. Kirkpatrick and D.D. Duncan. Optical elastography. In Saratov Fall Meeting
2000 Optical Technologies in Biophysics and Medicine II, pages 58–68. Society
of Photo-Optical Instrumentation Engineers, 2001.
[78] S.J. Kirkpatrick and D.D. Duncan. Acousto-optical elastography. In Laser-Tissue

Interaction XII: Photochemical, Photothermal, and Photomechanical, pages 426–
432. Society of Photo-Optical Instrumentation Engineers, 2001.
[79] M.P. Rothney, C.W. Washington, and M.I. Miga. Evaluation of a similarity-
based elastography technique using four similarity metrics. In IEEE International
Symposium on Biomedical Imaging, pages 696–699. IEEE, 2004.
[80] P. Barbone, N. Gokhale, M. Richards, A. Oberai, and M. Doyley. Simultaneous

elastic image registration and elastic modulus reconstruction. In Proceedings of
the 2004 IEEE International Symposium on Biomedical Imaging: From Nano to
Macro, pages 543–546, 2004.
[81] Y. Liu, L.Z. Sun, , and G. Wang. Tomography-based 3D anisotropic elastography

using boundary measurements. IEEE Trans. Medical Imaging, 24:1323–1333,
2005.
[82] J.H. Holland. Adaptation in Natural and Artificial Systems. MIT Press, USA,
1992.
[83] R. Salomon. Evolutionary algorithms and gradient search: Similarities and dif-
ferences. IEEE Trans. on Evolutionary Computation, 2:45–55, 1998.
178 REFERENCES
[84] J. Carnahan and R. Sinha. Nature’s algorithms. IEEE Potentials, 20:21–24, 2001.
[85] E. Mezura-Montes, C.A. Coello Coello, and R. Landa-Becerra. Engineering op-

timization using a simple evolutionary algorithm. In Proc. 15th IEEE Int. Conf.
on Tools with Artificial Intelligence. IEEE, 2003.
[86] A. Konak, D.W. Coit, and A.E. Smith. Multi-objective optimization using genetic
algorithms: A tutorial. Reliability Engineering and System Safety, 91:992–1007,
2006.
[87] S. Kirkpatrick, D.G. Gelatt Jr., and M.P. Vecchi. Optimization by simulated
annealing. Science, 220:671–680, 1983.
[88] R.A. Rutenbar. Simulated annealing algorithms: An overview. Circuits and

Devices Magazine, 5:19–26, 1989.
[89] T. Matsuoka, S. Yamamoto, and M. Takahara. Prediction of structures and

mechanical properties of composites using a genetic algorithm and finite element
method. Journal of Materials Science, 36:27–33, 2001.
[90] R. Olmi, M. Bini, and S. Priori. A genetic algorithm approach to image re-
construction in electrical impedance tomography. IEEE Trans. on Evolutionary
Computation, 4:83–88, 2000.
[91] Y. Zhang, L.O. Hall, D.B. Goldgof, and S. Sarkar. A constrained genetic approach
for computing material property of elastic objects. IEEE Trans. on Evolutionary
Computation, 10:341–357, 2006.
[92] R. Olmi, M. Bini, and S. Priori. A combined FEM/Genetic algorithm for vascular
soft tissue elasticity estimation. Cardiovascular Engineering, 6:93–102, 2006.
[93] F.S. Azar, D.N. Metaxas, and M.D. Schnall. A deformable finite element model of
the breast for predicting mechanical deformations under external perturbations.
Academic Radiology, 8:965–975, 2001.
[94] M.M Doyley, P.M. Meaney, and J. Bamber. Evaluation of an iterative recon-
struction method for quantitative elastography. Physics in Medicine and Biology,
45:1521–1540, 2000.
[95] A.J.H. Hii, C.E. Hann, J.G. Chase, and E.E.W. Van Houten. Fast normalized
cross corelation for motion tracking using basis functions. Computer Methods
and Programs in Biomedicine, 82:144–156, 2006.
[96] S. Wortmann. Soft tissue actuation and imaging-based measurement system.

Technical report, University of Canterbury, 2005.
REFERENCES 179
[97] T.J. Hall, M. Bilgen, M.F. Insana, and T.A. Krouskop. Phantom materials for
elastography. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control, 44:1355–1365, 1997.
[98] Gelita Group. Gelatin production. www.gelita.com, 2004.
[99] J.M. Blackall, D. Rueckert, C.R. Maurer Jr., G.P. Penney, D.L.G. Hill, and D.J.
Hawkes. An image registration approach to automated calibration for freehand
3d ultrasound. In Medical Image Computing and Computer-Assisted Intervention
MICCAI 2000. Springer, 2000.
[100] Dow Corning. The basics of silicon chemistry. www.dowcorning.com, 2007.
[101] J. Steel, J. Fincher, C. Murray, and A. Morrison. Diet soft tissue actuator.
Technical report, University of Canterbury, 2005.
[102] R. Klette, K. Schlüns, and A. Koschan. Computer Vision, Three-Dimensional

Data from Images. Springer-Verlag, Singapore, 1998.
[103] E. Trucco and A. Verri. Introductory Techniques for 3D Computer Vision.

Prentice-Hall, New Jersey, USA, 1998.
[104] A. Milsant and J. Rouzé. Digital image-based elasto-tomography: Motion sensing

system. Technical report, University of Canterbury, 2004.
[105] E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra,

J. Du Croz, A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen.
LAPACK Users’ Guide. Society for Industrial and Applied Mathematics, 1999.
[106] P.R. Amestoy, A. Guermouche, J-Y. L’Excellent, S. Pralet, G. Alessandrini,

A. Morassi, and E. Rosset. Hybrid scheduling for the parallel solution of lin-
ear systems. Parallel Computing, 32:136–156, 2006.
[107] K. Goto and R.A. can de Geijn. Anatomy of high-performance matrix multipli-
cation. ACM Transactions on Mathematical Software (in Press), 2006.
[108] D.W. Marquardt. An algorithm for least-squares estimation of nonlinear param-

eters. SIAM Journal of Applied Mathematics, 11:431–441, 1963.
[109] R.G. Brown, J.G. Chase, and E.E.W Van Houten. Discrete colour-based
euclidean-invariant signatures for feature tracking in a diet breast cancer screen-
ing system. In Proc. SPIE Int. Soc. Opt. Eng. SPIE, February 2007.
[110] J. Zhong, X. Hu, J. Zhang, and M. Gu. Comparison of performance between

different selection strategies on simple genetic algorithms. In Computational In-
telligence for Modelling, Control and Automation, pages 1115–1121. IEEE, 2005.
180 REFERENCES
[111] A. Oberai, N.H. Gokhale, and G.R. Feijóo. Solution of inverse problems in elas-
ticity imaging using the adjoint method. Inverse Problems, 19:297–313, 2003.
[112] G.E. Moore. Cramming more components onto integrated circuits. Electronics,
38:114–117, 1965.
[113] A. Sarvazyan. Moore’s law: past, present and future. IEEE Spectrum, 34:52–59,
1997.
[114] M. Lundstrom. Moore’s law forever? Science, 299:210–211, 2003.
[115] H-U. Berger, C.E. Hann, J.G. Chase, R.L. Broughton, and E.E.W Van Houten.
Boundary element methods in elastography – a first exploratory study. In Proc.
SPIE Int. Soc. Opt. Eng. SPIE, 2007.

Thesis Fulltext

Uploaded by

Copyright:

Available Formats

Thesis Fulltext

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Fulltext

Uploaded by

Copyright:

Available Formats

Digital Image Elasto-Tomography:

A thesis presented for the degree of

1 Introduction and Background 1

2 Proof of Concept Studies 25

3.5.1 Mesh Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4 Phantom Surface Motion Error Studies 69

5 Non-Linear Reconstruction Algorithms 101

5.2.2 Reconstruction Algorithms . . . . . . . . . . . . . . . . . . . . . 113

7 Future Work 147

Appendix A: The Finite Element Method 151

Appendix B: Silicone Material Data 159

2.5 A summary of the reconstruction results from the three-region study. . . 31

3.1 Granules of processed gelatine, a raw ingredient in the first experimental

3.20 The evolution of measured motion fitting techniques. . . . . . . . . . . . 63

4.1 The experimentally observed motion of the homogeneous gelatine phan-

5.1 An example of a gradient descent algorithm. . . . . . . . . . . . . . . . . 103

A.1 A three-dimensional linear tetrahedral element with nodes numbered

5.1 The reconstruction algorithms considered during development of the hy-

5.4 The total number of FE simulations required for each reconstruction

BEM Boundary element methods

C Stiffness contrast parameter

Γ Motion path error

δ Gradient descent step size parameter

Introduction and Background

1.1 Basic Cancer Biology

Sustained Tissue invasion

1.2 Breast Cancer

Figure 1.2 Anatomy of the female breast [6].

Following detection, the diagnosis of breast cancer is categorized based on the

Category Identifier and Description

Systemic treatment options are often used in addition to surgery in an attempt

1.3 Breast Cancer Screening

indicated that screening with mammography leads to a 20–30% reduction in breast

(a) Equipment and patient positioning for a craniocaudal mammogram.

(b) Craniocaudal projection. (c) Mediolateral oblique projection.

having radiological abnormalities, fatigue or lack of attention introduces significant

Computer-aided detection and the use of digital mammography is becoming more

A significant reason for non-compliance with mammography screening programs is

1.3.2 Alternative Breast Imaging Modalities

While not as prevalent as mammography, MRI has seen limited application as a

an image representing the internal tissue structure to be calculated and displayed.

(a) Technician performing a breast ultrasound (b) A color-enhanced ultrasound image

Figure 1.6 Diagnostic ultrasound of the breast [41].

is widespread, with both systems providing physicians with complimentary informa-

Despite the questionable effectiveness of manual palpation, the mechanical property

1.4.1 Mechanical Properties of Soft Tissue

Mechanical property measurement of in vivo or ex vivo human soft tissue is a challeng-

1.4.2 Ultrasound Elastography

In addition to traditional clinical use, ultrasound technology has been implemented as

model-based ultrasound elastography have yielded comparable results to the strain-

1.4.3 Magnetic Resonance Elastography

1.4.4 Novel Elastographic Techniques

1.5 Stochastic Optimization in Elastography

of several independent parameters with different units. Incorporating the naturally

Annealing is a metallurgical process involving several stages of controlled tempera-

Applications of CO in the field of elastic property reconstruction are limited, al-

1.6 Digital Image Elasto-Tomography

Digital Image Elasto-Tomography (DIET) is a novel technique intended to detect and

Breast with high