Image Analysis in Dual Modality Tomography

for Material Classification

I. Basarab-Horwath1, A. T. Daniels, R. G. Green
School of Engineering, Sheffield Hallam University,
City Campus, Sheffield, S1 1WB.
Abstract
A dual modality tomographic system is described for material classification in a simulated multicomponent flow regime. It combines two tomographic modalities, electrical current and light, to image
the interrogated area. Derived image parameters did not allow material classification. PCA analysis
was performed on this data set producing a new parameter set, which allowed material classification.
This procedure reduces the dimensionality of the data set and also offers a pre-processing technique
prior to analysis by another classifier.
Introduction
Dual-modality tomography is a technique that uses two modalities to produce two separate
tomographic images of the same interrogated area or object. Since, in principle, each modality
produces a mapping or image of the distribution of separate or different object properties, dualmodality tomography produces two (complimentary) images of an object which show the distribution of
two separate object properties. This paper describes preliminary work that demonstrates that it is
possible to combine the output from these two modalities to enable classification of the materials
flowing through that interrogated area.
There is also an increasing amount of work into spectroscopic methods. These are best classed as
quasi-multi-modal systems; only one modality is used but at different excitation modes in order to
differentiate between the various components within an object or interrogated area. Multi-spectral work
has been used in remote sensing, medicine and process tomography [1, 2, 3]. One rather interesting
application of this type of work is the use of a range or spectrum of X-ray energies, along with principal
component analysis (PCA), to identify materials in baggage handling [4]. Relatively little work has been
carried out on multi-modal tomography [5, 6, 7, 8, 9] and as such it is difficult to comment on exactly
which is the better method for material classification. In principal, dual modality tomography should
allow classification without the need to characterise each material over a spectrum of excitations for a
single modality. However, there are situations where only one modality can be easily applied e.g.
electrical impedance tomography (EIT) in medical/physiological measurements - and hence
spectroscopic techniques have a larger role in these applications.
An example of multi-component flow is the flow of materials in sewers. Typically, the materials flowing
in sewers, in addition to water, are perspex, plastic, gas bubbles, rubber products and paper. These
materials cannot be distinguished from one another using a single modality other than to determine
whether they are good or bad electrical conductors or to determine the degree to which they allow light
to pass through them. Thus material identification cannot be carried out with a single modality in such
an environment although it is possible to perform rather crude classification. Dual-modality tomography
should, given a suitable choice of modalities, permit the classification of these materials.
Method
The materials chosen for this experimental work were perspex, opaque plastic, paper and rubber [10].
Because the objective of this work was only to demonstrate the principal of dual modality tomography,
the experimental work was carried out in a tank with no movement of the materials. The materials were
in the form of cylinders; a bubble was simulated using an acetate sheet rolled into a cylinder of the
appropriate dimensions and sealed at both ends. The normalised radius of each cylinder of material
relative to the radius of the tank was 0.25, i.e. a normalised area of 0.0625. Measurements were also
made on combinations of these materials - three combinations were chosen: rubber and paper,
perspex and rubber and also perspex and paper. There was therefore a total of eight different materials
or combinations of materials and therefore eight measurements or images for each modality, giving a
total of sixteen images.

Please address all correspondence to Dr. I. Basarab.

1

The two modalities chosen to interrogate a 'model' pipeline or phantom were electrical current to
determine the electrical impedance distribution within the pipe and infra-red (IR) to determine the
'optical' transmission distribution or IR opacity. The electrical impedance system used was a standard
adjacent-pair sixteen electrode system operating at 8 kHz. The IR system consisted of two arrays of
eight LEDs, positioned so as to be mutually orthogonal to each other and to the central axis of the
pipe. The experimental arrangement illustrated in Figure 1 shows a side view of phantom and a plan
view of half the cross-section viewed along the vertical axis of the phantom, both figures show the
position of the electrodes and IR sensors.
Images of the distribution of electrical impedance were produced using a back-projection algorithm
[11]. Similarly, images of the IR transmittance distribution were constructed using a sensitivity
coefficient algorithm [12]. The IR image can be used to accurately place the object or objects in the
pipeline in one or more of 64 square pixels. The images cannot be combined sensibly using logical or
arithmetic functions [10]. One particular method is to treat corresponding pixel values from the two
images as elements in a 2D vector, that is, as if the electrical and IR images were orthogonal to each
other. In fact, each material and material combination did produce its own unique scatter diagram.
Unfortunately, material combinations did not have a scatter plot that was a combination of the scatter
plots of the separate materials. Thus cluster analysis could be applied to the dual modality scatter plots
of individual materials but not to combinations of materials. Also, it was clear that a scatter plot of a
material combination was not a convolution or cross-correlation of the scatter plots of the individual
materials. These effects are due to the non-linearity of a soft modality such as EIT, where the image is
not material independent, that is, not independent of the resistance distribution. Thus the statistics
chosen to analyse the images had to be material invariant. The method chosen to analyse the images
was to first construct the histogram of grey scale amplitudes for both images and then to perform an
analysis on these histograms [1]. Seven measures were calculated for each of the histograms from the
two images; the mean, variance, dispersion, mean square value (or energy), skewness, kurtosis and
median. There was therefore a total of fourteen measures representing the two images for every one of
the eight materials. Again, there was little correlation in terms of material properties between any of
these measures. One method used to look at possible correlations between the measures was to plot
each of the seven measures for each image (a total of eight sets of fourteen numbers; one set for each
material) against the corresponding values of another measure. There appeared to be no material
classification taking place.
Theoretical background
A set of uncorrelated variables that can describe the images better can be obtained from the original
set of measures using principal component analysis (PCA). PCA transforms a data set of several
correlated measurements for each of several individuals or objects into a set of uncorrelated variables.
In most cases the effect is to reduce the dimensionality of a problem and to bring out structure in the
data [13, 14]. PCA has also been used to look at tomographic images from mixing vessels and this has
been correlated with impeller speeds with a view to adjusting impeller speeds [15, 16]. In the work
described here the eight sets (one for each material combination) of fourteen variables (seven
variables per image) used to describe the sixteen images are replaced, after PCA, with a smaller
number of variables that are uncorrelated. Ideally it is desirable to only have two or three resulting
variables since the relationship between these new variables is then relatively easy to understand
graphically.
The singular value decomposition of the data matrix D gives D=URVT, where U and V are both
column-orthogonal matrices and R is a (square) diagonal matrix of singular values [17] of the
data matrix. The normalised transformed data is given by [14]
B=

1 T
U D = VT
R

where B here is an 8 by 8 matrix, each row being the transformed data for a particular material or
combination of materials. The amount of information contained in each column (element) is
related to the magnitude of the corresponding (diagonal) element in the matrix R since the
variance in the data 'explained' by any eigenvector is proportional to the root of the corresponding
eigenvalue. The eigenvector associated with the largest eigenvalue accounts for the maximum
variance in the data in a least square sense and so on in a decreasing fashion for decreasing
values of eigenvalue.

Results
The results of the analysis were stored in the B matrix. The first four eigenvalues, from the R matrix,
were 2.43, 1.38, 1.09, 0.74 and continued to decay virtually exponentially. However, it was interesting
to note that column B1, the first column in the B matrix and which corresponds to the largest
eigenvalue, does not allow sorting or classification of the materials. Combinations of other columns did
however allow classification. Figure 2 shows a plot of the 3rd and 4th columns of the B matrix, referred
to as B3 and B4. These columns correspond to the 3rd and 4th largest eigenvalues of the C matrix. It
can be seen in this projection that there is classification in the B 3, B4 plane. Plastic and rubber
(opaque, non-conducting) lie in the B3<0, B4<0 quadrant while perspex and the bubble lie in the B 4<0,
B3>0 quadrant (transparent, non-conducting). Paper is virtually on the B 4=0 line (B30). Interestingly,
paper will absorb water and becomes weakly conducting and partially transparent. The B 3<0, B4>0
quadrant contains the combinations of perspex/paper and rubber/paper while the perspex/rubber
combination is located in the B 3>0, B4>0 quadrant. It is important to note that these classifications or
groupings are not apparent in the raw data. Figure 3 shows the B 2 and B4 columns plotted against
each other. Here again, it can be seen that the materials have been classified, albeit with a different
classification from that in the B3, B4 plane.
Discussion and Conclusions
This analysis technique has shown that there are some data invariant statistics present in the images
produced by a simple back-projection algorithm. This method of analysis is able to produce
classification - but it has to be shown that these are the same co-ordinates for the same material in
each experiment. Although the hard field modalities will produce the same local image for each
embedded object within the interrogated area this cannot be true by definition for the soft field
modalities. This makes PCA sensitive to the details of the data. There may well be image parameters
that are only material dependent but not 'soft-field' dependant. Unfortunately, PCA is not image
invariant and will thus produce a different set of results for the same materials if the materials are
distributed differently within the interrogated area, i.e. the images are different. However, there is also
a problem in that if each material were represented by a direction in space one would expect
combinations of materials to lie between the pure materials. For example, given a direction for rubber
and another for paper then one would expect a rubber/paper combination to lie on or close to the plane
joining them. It is interesting to note from these results that this is not the case. However, it has been
shown in this work that a factor analysis technique such as PCA performed on image-derived data
reduces the dimensionality of the data set and can be used as a crude classifier of materials. What
these results do indicate is that the classifier statistics are not, counter-intuitively, some linear
combination of the basic image parameters, that is, resistivity/conductivity or IR transparency/opacity.
These results also show that this method of analysis can also be used as a pre-processing technique
prior to analysis by some other classifier, such as neural networks - the results presented here are
linearly separable and as such are amenable to analysis by relatively simple nets.
Acknowledgements
IBH would like to thank Professor D. Barber, University of Sheffield, for invaluable advice as well as
Professor B. W. Jervis and Drs. K. Dutton and R. McLaughlin, Sheffield Hallam University, for useful
discussions.
References
[1] Jain, A K, Fundamentals of Digital Image Processing, Prentice-Hall, 1989, ISBN 0-13-336165-9
[2] Brown, B H Barber, D C, Wei Wang, Liqin Lu, Leathard, A D, Smallwood, R H, Hampshire, A R,
Mackay, R, Hatzigalanis, K, Multi-frequency imaging and modelling of respiratory related electrical
impedance changes, Physiol. Meas., 15, Suppl. 2A, A1-A12, May, 1994
[3] Barlow, R K, Williams, R A, Bond, J, Oliver, K T, Hunt, P, Composition analysis of mixed mineral
suspensions using impedance spectroscopy, pp 131-136, Frontiers in Industrial Process Tomography 2,
April 8-12, 1997, Tech. Univ. Delft, Nederlands.
[4] Maitrejean, S, Perion, D, Sundermann, D, Non-destructive chemical identification using an x-ray
transmission function obtained with the multi-energy method, SPIE Conf. Hard X-Ray and GammaRay Detector Physics and Applications, San Diego, California, July, 1998, SPIE Vol. 3446, pp 134-152.

[5]Johansen, G A, Multi-sensor tomography, in 'Process Tomography - 1995, Implementation for

industrial processes', Proceedings of a Workshop held in Bergen, 6-8 April 1995, Eds. Beck, M S et al,
ISBN 0 9523165 2 8.
[6] Johansen, G A, Froystein, T, Hjertaker, B T, Olsen, O, A dual sensor flow imaging tomographic
system, Meas. Sci. Tech., 7, 3, 297-307, pp March, 1996.
[7] Hjertaker, B T, Static characterisation of a dual sensor flow imaging tomographic system, p 53,
Frontiers in Industrial Process Tomography 2, April 8-12, 1997, Tech. Univ. Delft, Nederlands.
[8] Schwartz, A, Multi-tomographic flame analysis with a schlieren apparatus, Meas. Sci. Tech., 7, 3,
406-413, pp March, 1996
[9] West, R M, Williams R A, Opportunities for Data Fusion in Multi-Modality Tomography, Proc.
1st World Congress on Industrial Process Tomography, Buxton, Greater Manchester 1417 April
1999, pp 195 200
[10] Daniels, A R, An Investigation into the use of dual modality tomography for the monitoring of
constituent volumes in multi-component flows, PhD thesis, Sheffield Hallam University, 1995.
[11] Gadd R, Record P, Rolfe P, A sensitivity region reconstruction algorithm using adjacent drive
current injection strategy, Clin. Phys. Physiol. Meas., 13, Suppl. A, pp 101-5, 1992.
[12] Dugdale, W P, An optical instrumentation system for the imaging of two component flows, PhD
thesis, Bolton Institute of Higher Education, 1994.
[13] Hotelling, H, Analysis of a complex of statistical variables into principal components, J.
Educational Psychology, 24, September, pp 417- 441 and October, pp 498 520, 1933.
[14] Martel, Anne L, The use of factor methods in the analysis of dynamic radionuclide studies,
PhD thesis, University of Sheffield, June 1992.
[15] Ni, X, Simons, S J R, Williams, R A, Jia, X, Use of PCA and neural networks to extract information
from tomographic images for process control, pp 309-314, Frontiers in Industrial Process Tomography
2, April 8-12, 1997, Tech. Univ. Delft, Nederlands.
[16] Tabe H, Simons S J R, Savery, J, West, R M, Williams, R A, Modelling of Multiphase
Processes Using Tomographic Data for Optimisation and Control, Proc. 1st World Congress on
Industrial Process Tomography, Buxton, Greater Manchester, April 14-17, 1999.
[17] Press W H, Teukolsky S A, Vetterling W T, Flannery B P, Numerical Recipes, CUP, 1992, 2 nd
Edition, ISBN 0 521 43064 X.

(a)

(b)
Figure 1: (a) Side view of phantom and (b) plan view of half the cross-section viewed along
the vertical axis of the phantom, both showing the position of the electrodes and IR sensors.

1.0
0.8

perspex/rubber

0.6

B3

0.4

perspex
bubble

0.2
0.0
-0.8

-0.6

-0.4

plastic

-0.2

0.0
-0.2

paper

0.2

0.4

perspex/paper

0.6

rubber/paper

rubber

-0.4

B4

Figure 2. Plot of the two principal components B3 and B4 for each of the eight material
combinations.

0.8
rubber/paper

0.6
paper

0.4
B2

perspex

0.2

bubble

0.0
-0.8

-0.6

-0.4 plastic

-0.2

-0.2

0.0

0.2

perspex/rubber

0.4

0.6

rubber

-0.4

perspex/paper

-0.6
B4

Figure 3. Plot of the two principal components B2 and B4 for each of the eight material
combinations.