2018 Book MedicalImagingSystems
2018 Book MedicalImagingSystems
Vincent Christlein
Joachim Hornegger (Eds.)
Tutorial
LNCS 11111
Medical
Imaging Systems
An Introductory Guide
Lecture Notes in Computer Science 11111
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Andreas Maier Stefan Steidl
•
Vincent Christlein
Joachim Hornegger (Eds.)
Medical
Imaging Systems
An Introductory Guide
Editors
Andreas Maier Vincent Christlein
Lehrstuhl für Mustererkennung Lehrstuhl für Mustererkennung
Friedrich-Alexander-Universität Friedrich-Alexander-Universität
Erlangen-Nürnberg Erlangen-Nürnberg
Erlangen Erlangen
Germany Germany
Stefan Steidl Joachim Hornegger
Lehrstuhl für Mustererkennung Lehrstuhl für Mustererkennung
Friedrich-Alexander-Universität Friedrich-Alexander-Universität
Erlangen-Nürnbergät Erlangen-Nürnbergät
Erlangen Erlangen
Germany Germany
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics
© The Editor(s) (if applicable) and The Author(s) 2018. This book is an open access publication.
Open Access This book is licensed under the terms of the Creative Commons Attribution 4.0 International
License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution
and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and
the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this book are included in the book's Creative Commons license,
unless indicated otherwise in a credit line to the material. If material is not included in the book's Creative
Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use,
you will need to obtain permission directly from the copyright holder.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Cover illustration: Graphical visualization of the Fourier slice theorem. LNCS 11111, p. 154. Used with
permission.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The present book is the result of four years of work that started in Winter 2014/15 and
was finally concluded in Summer 2018. As such, numerous hours of work went into
this manuscript by several authors, who were all affiliated with the Pattern Recognition
Lab of the Friedrich-Alexander-University Erlangen-Nuremberg. I truly appreciate the
dedication and the hard work of my colleagues that led to this final manuscript and,
although many already left the lab to take positions in academia and industry, they still
supported the finalization of this book.
While major parts of the book were already completed in Winter 2016/17, Springer
gave us the opportunity to rework the book with new concepts like the geek boxes and
new figures in order to adapt the book to a broader audience. With the present concepts,
we hope that the book is suited to early-stage undergraduate students as well as stu-
dents who already completed fundamental math classes and want to deepen their
knowledge on medical imaging. We believe, the time to improve the manuscript was
well spent and the final polish gave rise to a textbook with a coherent story line. In
particular, we break with the historical development of the described imaging devices
and present, e. g., magnetic resonance imaging before computed tomography, although
they were developed in opposite order. A closer look reveals that this change of order is
reasonable for didactical purposes: magnetic resonance imaging relies mainly on the
Fourier transform, while computed tomography requires understanding of the Fourier
slice theorem discovered by Johann Radon. These observations then also mend the
apparent historical disorder, as we celebrate Joseph Fourier’s 250th birthday this year
and celebrated the 100th birthday of the Radon transform last year.
We also tried to find many graphical explanations for many of the mathematical
operations such that the book does not require complete understanding of all mathe-
matical details. Yet, we also offer details and references to further literature in the
previously mentioned geek boxes as students in the later semesters also need to be
familiar with these concepts. In conclusion, we hope that we created a useful textbook
that will be accessible to many readers. In order to improve this ease of access further,
we chose to publish the entire manuscript as open access book under Creative Com-
mons Attribution 4.0 International License. Thus, any information in this book can
shared, copied, adapted, or remixed even for commercial purposes as long as the
original source is appropriately referenced and a link to the license is provided.
2 System Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1 Signals and Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Convolution and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Complex Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1 Types of Fourier Transforms . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Convolution Theorem & Properties . . . . . . . . . . . . . . . . . 25
2.4 Discrete System Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Sampling Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.3 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1 Images and Histograms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.1 Images as Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.2 Histograms of Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Image Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1 Window and Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.2 Gamma Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.3 Histogram Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Edge Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Image Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.1 Filtering – Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.2 Linear Shift-invariant Filters in Image Processing . . . . . 44
3.4.3 Nonlinear Filters – the Median Filter . . . . . . . . . . . . . . . 47
3.5 Morphological Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6 Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4 Contents
4 Endoscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1 Minimally Invasive Surgery and Open Surgery . . . . . . . . . . . . . 57
4.2 Minimally Invasive Abdominal Surgery . . . . . . . . . . . . . . . . . . . . 58
4.3 Assistance Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.4 Range Imaging in Abdominal Surgery . . . . . . . . . . . . . . . . . . . . . 63
4.4.1 Stereo Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.4.2 Structured Light . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.4.3 Time-of-Flight (TOF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1 Image Formation in a Thin Lens . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2 Compound Microscope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Bright Field Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.4 Fluorescence Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.5 Phase Contrast Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.6 Quantitative Phase Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.7 Limitation of Light Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.8 Beyond Light Microscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.9 Light Microscopy Beyond the Diffraction Limit . . . . . . . . . . . . . 88
11 Ultrasound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
11.2 Physics of Sound Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
11.2.1 Sound Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
11.2.2 Sound Wave Characteristics at Boundaries . . . . . . . . . . 239
11.2.3 Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
11.3 Image Acquisition for Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . 243
11.3.1 Transducers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
11.3.2 Piezoelectric Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
11.3.3 Spatial Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
11.3.4 Imaging Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
11.4 Safety Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Chapter 1
Introduction
covers the basics of image processing and explains how different image trans-
formations such as edge detection and blurring are implemented as image
filters using convolution.
The following chapters cover examples for imaging devices using stan-
dard optics. In this book, endoscopy and microscopy are discussed as typical
modalities of this genre. Endoscopes, see Chap. 4, were among the first med-
ical imaging devices that were used. Images can be acquired by using long
and flexible optical fibers that are able to transport visible light through the
body of a patient.
Microscopes also use visible light. However, tissue samples or cells have to
be extracted from the body first, e. g., in a biopsy. Then the microscope’s op-
tics are used to acquire images at high magnifications that allow the imaging
of individual cells and even smaller structures. Microscopes and the principles
of optics are described in Chap. 5.
Magnetic resonance imaging (MRI), see Chap. 6 uses electromagnetic
waves to excite water atoms inside the human body. Once the excitation
is stopped, the atoms return to their normal state and by doing so emit
the same electromagnetic radio wave that was used to excite them. This ef-
fect is called nuclear magnetic resonance. Using this effect, an MRI image is
obtained. Fig. 1.1 shows a state-of-the-art MR scanner.
X-ray imaging devices, see Chap. 7, use light of very high energy. However,
the light is no longer visible for the human eye. The higher energy of the light
allows for a deeper penetration of the body. Due to different absorption rates
of X-rays, different body tissues can be distinguished on X-ray images. Tissues
with high X-ray absorption, e. g., bones, become visible as bright structures
in X-ray projection images. Today, X-rays are among the most widely spread
1 Introduction 9
Figure 1.1: MRI is based on nuclear magnetic resonance which does not
involve ionizing radiation. For this reason MRI is often used in pediatric
applications. Image courtesy of Siemens Healthineers AG.
Figure 1.2: X-ray projection images are one of the most wide-spread imaging
modalities. Image courtesy of Siemens Healthineers AG.
10 1 Introduction
Figure 1.3: Modern CT systems allow even scanning of the beating heart.
Image courtesy of Siemens Healthineers AG.
Figure 1.4: An X-ray dark-field setup can be used to reconstruct the ori-
entation of fibers that are smaller than the detector resolution. The image
on the left shows the reconstructed fiber orientation in different layers of a
peanut. The image on the right shows a microscopic visualization of the waist
of the peanut (picture courtesy of ECAP Erlangen).
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 2
System Theory
In the digital age, any medical image needs to be transformed from contin-
uous domain to discrete domain (i. e. 1’s and 0’s) in order to be represented
in a computer. To do so, we have to understand what a continuous and a
discrete signal is. Both of them are handled by systems which will also
be introduced in this chapter. Another fundamental concept is the Fourier
transform as it allows us to represent any time domain signal in frequency
space. In particular, we will find that both representations – time domain
and frequency domain – are equivalent and can be converted into each other.
Having found this important relationship, we can then determine conditions
which will guarantee that also conversion from continuous to discrete domain
and vice versa is possible without loss of information. On the way, we will
introduce several other important concepts that will also find repeated use
later in this book.
2.1.1 Signals
Figure 2.1: A system H{.} with the input signal f (t) and the output signal
g(t).
2.1.2 Systems
for the shift τ . This means that shifting the signal by τ followed by
processing with the system is identical to processing the signal with
the system followed by a shift with τ .
before the input actually arrives in the system. This is especially important
for signals with time as the independent parameter. However, non-causal
systems do not pose a problem for the independent parameter space, e. g.,
image filters that use information from the left and right of a pixel. Geek
Box 2.3 presents examples for the combination of different system properties.
Linear shift-invariant systems are important in practice and have conve-
nient properties and a rich theory. For linear shift-invariant systems, the ab-
stract operator H{·} can be described completely using the impulse response
h (t) (cf. Sec. 2.2.2) or transfer function H (ξ) (cf. Sec. 2.3.2). The impulse
response is combined with the signal by the operation of convolution. This is
sufficient to describe all linear shift-invariant systems.
2.2 Convolution and Correlation 17
a = Re (z) (2.6)
b = Im (z) . (2.7)
a = A cos φ
b = A sin φ
polar coordinates consists of magnitude A and angle φ (cf. Geek Box 2.4). For
system theory, an important property of complex numbers is Euler’s formula
Using this relation, a complex sum of sine and cosine can be expressed con-
veniently using a single exponential function. This leads directly to the ex-
ponential notation of complex numbers z = Aeiφ . We will use the complex
numbers and different notations in Sec. 2.3.
2.2 Convolution and Correlation 19
Description Equation
2.2.2 Convolution
1.4
Input signal f (t)
1.2 Impulse response h(t)
1.0 Output signal g(t)
0.8
0.6
0.4
0.2
0.0
−1 0 1 2 3 4
For the definition of the square function, the Heaviside step function
is useful to shorten the notation
(
0, if t < 0
H (t) = .
1, otherwise
Z ∞
f (t)δ(t − T ) dt = f (T ).
−∞
With the sifting property, the element at t = T can be selected from the
function, which is equivalent to sampling the function at that time point.
The sift property is useful for convolution of an arbitrary function and the
Dirac function.
Z ∞
f (t) ∗ δ(t − T ) = f (τ )δ(t − T − τ ) dτ = f (t − T ) (2.11)
−∞
Figure 2.2: Graphical representation of the Dirac function δ(t). The arrow
symbolizes infinity.
Figure 2.3: Laboratory sifters are used to remove undesired parts from
discrete signals. Sequences of Dirac pulses can be applied in a similar way.
Image courtesy of BMK Wikimedia.
2.2.3 Correlation
Here, we use x̄, ȳ, σx , and σy to denote the respective mean values and
standard deviations. If we assume the standard deviations to be equal
to 1 and the means equal to 0, we arrive at the following equation:
X
r(x, y) = x n · yn
n
0.8 1.75
Fourier coefficients of f(t) Coeff. at 1/2 Original signal f(t) Frequs. 0, 1/2
0.7 Coefficient at 0 (offset) Coeff. at 3/2 1.50 Frequency 0 (offset) Frequs. 0, 1/2, 3/2
0.6 1.25
0.5 1.00
0.4 0.75
0.3 0.50
0.2 0.25
0.1 0.00
0.0 0.25
3 2 1 0 1 2 3 3 2 1 0 1 2 3
(a) Fourier coefficients, weights of trigono- (b) Periodic signal and approximations us-
metric functions approximating the signal ing different numbers of Fourier coefficients
f (t)
f (t) = A · cos(2πξt + ϕ) A, ϕ ∈ R
= a · cos(2πξt) + b · sin(2πξt) a, b ∈ R
= c · e2πiξt + c̄ · e−2πiξt c∈C
In Geek Box 2.7, we show how the parameters a, b, and c are related to A
and ϕ.
A Fourier series (cf. Geek Box 2.8) is used to represent a continuous
signal using only discrete frequencies. As such a Fourier series is able to ap-
proximate any signal as a superposition of sine and cosine waves. Fig. 2.4(b)
shows a rectangular signal of time. The absolute values of its Fourier coeffi-
cients are depicted in Fig. 2.4(a). As can be seen in Fig. 2.4(a), the Fourier
coefficients decrease as the frequency increases. It is therefore possible to ap-
proximate the signal by setting the coefficients to 0 for all high frequencies.
Fig. 2.4(b) includes the approximations for three different choices of sets of
frequencies.
The Fourier series, which works on periodic signals, can be extended to
aperiodic signals by increasing the period length to infinity. The resulting
transform is called continuous Fourier transform (or simply Fourier trans-
form, cf. Geek Box 2.9). Fig. 2.5(b) shows the Fourier transform of a rectan-
gular function, which is identical to the Fourier coefficients at the respective
frequencies up to scaling (see Fig. 2.5(a)).
The counter part to the Fourier series for cases in which time domain is
discrete and the frequency domain is continuous is called the discrete time
Fourier transform (cf. Geek Box 2.10). It forms a step towards the dis-
crete Fourier transform (cf. Geek Box 2.11) which allows us to perform
all previous operations also in a digital signal processing system. In discrete
space, we can interpret the Fourier transform simply as a matrix multiplica-
tion with a complex matrix F
2.3 Fourier Transform 25
Table 2.2: Fourier transforms of popular functions. Here we use the defini-
tion sinc(x) = sin(πx)
πx . Note that a convolution of two rectangular functions
yields a triangular function as F [rect(t) ∗ rect(t)] = sinc2 (ξ).
k = Fn (2.14)
where the signal n and the discrete spectrum k are vectors of complex values.
The inverse operation is then readily found as
n = FHk (2.15)
F{f ∗ g} = F · G (2.24)
F{f · g} = F ∗ G (2.25)
26 2 System Theory
+ =2
= Re 2c · e2πiξt .
2.3 Fourier Transform 27
1 d+T
Z
c[k] = f (t) e−2πitk/T dt k∈Z (2.16)
T d
∞
X
f (t) = c[k] e2πitk/T t∈R (2.17)
k=−∞
The coefficients c[k] and c[−k] together form a shifted cosine wave with
frequency ξ = |k|
T (see Geek Box 2.7). It follows that c[−k] = c[k]:
Fig. 2.5(c) shows the DTFT of a band-limited function and the Fourier
transform. The DTFT is identical to the Fourier transform up to
scaling except that it is periodic with period 1/T .
Fig. 2.5(d) shows the DFT and the Fourier series of a band-limited
signal. The DFT is identical to the Fourier series up to scaling except
that it is periodic with period 1/N .
2.3 Fourier Transform 29
f [0]
f [1]
F [k] = (e0 , e−2πnk/N , . . . , e−2πi(N −1)k/N ) f [2] = ξ >
kn
..
.
f [N − 1]
F [0] ξ>
0
F [1] ξ>1
k= .. = .. n = Fn
. .
F [K − 1] ξ>
K−1
1.6
Fourier transform of f(t) 1.4 Signal f(t)
1.4
1.2
1.2
1.0 1.0
(b) Fourier 0.8 0.8
0.6 0.6
transform
0.4 0.4
0.2 0.2
0.0 0.0
4 3 2 1 0 1 2 3 4 3 2 1 0 1 2 3
2.00
Fourier transform of f(t) 1.4 Bandlimited signal f(t)
1.75 DTFT of 1T f[n] f(n T) = f[n]
1.2
1.50
1.25 1.0
(c) DTFT 1.00 0.8
0.75 0.6
(band-limited)
0.50 0.4
0.25 0.2
0.00 0.0
4 3 2 1 0 1 2 3 4 3 2 1 0 1 2 3
0.6 3.0
Fourier coefficients of f(t) Bandlimited, periodic signal f(t)
0.5 DFT of N f[n] 2.5 f(n T) = fN[n]
f[n] (single period of fN[n])
0.4 2.0
(d) DFT 1.5
0.3
(band-limited 1.0
0.2
and periodic) 0.5
0.1
0.0
0.0
3 2 1 0 1 2 3 3 2 1 0 1 2 3
A similar theorem exists for the DFT. Let C h denote the matrix that per-
forms the convolution with discrete impulse response h , and f be a discrete
input signal. Then system output g is obtained as
g = h ∗ f = C h f = F H H Ff .
1 ξ
Scaling f (at) F
|a| a
2.4.1 Motivation
As already indicated in the introduction, discrete signals and systems are very
important in practice. All signals can only be stored and processed at fixed
discrete time instances in a digital computer. The process of transforming
a continuous time signal to a discrete time signal is called sampling. In
the simplest and most common case, the continuous signal is sampled at
regular intervals, which is called uniform sampling. The current value of the
continuous signal is stored exactly at the time instance where the discrete
32 2 System Theory
0 1 2 3 4 5
Figure 2.7: Sampling a sine signal with a frequency below 2 ξB will cause
aliasing. The reconstructed sine wave shown with blue dashes does not match
the original frequency shown in red.
Formally, the theorem can be derived using the periodicity of the DTFT
(see Fig. 2.5(c)). The DTFT spectrum is a periodic summation of the original
spectrum, and the periodic spectra do not overlap as long as the sampling
theorem is fulfilled. It is therefore possible to obtain the original spectrum
by setting the DTFT spectrum to zero for frequencies larger than B. The
signal can then be reconstructed by applying the inverse Fourier transform.
We refer to [3] for a more detailed description of this topic.
So far, we have not discussed how the actual sampling frequency 2 ξB is de-
termined. Luckily such a band limitation can be found for most applications.
For example, even the most sensitive ears cannot perceive frequencies above
22 kHz. As a result, the sampling frequency of the compact disc (CD) was
determined at 44.1 kHz. For the eye, typically 300 dots per inch in printing
or 300 pixels per inch for displays are considered as sufficient to prevent any
visible distortions. In videos and films, a frame rate of 50 Hz is often used to
diminish flicker. High fidelity devices may support up to 100 Hz.
If the sampling theorem is not respected, aliasing occurs. Frequencies
above the Nyquist frequency are wrapped around due to the periodicity and
appear as lower frequencies. Then, these high frequencies are indistinguish-
able from the true low frequencies. Fig. 2.7 demonstrates this effect visually.
2.4.3 Noise
1.6
1.4 Autocorrelation of white noise Spectrum of white noise
1.4
1.2
1.2
1.0 1.0
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 4 3 2 1 0 1 2 3 4
2.5 Examples
To conclude this chapter, we want to show the introduced concepts of con-
volution and Fourier transform on two example systems. A simple system is
a smoothing filter, that allows only slow changes of the signal. This is called
a low-pass filter. It is an important building block in many applications, for
example to remove high-frequency noise from a signal or to remove signal
parts with high-frequency before down-sampling to avoid aliasing.
The filter coefficients of a low-pass filter are visualized in Fig. 2.9(a). The
low-pass filter has a cutoff frequency of π2 samplerad
and a length of 81 coef-
ficients. The true properties of the low-pass filter are best perceived in the
frequency domain, as displayed in Fig. 2.9(b). Note that the scale of the y-axis
is logarithmic. In this context, values of 0 indicate that the signal can pass
unaltered. Small values indicate that the signal components are damped. In
this example, high frequencies are suppressed by several orders of magnitude.
An ideal low-pass filter is a rectangle in the Fourier domain, i. e., all values
below the cutoff frequency are passed unaltered and all values above are set
2.5 Examples 35
1.0 0
0.8
20
0.6
40
0.4
Amplitude [dB]
0.2 60
0.0 80
0.2
100
0.4
120
0 10 20 30 40 50 60 70 80 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Frequency [rad/sample]
1.0 0
0.8
20
0.6
40
0.4
Amplitude [dB]
0.2 60
0.0 80
0.2
100
0.4
120
0 10 20 30 40 50 60 70 80 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Frequency [rad/sample]
to 0. In our discrete filter, we can only approximate this shape. In the time-
domain, the coefficients are samples of a sinc function, which is the inverse
Fourier transform of a rectangular function in Fourier domain (cf. Tab. 2.2).
The opposite of the low-pass filter is the high-pass filter, shown in Fig. 2.10.
Here, frequencies below the cutoff frequency are suppressed, whereas frequen-
cies above are unaltered. Note that the time domain versions of high- and
low-pass filters are difficult to differentiate.
Finally, we show how a signal with high and low frequency components
is transformed after convolution with a high-pass and a low-pass filter. The
signal in Fig. 2.11 is a sine with additive white noise. Thus, noise is distributed
equally in the whole frequency domain. A large portion of the noise can be
removed by suppressing frequency components where no signal is present.
Consequently, the cutoff frequency of the filters is slightly above the frequency
of the sine function. As a result, the output of the high-pass filter is similar
to the noise and the output of the low-pass filter is similar to the sine. In our
36 2 System Theory
2.5
Input signal f(t)
2.0 High-passed g2(t)
Low-passed g1(t)
1.5
1.0
0.5
0.0
0.5
1.0
2 1 0 1 2 3 4 5
Figure 2.11: Sine signal with additive noise after processing with a low-pass
filter and a high-pass filter.
example, we chose a causal filter which introduces a time delay in the filter
output. A causal filter can only react to past inputs and needs to collect a
certain amount of samples before the filtered result appears at the output.
Further Reading
[1] Ronald N Bracewell. The Fourier transform and its applications. McGraw-
Hill, New York, 1986.
[2] Olaf Dössel. Bildgebende Verfahren in der Medizin. Von der Technik
zur medizinischen Anwendung. Vol. 1. 2000.
[3] Bernd Girod, Rudolf Rabenstein, and Alexander Stenger. Einführung
in die Systemtheorie. Vol. 4. Teubner Stuttgart, 1997.
[4] Andreas Maier. Speech of children with cleft lip and palate: Automatic
assessment. Logos-Verlag, 2009.
[5] Karl Pearson. “Mathematical contributions to the theory of evolution.—
on a form of spurious correlation which may arise when indices are used
in the measurement of organs”. In: Proceedings of the royal society of
london 60.359-367 (1897), pp. 489–498.
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 3
Image Processing
3,000 40,000
30,000
2,000
20,000
ni
ni
1,000 10,000
0 0
0 64 128 192 0 64 128 192
Intensity Intensity
(a) Example grayscale im- (b) Histogram of the ex- (c) Histogram of the example
age. ample image with 256 bins. image with 16 bins.
Figure 3.2: Effect of window and level on a C-arm CT slice: The image
on the left shows a slice of an animal experiment displaying −1000 to 1000
HU. In this range, the materials air, soft tissue, contrast agent, and bone are
displayed. The image on the right shows the same slice in the range from 200
to 500 HU. Now, only contrast agent and bone are visible. Image courtesy of
Stanford University.
In images that have significantly more gray values than 8 bit, semi-automatic
adjustment of the display using window and level functions is common. In
CT, the gray values have known physical properties and allow interpretation
of the material. This effect is diplayed in Fig. 3.2. The image on the left hand
side displays the Hounsfield unit (HU) from −1000 to 1000 (cf. Tab. 8.1).
This range covers the materials air, soft tissue, contrast agent, and bone.
The image on the right shows exactly the same slice. The displayed range is
now from 200 to 500 HU which shows only contrast agent in the heart and
cortical bone. Note that the image was obtained from a C-arm system which
shows significantly lower image quality than conventional CT.
40 3 Image Processing
1
0.010
cdf (i)
0.5
ni
0.005
0 0
0 64 128 192 0 64 128 192
Intensity Intensity
(a) Input image. (b) Normalized histogram (c) CDF of the histogram
of the input image. Note of the input image.
the large empty areas for
high and low intensity val-
ues.
0.015 1
0.010
cdf (i)
0.5
ni
0.005
0 0
0 64 128 192 0 64 128 192
Intensity Intensity
(d) The resulting image (e) The histogram of the (f) CDF of the equalized
after applying a histogram image after histogram histogram, which is now
equalization. equalization. Note how linear.
the intensity values now
occupy a larger range of
values.
f (x + h) − f (x)
f 0 (x) = lim . (3.2)
h→0 h
An image, however, is defined as a function over discrete coordinates, where
the above formula is not defined. In this case, a derivative can be calculated
by using finite differences, an approximation that is similar to the difference
quotient. The simplest derivatives we can calculate is the forward difference:
Comparing the continuous derivative and the forward difference, we can see
that only the meaning of h has changed. While for the derivative of a con-
tinuous function, h goes to 0, it is replaced by the constant spacing (usually
h = 1) in the discrete approximation.
In contrast to the continuous case, many different approximations can
be used for the discrete derivative. Along the same line as for the forward
difference, we can define the backward- and central differences
δx f (x, y) = −1 · f (x − 1, y) + 0 · f (x, y) + 1 · f (x + 1, y)
−1 0 1
Figure 3.6: Construction of the central difference filter from Eq. (3.5).
Filters can be applied to images in order to process them, e. g., for noise
reduction. The transformation of the image is determined by the filter kernel
h(i, j), a rectangular patch of size wk ×hk (the point (0, 0) is assumed to be at
the center of the kernel). At each location in the image, the kernel is applied,
and the resulting value is the new value at the center position of the kernel.
Fig. 3.6 shows the filter kernel corresponding to the central difference along
the direction of the x-axis, and how it is constructed from Eq. (3.5). Some
other possible filter kernels for calculating derivatives are shown in Fig. 3.7.
In mathematical terms, a filter H is treated as an operator that is applied
to the input image
H(f (x, y)) = r(x, y) , (3.7)
where r is the resulting filtered image.
When filters are linear and shift-invariant (cf. Geek Box 3.1), they can be
efficiently calculated by convolution. The filtering of an image f (x, y) with
the filter kernel k(i, j) can then be expressed as
wk hk
2
X 2
X
H{f (x, y)} = f ∗ h = f (x − i, y − j)h(i, j) . (3.8)
wk h
i=− 2 j=− 2k
44 3 Image Processing
−1 0 1 −1 1 −1 1 1 −2 1
(a) Central dif- (b) Forward dif- (c) Backward (d) Second order
ference ference difference
Figure 3.7: The derivative filters corresponding to the finite differences in-
troduced in Sec. 3.3. The red circles mark the support of the filter.
Filters are shift-invariant, when the filter does not change when we
shift it over an image, i. e., the filter is independent of the position in
the image it is applied to
H{(f (x − x0 , y − y0 )} = r(x − x0 , y − y0 )
For small images and kernels, this expression is usually calculated for each
pixel of the filtered image individually. For large images, however, we can use
the property of the Fourier transformation that a convolution in the spatial
domain is a multiplication in the frequency domain (cf. Geek Box 3.2).
We now take a look at some linear shift-invariant filters that are often used
in image processing.
This is the most basic filter. Each element of the kernel has the same value.
In order to prevent a change in the range of intensities, the kernel should be
3.4 Image Filtering 45
In the previous chapter, the Fourier transform for 1-D signals (cf.
Sec. 2.3) was introduced. The transformation can also be defined for
2-D signals, i. e., images. The Fourier transform F (µ, ν) and its inverse
transform for continuous functions f (x, y) is defined by
ZZ ∞
F (µ, ν) = f (x, y)e−2πi(µx+νy) dx dy
−∞
ZZ ∞
f (x, y) = F (µ, ν)e2πi(µx+νy) dµ dν .
−∞
As before in the 1-D case, this transform can also be discretized, which
is needed for transforming images. For an image f of size wf × hf ,
the discrete Fourier transform and the inverse Fourier transform are
given by
wf −1 hf −1
X X
F (µ, ν) = f (x, y)e−2πi(µx/wf +νy/hf )
x=0 y=0
wf −1 hf −1
1 X X
f (x, y) = F (µ, ν)e2πi(µx/wf +νy/hf )
wf hf µ=0 ν=0
The convolution theorem that was introduced for the 1-D Fourier
transform also exists for the 2-D case
F{f ∗ g} = F · G
F{F ∗ G} = f · G
DFT{f ∗ g} = F · G
normalized, which leads to the filter kernel shown in Fig. 3.8. The averaging
(or mean filter or box filter) blurs an image and can be used to reduce noise
in an image. There are, however, better filter choices for this task, as the
averaging filter leads to “ringing” artifacts near edges.
46 3 Image Processing
1 1 1
9 9 9
1 1 1
9 9 9
1 1 1
9 9 9
1 1 1
16 8 16
1 1 1
8 4 8
1 1 1
16 8 16
(a) Gaussian filter kernel (b) Gaussian filter of size (c) Gaussian filter of size
of size 3 × 3 15x15 and σ = 1 applied. 15x15 and σ = 5 applied.
Gaussian filter
A better choice for blurring an image than the averaging filter is the Gaussian
filter. This filter puts a higher weight to values near its center. The values
of the kernel can be calculated by the isotropic, zero-mean 2-D Gaussian
function
i2 +j 2
h(i, j) = N · e− 2σ2 (3.9)
where N is used to normalize the kernel. The parameter σ determines how
strong the image is blurred. A small value puts the emphasis on the central
pixels when applying the filter, whereas for a large value the weights of neigh-
boring pixels are larger. A typical Gaussian filter and the effect of different
values of σ are shown in Fig. 3.9.
3.4 Image Filtering 47
−1 0 1 −1 0 1
−1 0 1 −2 0 2
−1 0 1 −1 0 1
8 6 1 5 2 1 2 5 6 8
input sorted
Figure 3.11: Calculating the median (red circle) of an array. The array is
first sorted. The median then is the middle value of the array.
The Prewitt and Sobel filters are combinations of a blurring and a derivative
filter. In practice, image noise leads to many falsely detected edges, when one
of the derivative filters in Fig. 3.7 is used. It therefore either makes sense to
slightly blur the image before the edges are found (not too much, otherwise
the edges vanish), or to use the Prewitt and Sobel filters.
The Prewitt filter corresponds to blurring the image with the averaging
filter before calculating the derivative. The Sobel filter can be thought of as
a combination of a Gaussian filter and the central derivative. Example filter
kernels are shown in Fig. 3.10(a) and Fig. 3.10(b), respectively.
There exist of course other filters that are not linear and shift-invariant. The
most often used one is the median filter. The median of a set of numbers
48 3 Image Processing
is the value that is at the center of the sorted set (cf. Fig. 3.11) and the
median filter sorts the intensity values within its kernel and places the median
value at the center. It is for example used to reduce salt-and-pepper noise
(sparsely appearing black and white pixels, cf. Fig. 3.12) or in general for
noise reduction as it preserves edges better than the averaging or Gaussian
filter. However, since the median filter is not linear and shift-invariant, it
cannot be applied using convolution and is therefore not as computationally
efficient. In fact, for each location in the image, the surrounding pixels have
to be sorted individually to calculate the median value.
(3.10)
A morphological operator now consists of a structuring element (which is
an image/set) and the operation itself. Structuring elements are sets, where
the values usually have binary values corresponding to back- and foreground:
n o
S = x0 , y 0 , x1 , y 1 , . . . , (3.11)
3.5 Morphological Operators 49
(a) Block (b) Cross (c) Diamond (d) Bar (e) L-shaped
Depending on the goal, different shapes and sizes are used, of which some
common ones are shown in Fig. 3.13. As operations themselves, there are
four basic ones: erosion, dilation, opening, and closing, where the opening
and closing operations actually are compositions of erosion and dilation.
Morphological operators are often applied to binary images, i. e., seg-
mented images, where only foreground and background are distinguished
(cf. Sec. 3.6). Thus, images that only contain pixels that are marked with
a specific value, i. e.,
n o
Fbin = x, y | f (x, y) = 1 (3.12)
The basic idea behind binary morphological operations is, that the struc-
turing element is shifted over the binary image. The image is then trans-
formed by the operation, depending on how the structuring element fits to
the shapes visible in the image, or how it misses them. For all operations, the
structuring element S has to be shifted over the image and we will denote
the structuring element at image coordinates (x, y) as
n o
S(x,y) = x + i, y + j | (i, j) ∈ S . (3.13)
Binary Erosion
Using the set notation for images and structuring elements, erosion of an
image F with the structuring element S can be written as
This means that the eroded image contains foreground values only where the
whole structuring element falls in the foreground region of the image. This
is illustrated in Fig. 3.14 where a rectangular shape is eroded with a cross
shaped structuring element.
50 3 Image Processing
Figure 3.14: An example for binary erosion with a cross shaped structuring
element. (b) shows the structuring element overlaid at several positions. A
green structuring element means that the center of the structuring element
will be foreground in the eroded image, a red structuring element means that
it will be a background pixel. Note that the structuring element is only green
when it is completely covered by the block.
Binary Dilation
Using the set notation, binary dilation of an image F with the structuring
element S can be written as
This means that in the dilated image all pixel locations where the overlap
between the structuring element and the foreground pixels in the image is
not empty will be a foreground pixel. This is illustrated with an example in
Fig. 3.15.
Binary Opening
This operation can be used to remove single noise pixels, or small extensions
of larger structures (cf. Fig. 3.16).
3.5 Morphological Operators 51
Figure 3.15: An example for binary dilation with a cross shaped structuring
element. (b) shows the structuring element overlaid at several positions. A
green structuring element means that the center of the structuring element
will be foreground in the eroded image, a red structuring element means that
is will be a background pixel. Note that the structuring element is green even
when only a part of it overlaps with the block.
(a) Input image with (b) Eroded image with (c) Opened image after
structuring element over- structuring element over- dilating the eroded image.
laid for the erosion. laid for the dilation.
Figure 3.16: An example for opening an image with a block shaped struc-
turing element. The image is first eroded and the result is then dilated. Note
how the thin extension of the input shape at its bottom is removed by the
opening operation.
Binary Closing
(a) Input image with the (b) Dilated image with (c) Closed image after
structuring element over- structuring element over- eroding the dilated image.
laid for the dilation. laid for the erosion.
Figure 3.17: An example for closing an image with a block shaped struc-
turing element. The image is first dilated and then eroded. Note how the gap
between the two blocks is filled by the closing operation.
Closing can be used to fill small gaps in shapes (cf. Fig. 3.17). Note that all
of all morphological operators can also be applied on grayscale images (cf.
Geek Box 3.3).
The only question then is how to choose the threshold value θ. Besides trying
different values for θ by hand, the histogram of the image often contains useful
information on how to choose the threshold. There are also some algorithms
that determine the threshold automatically.
3.6 Image Segmentation 53
Grayscale Erosion
Like in the binary case, the structuring element is again shifted over
the whole image. The eroded image at location (x, y) is now defined
as the minimum in the overlap of the structuring element S at this
position and the image F
Grayscale Dilation
as a function of the threshold value. We can do the same for the mean values
of the intensities belonging to each class
θ
1 X
µ1 (θ) = (i + 1)h(i) (3.24)
P1 (θ) i=0
L−1
1 X
µ2 (θ) = (i + 1)h(i) . (3.25)
P2 (θ)
i=θ+1
10,000
5,000
ni
0
0 64 128 192
Intensity
(a) Input image (b) Histogram of the in- (c) Thresholded image
put image. The thresh-
old value determined with
Otsu’s method is marked
by a red line.
So maximizing σb2 means that the distance between the mean values of fore-
ground and background is maximized. The optimal threshold θ∗ is found
by
θ∗ = argmax σb2 (θ), θ ∈ [0, L). (3.27)
θ
Further Reading
[1] Rafael C. Gonzalez and Richard E. Woods. Digital Image Processing
(3rd Edition). Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 2006.
isbn: 013168728X.
[2] Richard Szeliski. Computer Vision: Algorithms and Applications. 1st.
New York, NY, USA: Springer-Verlag New York, Inc., 2010. isbn:
9781848829343.
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 4
Endoscopy
This chapter points out the key aspects of minimally invasive surgery
with particular focus on abdominal surgery using endoscopes. The compari-
son between minimally invasive and conventional open surgery is illustrated
and several procedures are detailed. Moreover, this chapter introduces the
term Image Guidance and its benefits. Finally, this leads to the motiva-
tion why range imaging is of special interest for the next step of modern
surgery. As endoscopes are more or less regular cameras that are inserted
into the body, we also use this section to introduce the pinhole camera in
Geek Box 4.1, elaborate on fundamental mathematical concepts of projec-
tion in Geek Box 4.2, and introduce homogeneous coordinates and perspec-
tive spaces in Geek Box 4.3. Fig. 4.1 displays an endoscopic image taken in
the sigmoid colon which is the part of the large intestine that is closest to
the rectum.
Figure 4.1: View into the colon using an endoscope. Here, the so-called
sigmoid colon is under investigation for abdominal pain. The inner walls of
the colon appear healthy without any signs of inflammation. Image data
courtesy of University Hospital Erlangen, Germany.
parison to conventional open surgery, this leads to a shorter recovery time for
the patient and thereby to a reduced hospital stay. In recent years, a variety
of minimally invasive alternatives to conventional open surgery have evolved
with special focus on pathologies of the heart and the abdomen. In contrast to
open surgery, the physician has no direct access to the organs or structures
of the human body. On the one hand, this means fewer and smaller scars
and less pain for the patient, but on the other hand, without direct access
the physician has a limited sense of orientation and usually has to rely on
additional imaging techniques. For some medical procedures minimally inva-
sive alternatives are not available as the incision is just too small, e. g., the
removal of larger organs or transplantations. For smaller organs such as the
kidney or gallbladder, laparoscopic interventions are already performed as a
common routine. In terms of operative time, minimally invasive surgery usu-
ally takes longer due to the smaller incision and worse orientation. Both open
and in most cases minimally invasive surgery require anesthesia during the
intervention. Statistical comparison of open and minimally invasive surgery
in terms of quality-of-life shows an overall improved result for laparoscopic
interventions.
The pinhole camera model is a simple yet powerful model that allows
us to describe various effects of the projection process. The model
assumes the the outside world is observed using a pinhole. As such
the outside world is projected onto a screen on the opposite side of
the pinhole. Due to the pinhole, the image of the world is a scaled,
upside-down version of the outside world. Doing so, the model al-
lows us to describe perspective, i. e., that the size of the projection is
dependent on the distance. In medical imaging and computer vision
[2], the pinhole camera model is a widely used assumption. Another
convention that is also commonly used is that the screen is virtually
placed outside the camera for graphical simplification.
Virtual Screen
x ix
x WC
Orthographic
Weak perspective
Perspective
f z
In Geek Box 4.2, we have seen that projective transforms cannot be ex-
pressed by means of linear transforms in a Cartesian space. However,
there exists a class of mathematical spaces in which this is possible.
In order to do so, we virtually extend the dimension of our vector
x 0 = (x0 , y 0 )> by an additional component. This way, we create a new
homogeneous vector
λ · ix
ĩ = λ · iy
λ
with λ 6= 0. Note that i can easily be lifted to homogeneous space by
selecting, e. g. λ = 1. For conversion back to a Cartesian space, one
selects simply the last component of a homogenous vector and divides
all components by it. Then, the last component can be omitted and
the point can be mapped back to the original space. This projective
space has a number of additional properties, e. g., all points are only
equal (denoted by =),
˜ if they map back to the original Cartesian point
and points with 0 in the last component cannot easily be mapped back
to the original space as they lie at infinity. A very good introduction
is given in [2].
Using this powerful concept, we can now return to our previous prob-
lem and observe
f 00
f ·x f ·x x
ix
i= = fz·y = ˜ f · y = 0 f 0y .
iy z z 001 z
Figure 4.3: The da Vinci assistance system (©2014 Intuitive Surgical, Inc)
with the workstation on the left and the distant robotic system on the right.
Camera2
Camera1
line
Base
i1 i2
xWC
Figure 4.4: The illustration on the left hand side describes a stereo vision
setup with two cameras Camera1 and Camera2 that observe a 3-D point xWC .
The projection of this point onto each image plane results in a pixel index
i1 and i2 which can be computed by a simple pinhole camera model. The
right hand side image shows the front view of a stereo endoscope with two
apertures to acquire images from two different field of views (close-up view
courtesy by Prof. Speidel, National Center for Tumor Diseases, Dresden)).
Stereo endoscopes are the most investigated range image acquiring setups in
minimally invasive surgery. The principle is also implemented in the da Vinci
assistance system. Stereo vision describes an intuitive acquisition technique
that is similar to the human vision and depth estimation.
The core concept behind stereo endoscopy is to estimate range information
by observing a scene from two different perspectives. Given a known baseline,
the framework has to detect the 2-D projections of a 3-D point in both image
planes. In theory, using basic trigonometry, the range information of these
points can then be calculated by triangulation, see Fig. 4.4. In practice, both
lines will probably not intersect and minimizing the distance of both lines
will estimate the position of the 3-D point.
The requirements for stereo endoscopy are on the one hand a precisely cal-
ibrated device and on the other hand diversified texture information of the
observed scene. Accuracy is increased with a wider baseline between both
sensors. As this baseline is limited by the diameter of the endoscope, the
improved accuracy has to be gained by calculating the corresponding points
in both images with higher precision. Corresponding points are calculated by
detecting features in both images, e. g., by applying the scale-invariant fea-
ture transform (SIFT) or by computing speeded-up robust features (SURF).
Matching those feature points results in point pairs that correspond to the
same 3-D point in the observed scene. Therefore, the output of a stereo en-
doscope highly depends on the quality of the two images and on the speed
4.4 Range Imaging in Abdominal Surgery 65
and robustness of the feature detection and matching. The estimated range
images of a stereo setup, also called disparity maps, have a scene dependent
image resolution.
The bottleneck of this range image acquisition technique is the calcu-
lation of corresponding feature points. Hardware manufacturers tackle this
problem by increasing image resolution in the sensor domain. This leads to
more details even in almost homogeneous regions in the acquired images, but
also induces more computational effort to calculate features on both images.
Therefore, estimating accurate 3-D range data in real-time is a major issue
in stereo endoscopy. Stereo endoscopes are currently the only commercially
available and CE certified 3-D endoscopes.
Projector
Camera
line
Base
xWC
Figure 4.5: The illustration on the left hand side describes a structured light
setup with one camera and one projector that generates a known pattern
onto the scene. The projection of this point onto the image plane results in
a pixel index i. Together with the projection pattern the 3-D world xWC
can be reconstructed using triangulation. The right hand side image shows a
projection on animal organs.
TOF technology [4] tackles the topic of 3-D reconstruction from a completely
different field of view. Instead of using acquired color images to find any
distinctive structures, reflection characteristics are exploited to physically
measure the distances of the observed scene. The first work to introduce this
concept is found in [6].
The concept behind TOF technology is to measure a frequency modulated
light ray that is sent out by an illumination unit and received by a TOF sen-
sor, see Fig. 4.6. The received sinusoidal signal is sampled at four timestamps
to estimate the phase shift Φ between the emitted and the received signal.
The radial distance d is then computed by:
c Φ
d= · , (4.1)
2fmod 2π
where c denotes the speed of light and fmod the modulation frequency.
As the illumination unit can be realized by an light-emitting diode (LED)
and the sensor is a simple CMOS or CCD chip, production costs of TOF
sensors are rather low. However, due to their novelty compared to stereo vi-
sion, current TOF devices exhibit low data quality and low image resolution.
Besides the range image, most TOF devices provide additional data, e. g.,
photometric data, often denoted as the amplitude image and a binary va-
lidity mask. Due to its measurement technique, TOF setups do not require
4.4 Range Imaging in Abdominal Surgery 67
Illumination unit
ToF sensor
xWC
Figure 4.6: The left hand side illustration describes a TOF setup with a
single camera and an illumination unit that sends out the modulated light.
The reflected signal is then received by the TOF sensor at pixel index i. After
calculating the phase shift, the radial distance is computed. The right hand
side image shows a prototype. The setup includes two separated light sources
for the color and the TOF acquisition.
a baseline between the illumination unit and the measuring sensor, which is
beneficial for the use in minimally invasive surgery.
Besides the low resolution, systematic errors reduce the data quality ex-
tremely. The error sources range from color dependent range measurements
over temperature issues of the devices to flying pixels at object boundaries.
In minimally invasive procedures, two major issues occur. First, multiple re-
flection within the abdominal cavity corrupts TOF measurements. Second,
inhomogeneous illumination caused by the endoscopic optic hinders accurate
range measurements. Still the technology is real-time capable and thereby
also suited for live broadcasting [9].
Further Reading
[1] Sven Haase et al. “ToF/RGB Sensor Fusion for 3-D Endoscopy”. In:
Current Medical Imaging Reviews 9.2 (2013), pp. 113–119.
[2] Richard Hartley and Andrew Zisserman. Multiple view geometry in
computer vision. Cambridge university press, 2003.
[3] K. Kremer. Minimally Invasive Abdominal Surgery. Minimally Invasive
Abdominal Surgery. Georg Thieme Verlag, 2001. isbn: 9783131081919.
[4] R. Lange and P. Seitz. “Solid-state time-of-flight range camera”. In:
Quantum Electronics, IEEE Journal of 37.3 (Mar. 2001), pp. 390–397.
issn: 0018-9197.
68 4 Endoscopy
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 5
Microscopy
We perceive the physical world around us using our eyes, but only down to
a certain limit. Objects with a diameter smaller than 75 µm cannot be recog-
nized by the naked eye, and due to this reason, they remained undiscovered
for the most of human history. Entities which belong to this category include
cells (diameter of 10 µm), bacteria (1 µm), viruses (100 nm), molecules (2
nm), and atoms (0.3 nm)1 . In fact, the importance of these micro/nano en-
tities in almost every aspect of our life cannot be sufficiently appreciated.
Microscopes are the tools which enable us to extend our vision to the micro-
world and, despite the prefix micro- in the name, to the nano-world, too. This
chapter takes the reader through the basic principles of the most widely-used
light microscopy techniques, their advantages, and their inherent limitations.
Further microscope types such as scanning tunneling microscopes or atomic
force microscopes are beyond the focus of this text. In contrast to the pre-
vious chapter, a pinhole projection model is no longer sufficient to explain
1
The diameter measurements given here are for a blood cell, a typical bacterium, an
influenza virus, a DNA molecule, and a uranium atom.
d d0
h
B0
B F O F0
h0
L
T0
f f
Figure 5.1: Image formation in a converging lens for an object whose dis-
tance to the lens is larger than the focal length.
h0 d0
= (5.1)
h d
The same applies for triangles T F B and F OL:
h0 f
= (5.2)
h d−f
f d0
=
d−f d
f d = d0 d − d0 f
f d + d0 f = d0 d
T0
T
h0
h
F O F0
d0
f f
Figure 5.2: Image formation in a converging lens for an object whose dis-
tance to the lens is smaller than the focal length.
h
h0
F0 O F
d0
f f
1 1 1
+ = (5.3)
d d0 f
Eq. (5.3) was derived in this text for real images in a converging lens. Never-
theless, it can be also used for virtual images and/or diverging lenses under
the following sign conventions: 1) d0 is negative when the image is at the ob-
ject side of the lens (similar to the case in Figure 5.2), otherwise it is positive.
2) f is negative for diverging lenses. Moreover, if we add a third sign con-
vention stating that h0 is positive for upright images and negative otherwise,
then Eq. (5.1) and Eq. (5.2) can be generalized to the following form:
h0 f d0
M= =− =− (5.4)
h d−f d
Based on the above-mentioned sign conventions, the magnification M is posi-
tive for upright images and negative for upside-down images. This generaliza-
tion, i. e., Eq. (5.3) and Eq. (5.4), can be proved to be correct by applying the
three rules of geometric image formation and employing triangle similarity for
each specific setup. Moreover, based on Eq. (5.4), the following conclusions
can be drawn:
• The image of an object in a converging lens is magnified (|M | > 1) when
d < 2f , has the same size of the object when d = 2f , and demagnified
(|M | < 1) when d > 2f .
• The image of an object in a diverging lens (f < 0) is demagnified.
fo fe
do d0o de < f e
h Fe
Fo Fo0 Fe0
h0
Objective Go
00
h
Eyepiece
Ge
d0e
Medium (n)
F F0
f f
Figure 5.5: The numerical aperture is determined by θ the half angle of the
maximum light cone and n the refractive index of the medium between lens
and specimen.
5.3 Bright Field Microscopy 75
Eye Eye
Light source
Eyepiece
Eyepiece
Emission filter
Specimen
Condenser Objective
Eq. (5.4). The numerical aperture quantifies the capability of a lens to gather
light. It is defined as follows:
NA = n sin θ, (5.5)
where n is the refractive index of the medium between objective lens and
specimen (nair ≈ 1) and θ is the half angle of the maximum light cone which
the lens can collect (cf. Figure 5.5). Since the image formed by the objective
lens is real, it can be captured by a physical detector. For instance, it can be
recorded by a CCD chip, and hence, the magnified view can be saved as a
digital image which can be further processed by a digital computer.
The principle of compound microscope models the magnification mecha-
nism. Additionally, depending on how the sample is illuminated and which
kind of information is carried by light rays, light microscopes can be further
classified into subcategories: bright field, fluorescence, phase contrast, quanti-
tative phase, and others. In the following sections, more details will be given
about each of the aforementioned microscopic modalities.
Figure 5.7: A microscopic image of a cell culture: The image was acquired
using a Nikon Eclipse TE2000U microscope with a bright field objective of
magnification 10× and NA = 0.3.
(a) A bright field image of Chinese hamster (b) The same scene at the left-hand side
ovary (CHO) cells. but seen under a fluorescent channel. Red
spots indicate dead cells.
schematically shows how this fact can be utilized in a microscopic setup. The
condenser shown in the figure plays the role of concentrating light coming
from a light source at the specimen. The specimen information is encoded in
the intensity of light wave which reaches the objective. Background or the
part of the scene which does not contain dense objects tends to be bright in
the resulting image. This observer impression gave the technique its name.
Bright field setup is the number-one choice whenever minimization of ex-
penditure or implementation difficulties are main concerns. An example of a
bright field image of cells is shown in Figure 5.7.
In clinical routine cells in suspension are only investigated infrequently. In-
stead, the most common investigation techniques for bright field microscopy
are cytology, where cells and their inner structure are investigated, and his-
tology, where the embedding of cells into the surrounding tissue architecture
5.3 Bright Field Microscopy 77
laser
CLE Image of healthy epithelial tissue of the vocal folds (left) and
with squamous cell carcinoma (right). Images courtesy of University
Hospital Erlangen, Germany.
80 5 Microscopy
(a) A bright field image (b) A bright field image (c) A phase contrast im-
dominated by amplitude ob- dominated by phase objects: age of the same scene shown
jects: CHO cells in suspen- adherent ultra-thin CHO in 5.9(b). In comparison to
sion. cells. 5.9(b), cells are clearly vis-
ible, albeit surrounded by
halo artifacts.
Typical light detectors such as CCD chips or retina in our eyes can recog-
nize amplitude variations but they are insensitive to phase distortion. In the
1930s, the Dutch physicist Frits Zernike came up with a brilliant trick for con-
verting the invisible phase shift to a visible amplitude change using an optical
filter. His contribution is the basis for a long-established technique in labora-
tories today known as phase contrast. Figure 5.9(a) shows a bright field image
of a sample dominated by amplitude objects. In this particular example, they
are cells in suspension. Figure 5.9(b) also shows a bright field image, but of
a sample dominated by phase objects. The sample contains ultra-thin adher-
ent cells. In Figure 5.9(c), the same specimen of Figure 5.9(b) is shown, but
under a phase contrast microscope. A considerable improvement in contrast
and information content can be clearly seen in the phase contrast image.
The term U (r) encodes both amplitude change A(r) and phase shift
φ(r) as a complex number, and is thus called complex amplitude of
the wave. Eq. (5.6) is not sufficient to describe a wave unless ψ fulfills
the celebrated wave equation:
∂2ψ
= c2 ∇2 ψ, (5.7)
∂t2
where c is the speed of light in the propagation medium, and ∇2 =
∂2 ∂2 ∂2
∂x2 + ∂y 2 + ∂z 2 is the spatial Laplacian. Assuming that ψ can be
factorized as ψ(r, t) = ψr (r)ψt (t) (which is the case in Eq. (5.6)),
one can derive the time-independent wave equation, also known as
Helmholtz’s equation:
a) Wave
λ
b) Amplitude Object
Absorption
c) Phase Object
Phase Shift
where q (x, y) is the object thickness at (x, y). The product of refrac-
tive index with the geometric length of light path is usually termed
optical path length. In addition, the difference of two optical path
lengths is called optical path difference. Therefore, the numerical value
of phase is interpreted as optical path difference between the object
and the surrounding medium. The constant k is typically ignored.
(a) A defocused bright field image (b) A bright field image of the cell
of the cell culture: ∆z = −15 µm. culture at-focus: ∆z = 0.
(c) A defocused bright field image (d) A quantitative phase map ob-
of the cell culture: ∆z = +15 µm. tained by solving the TIE. The bias
field was partially corrected using a
bias-correction algorithm.
tains two distinct peaks. However, the condition under which the two peaks
are considered distinct, can be defined in several ways. This led to different,
but similar, definitions of the resolving power. According to Rayleigh, it is
given by the radius of Airy disk dmin = dAiry (cf. Figure 5.12(b)). A slightly
different definition, known as Abbe criterion, is given as dmin = 0.5 NA
λ
.
In order to enhance microscopic resolution, one needs to employ light of
shorter wavelength and/or an objective of higher numerical aperture. Using
shorter wavelengths will be considered in the next section. The numerical
aperture, as revealed by Eq. (5.5), is theoretically upper-limited by unity
when air (nair ≈ 1) is the medium between the specimen and the objective. In
order to go beyond this limit, microscope manufacturers designed objectives
which can function when a medium of higher refractive index such as water
(nwater ≈ 1.33) or oil (noil ≈ 1.51) is embedded between the specimen and
the objective. This led to the development of water immersion objectives and
oil immersion objectives.
If we set the wavelength in Eq. (5.14) to the wavelength at the center of the
visible spectrum λvisible ≈ 550 nm and numerical aperture to the theoretical
86 5 Microscopy
Intensity Intensity
dmin
Space Space
dAiry
(a) Airy pattern composed of Airy peak (b) Rayleigh criterion: Two features with
with radius dAiry surrounded by diffrac- distance less than dmin = dAiry will be re-
tion rings. solved as a single feature.
X-ray (10 – 1 nm), hard X-ray (below 1 nm)3 , or electron beams (wavelengths
below 5 pm are achievable). Each wavelength range allows us to explore a
part of the nano-world, but also imposes a new type of challenges for both
microscope manufacturers and users. At the UV wavelengths, glass strongly
absorbs light radiation, and thus, in UV microscopy, the lenses are made of
UV-transparent materials such as quartz. Moreover, at the wavelengths of
X-ray radiation, the refractive index of solid substances is very close to the
refractive index of air. Since the light-focusing performed by a visible-light
lens is inherently a refraction process, these lenses cannot be used to focus
X-ray beams. In fact, in X-ray microscopy, expensive and impractical devices
which are based on diffraction instead of refraction are employed to replace
the typical optical lenses. Electron microscopy utilizes electromagnetic lenses
and cathode rays in order to achieve a drastic improvement in resolution
compared to light microscopy. Unlike ultraviolet and X-ray radiation, cathode
rays, being electron beams of measurable mass and negative charge, do not
belong to the electromagnetic radiation. Therefore, the photon-wave duality,
and hence the conception of wavelength, are not directly applicable. One of
the major contributions which led to the development of electron microscopy
is the theory of Louis de Broglie who stated in his PhD thesis that the
particle-wave duality is also valid for matter. According to de Broglie, the
wavelength of an electron of mass me and speed ce is given by:
ρ
λe = , (5.15)
me · c e
where ρ is Planck constant. As an alternative for reflection in optical lenses,
in electromagnetic lenses, deflection of electron beams by magnetic fields was
exploited to focus the beams. In an electron microscope, similar to a cathode-
ray tube, an electron beam is emitted into vacuum by heating the cathode
and then accelerated by applying a voltage between the cathode and the an-
ode. The speed of the electrons, and hence the wavelength (cf. Eq. (5.15)),
can be controlled by varying the voltage. The first electron microscopes were
very similar from a schematic point of view to bright field microscopes. The
acquired image is based on the specimen absorption of electrons when trans-
mitted into the sample, and hence, they were given the name transmission
electron microscopes. A resolution as high as 0.2 nm is achieved by the trans-
mission electron microscopes. A major limitation of this scheme, however,
is that only very thin samples can be imaged. Scanning electron microscopy
was developed to cope with this difficulty. To do so, a primary electron beam
is focused by an electromagnetic lens on a very small part of the specimen.
This primary beam incites the emission of a secondary electron beam. The
intensity of this secondary beam is recorded. Afterwards, the primary beam
is moved to another part of the specimen, and the same process is applied.
3
X-ray and UV radiation, being a part of the electromagnetic spectrum, belong to
invisible light. The term light microscopy is, however, restricted to visible light in this
text.
88 5 Microscopy
This is repeated so that the entire specimen is scanned in a raster pattern and
the final image is obtained from the recorded values of the secondary beam
intensities. Scanning electron microscopy can be used to image thick samples,
even though it captures only the surface details. In addition, the secondary
beam is accompanied with X-ray emission characteristic to the material which
emitted it. Therefore, it is employed to reveal the chemical composition of
specimens. Both scanning electron microscopy and transmission electron mi-
croscopes work in a vacuum. Consequently, they can be used only for dead
specimens. From this perspective, X-ray and traditional light microscopy are
preferred over electron microscopy. Although X-ray and electron microscopes
provide a considerable improvement of resolution over light microscopes, they
are extremely expensive, require large hardware, and mostly involve compli-
cated sample preparation.
Further Readings
[2] R. Ali et al. “On the Use of Low-Pass Filters for Image Processing with
Inverse Laplacian Models”. In: Journal of Mathematical Imaging and
Vision (2010), pp. 1–10.
[3] Marc Aubreville et al. “Automatic Classification of Cancerous Tissue in
Laserendomicroscopy Images of the Oral Cavity using Deep Learning”.
In: Scientific Reports 7.1 (2017), s41598–017.
[4] Douglas E Chandler and Robert W Roberson. Bioimaging: current con-
cepts in light and electron microscopy. Jones & Bartlett Publishers,
2009.
[5] Guy Cox. Optical imaging techniques in cell biology. CRC Press, 2012.
isbn: 978-1-4398-4825-8.
[6] Louis De Broglie. “The wave nature of the electron”. In: Nobel lectures
12 (1929), pp. 244–256.
[7] Charles A DiMarzio. Optics for Engineers. Crc Press, 2011.
[8] Richard Feynman, Robert Leighton, and Matthew Sands. The Feynman
Lectures on Physics. Second. Vol. 1. Boston: Addison-Wesley, 1963.
[9] Sebastian Foersch et al. “Confocal Laser Endomicroscopy for Diagnosis
and Histomorphologic Imaging of Brain Tumors In Vivo”. In: PloS one
7.7 (July 2012), e41760.
[10] Miguel Goncalves et al. “Value of confocal laser endomicroscopy in the
diagnosis of vocal cord lesions”. In: European Review for Medical and
Pharmacological Sciences 21 (2017), pp. 3990–3997.
[11] J.W. Goodman. Introduction to Fourier Optics. Second. MaGraw-Hill,
1996.
[12] Stefan W Hell. “Microscopy and its focal switch”. In: Nature methods
6.1 (2008), pp. 24–32.
[13] A Hoffman et al. “Confocal laser endomicroscopy: technical status and
current indications”. In: Endoscopy 38.12 (Dec. 2006), pp. 1275–1283.
[14] Alan Lacey. Light Microscopy in Biology: A Practical Approach. Sec-
ond. Oxford University Press, 1999.
[15] Elisabeth Laemmel et al. “Fibered confocal fluorescence microscopy
(Cell-viZio) facilitates extended imaging in the field of microcircula-
tion. A comparison with intravital microscopy.” In: Journal of vascular
research 41.5 (Sept. 2004), pp. 400–411.
[16] A.E. Mirsky and J. Brachet. Cell: Biochemistry, Physiology Morphol-
ogy. Edited by Jean Brachet [and] Alfred E Mirsky. Academic Press,
1959.
[17] Firas Mualla et al. “Automatic Cell Detection in Bright-Field Micro-
scope Images Using SIFT, Random Forests, and Hierarchical Cluster-
ing”. In: IEEE Transactions on Medical Imaging 32.12 (2013), pp. 2274–
2286. doi: 10.1109/TMI.2013.2280380.
[18] Firas Mualla et al. “Unsupervised Unstained Cell Detection by SIFT
Keypoint Clustering and Self-labeling Algorithm”. In: Lecture Notes in
Computer Science, Volume 8675, MICCAI 2014 Proceedings, Part III.
Ed. by P. Golland et al. Boston, MA, USA, 2014, pp. 377–384.
90 5 Microscopy
[19] Firas Mualla et al. “Using the Low-Pass Monogenic Signal Framework
for Cell/Background Classification on Multiple Cell Lines in Bright-
Field Microscope Images”. In: International Journal of Computer As-
sisted Radiology and Surgery 9.3 (2014), pp. 379–386. doi: 10.1007/
s11548-013-0969-5.
[20] Douglas B. Murphy. Fundamentals of Light Microscopy and Electronic
Imaging. John Wiley & Sons, Inc., 2002. isbn: 9780471234296.
[21] Nicolai Oetter et al. “Development and validation of a classification
and scoring system for the diagnosis of oral squamous cell carcinomas
through confocal laser endomicroscopy”. In: Journal of Translational
Medicine 14.1 (July 2016), pp. 1–11.
[22] Frank L. Pedrotti, Leno M. Pedrotti, and Leno S. Pedrotti. Introduction
to Optics. 3rd ed. Benjamin Cummings, 2006. isbn: 0131499335.
[23] Gabriel Popescu. Quantitative Phase Imaging of Cells and Tissues.
McGraw-Hill, 2011.
[24] Simon Schöll et al. “Influence of the phase effect on gradient-based and
statistics-based focus measures in bright field microscopy”. In: Journal
of Microscopy 254.2 (2014), pp. 65–74. doi: 10.1111/jmi.12118.
[25] Michael Reed Teague. “Deterministic phase retrieval: a Green0 s func-
tion solution”. In: JOSA 73.11 (1983), pp. 1434–1441.
[26] Laura Waller, Lei Tian, and George Barbastathis. “Transport of inten-
sity phase-amplitude imaging with higher order intensity derivatives”.
In: Opt. Express 18.12 (June 2010), pp. 12552–12561.
[27] Qiang Wu, Fatima A. Merchant, and Kenneth R. Castleman. Micro-
scope Image Processing. Burlington: Academic Press, 2008.
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 6
Magnetic Resonance
Imaging
Modern MRI systems allow physicians to look inside the body without
the use of ionizing radiation (see Fig. 6.1). They provide excellent soft-tissue
contrast for morphological imaging as well as a range of possibilities for func-
tional imaging, e. g., for visualizing blood flow, tissue perfusion or diffusion
processes. In the following chapter, we will outline the physical fundamentals
of magnetic resonance (MR), concepts for imaging, common pulse sequences
to produce different contrasts as well as some advanced topics related to
speeding up the acquisition and for functional imaging.
Figure 6.1: A modern MRI scanner can provide both morphological and
functional imaging. Image courtesy of Siemens Healthineers AG.
N N N N
+ + + +
S S S S
Figure 6.3: Hydrogen nuclei are used for MRI because of their magnetic
susceptibility and their vast amount in the human body. An intrinsic property
of the hydrogen nuclei is their rotation (spin) which makes them magnetic
along the rotational axis.
B0
M
B0
(a) Random spins (b) Partially aligned (c) Precession
spins
Figure 6.4: Nuclei axes within the body will point randomly (a) until a
strong magnetic field forces their rotation axes to partially align with the
applied magnetic field (b). The accumulated magnetization of all spins M
precesses around B 0 (c).
aligned. The same motion can be observed in reverse with a spinning top,
with the spin axis corresponding to the tilt of the spinning top and the axis
of B 0 corresponding to the direction of gravity2 . As with the oscillation of
the compass needle, the precession causes the emission of RF waves.
We said for the magnetic needle that the frequency of the oscillation is
determined by the strength of the magnetic field and the properties of the
needle and that it remains fixed over time. The same holds for nuclear spin
precession, and the frequency of the precession is
f` = γ · kB 0 k, (6.1)
also known as the Larmor frequency. The gyromagnetic ratio γ is the field
strength dependent ratio for a specific nucleus, which is 42.576 MHz/T for
hydrogen. Using, for instance, a 1.5 T field strength, protons will resonate
with a frequency of about 64 MHz. Please note that precession should not
be confused with spin, being the rotation of a single nucleus around its own
axis.
We will now abstract away from the behavior of individual spins and only
consider the net magnetization M . Analogously to our compass needle, we
will push M out of its stable position, a process which we call excitation, to
cause the emission of RF waves from the body. The direction of M can be
modified through a weaker magnetic field B 1 in a direction orthogonal to B 0
by applying RF waves from a coil in the resonance frequency of M .
In the case of M , this is not as straightforward as it was for the compass
needle. In the 2-D plane of the compass, there is only one choice for the direc-
2
MIT Physics Demo: https://www.youtube.com/watch?reload=9&v=8H98BgRzpOM
6.1 Nuclear Magnetic Resonance (NMR) 95
B1 B1
tion of the second magnetic field, which is orthogonal to the main magnetic
field. M precesses in a 3-D motion around B 0 , so the second magnetic field
B 1 must be orthogonal to B 0 as well as aligned with the current rotation
angle of M . But we can simplify things if we change our point of view. Imag-
ine the precession of M by picturing that the vector M is “attached” to a
turntable which is rotating around B 0 . The motion of M as it is being pushed
out of balance and precessing at the same time seems complicated, as does
the direction of B 1 that needs to be applied, see Fig. 6.5(a). Now if we step
onto the turntable and look at the motion again, it is much simpler. In this
rotating frame of reference, the direction of B 1 is constant, and we cannot
see the precession motion, only the excitation caused by B 1 , see Fig. 6.5(b).
Once the secondary magnetic field B 1 is turned off, the magnetization
vector M will slowly return to the equilibrium position by a process called
relaxation, which is described in the next subsection. During this process,
RF waves are emitted from the body and can be received with coils placed
near the body surface. This is the signal from which MR images are then
generated, as outlined in Sec. 6.2.
This section aims at explaining the concept of relaxation, which is the origin
of contrast between different tissues in the resulting images. Relaxation is the
process that causes the net magnetization to constantly approach equilibrium,
96 6 Magnetic Resonance Imaging
6.1.2.1 Relaxation
Let’s first recap the situation before relaxation starts: The per-voxel magne-
tization vectors are initially aligned with the main magnetic field B 0 , which
we will now assume to be aligned with the z axis of our coordinate system.
As such, each magnetization vector can be split into a longitudinal compo-
nent M z and a transversal component M xy such that M = M xy + M z .
So initially, kM xy k = 0 and kM z k is a positive number dependent on the
number of hydrogen nuclei contained within the voxel. During excitation, an
RF pulse tips the magnetization vector into the transversal plane, such that
kM z k = 0 and kM xy k is maximal. We call such an RF wave a 90 ◦ pulse
because it changes the angle of the magnetization vector by 90◦ .
Once the 90◦ pulse ends, relaxation occurs in the form of two indepen-
dent processes to get back to the equilibrium state. The magnetization vec-
tor recovers its longitudinal component, i. e., kM z k tends towards its origi-
nal magnitude. And, usually much faster, the magnetization vector loses its
transversal component, i. e., kM xy k tends towards 0. These independent pro-
cesses happen on different time scales, meaning that the magnitude of the
magnetization vector is not constant over time. Fig. 6.6 visualizes the tra-
jectory of a magnetization vector during relaxation. The physical reasons for
both relaxation processes are outlined in the following paragraphs.
point in time when the process is considered “finished”. The common wisdom
is that after 5T1 , which corresponds to a 99.3 % recovery, the process is as
good as done. As an example, white matter has a T1 of 510 ms while it is
2500 ms for arterial blood. The magnetization recovery for arterial blood is
plotted in Fig. 6.7(a).
t t
T1 = 2.5 s 5T1 = 12.5 s T2 = 45 ms 5T2 = 225 ms
(a) Longitudinal magnetization recovery (b) Transverse magnetization decay
Figure 6.7: Plots of the recovery of longitudinal magnetization (a) and the
decay of transverse magnetization (b) for arterial blood with T1 = 2.5 s and
T2 = 45 ms. As a rule of thumb, the process is considered completed after five
times the respective time constant, with over 99 % of the longitudinal mag-
netization restored after 5T1 and less than 1 % of transverse magnetization
remaining after 5T2 .
The actual decay happens more quickly, i. e., the received signal in an MRI
acquisition decays faster than predicted by T2 . This is due to imperfections
in the homogeneity of the main magnetic field B 0 , which is related to various
effects like magnetic susceptibilities and magnet manufacture. T2∗ ( T two
”
star“) refers to the relaxation which includes the ideal tissue-dependent relax-
ation due to random interactions between nuclei (T2 ) plus the additional loss
of signal due to field imperfections. Note that T2∗ is also affected by tissue-
dependent magnetic susceptibilities and is always shorter than T2 . However,
there exists a measurement method, called spin echo that can recover the
signal lost through fi eld dependent dephasing of nuclei via refocusing pulses
(see Sec. 6.3.1).
Fig. 6.7 compares the exponential course of T1 recovery and T2 decay,
produced by the signal of arterial blood. Now, we can also fully explain the
3-D visualization of relaxation in Fig. 6.6 which is described by the blue
trajectory: the net magnetization vector M 0 (green) has been completely
tipped into the xy-plane (red vector). As relaxation starts, imagine the length
of the red vector to decrease (M xy ) with an exponential decay defined by T2 ,
while at the same time the length of the green vector (M xy ) grows with a
speed defined by T1 . Above all, the magnetization vector rotates around the
main magnetic field B 0 (here, z-direction) with the field strength dependent
Larmor frequency.
6.1 Nuclear Magnetic Resonance (NMR) 99
6.1.2.2 Contrasts
short TE long TE
short TR T1 weighting —
long TR PD weighting T2 weighting
T1 weighting is determined by a short TE to minimize T2 effects and a short
TR (∼ T1 ) at which the longitudinal magnetization has not yet recovered.
T2 weighting can be achieved by a long TR to reduce T1 impact and a long
TE (∼ T2 ) to allow the differences in T2 decay to appear.
PD weighting is given by a long TR such that the magnetization can reach
equilibrium and measured immediately after the RF pulse (short TE).
The missing combination is long TE and short TR, which would result in
a contrast mixture of T1 and T2 with no clinical use and, additionally, a weak
signal amplitude.
f` (z)
f` (z2 )
f` (z1 )
z1 z2
z
head feet
slice thickness
head. Now modifying the frequency of the excitation RF wave allows a slice-
selective excitation, only nuclei whose Larmor frequency matches that of the
wave will be excited.
If the excitation contains only a single frequency, the corresponding excited
slice will be infinitely thin and not enough nuclei will resonate to produce a
measurable signal. By emitting a wave containing a range of frequencies, the
thickness of a slice can be chosen to provide a good trade-off between spatial
resolution and signal-to-noise ratio. A visualization of this concept is shown
in Fig. 6.8.
We will illustrate this with a 1-D example “image” and later extend the
concept to multiple dimensions. Within our example slice, there are more
hydrogen nuclei toward the boundaries and less in the middle, represented
by the magnitude of the magnetization vectors within each voxel.
Directly after excitation, all spins within the excited slice point in the
same direction. The quantity measurable by our MRI system is the net mag-
netization, the sum of all spin magnetization vectors within the voxels of the
excited slice:
=
P
P
net magnetization = magnetizations within voxels
In the absence of any relaxation, all spins would precess at the Larmor
frequency implied by the magnetic field strength. Assuming a homogeneous
magnetic field, the magnitude of the net magnetization vector would not
change:
=
P
=
P
=
P
..
.
By applying a linear gradient, the precession frequency of spins changes
along the direction of the gradient, i. e., spins to the right rotate faster than
those on the left. Spins are no longer in phase, and the phase shift between
adjacent voxels is dependent on the strength of the applied field. This phase
shift has an influence on the magnitude of the net magnetization vector and
it may even become zero. A graph of the net magnetization magnitude for
different gradient field strengths is shown below, with two example “images”
showing different phase shifts.
6.2 Principles of Magnetic Resonance Imaging 103
=
P
Net magnetization
magnitude
0 =
P
Gradient field
strength
stronger gradient field → larger phase shift
This may seem counterproductive at first sight, but it is the core principle
of spatial encoding. If the hydrogen nuclei distribution of the measured tissue
matches the “pattern” implied by the phase shift, there will be a measurable
net magnetization. The better the match, the higher the magnitude of the
net magnetization will be, as shown here:
pattern:
=
P
no hydrogen no hydrogen
nuclei here nuclei here
pattern:
=
P
We can now see that the image f (x) we want to compute is the
Fourier transform of the measured net magnetization m(k), which is
how it can be reconstructed. In other words, phase encoding performs
a Fourier decomposition of our image.
The concept can also be expanded to n-D and comprises again a Fourier
transform (cf. Geek Box 6.2)
6.2 Principles of Magnetic Resonance Imaging 105
6.2.3 k-space
Having understood the concepts of slice selection and spatial encoding, two
options present themselves for 3-D imaging:
106 6 Magnetic Resonance Imaging
ky
kx
Figure 6.9: Example of a filled 2-D k-space, with grey values encoding
similarity as dark = low and bright = high, with associated phase patterns
for some k-space positions.
The spin echo (SE) pulse sequence is widely used as it allows to regain the
signal loss due to field imperfections. As we discussed earlier in Sec. 6.1.2,
the decay of transverse magnetization after excitation is subject to dephasing
due to random interactions and external field inhomogeneities. Apparently,
we cannot influence the random interactions between nuclei but we can alter
the phase of nuclei using an 180◦ inversion pulse such that dephased nuclei
can end up in phase again. For this to work, constant field imperfections are
assumed which holds in practice for conventional MRI acquisitions. The prin-
ciple of the spin echo sequence is shown in Fig. 6.10 and can be summarized
to:
1. A 90◦ RF pulse flips the magnetization into the transversal plane.
2. Dephasing due to random nuclei interactions and field inhomogeneities
sets in. Some nuclei will spin slightly faster and others slower due to local
field variations. The phases of these nuclei will further diverge over time
such that their magnetic moments will cancel each other, resulting in a
decay of transversal magnetization.
3. After a waiting period of TE/2, a 180◦ pulse inverts the magnetization by
flipping the dephased vectors along the x (or y) axis. In consequence, the
nuclei whose phase trailed behind are now ahead of the main magnetic
moment and vice versa.
4. After another waiting period of TE/2, all magnetic moments have refo-
cused and are in phase again as the faster spins have caught up with the
lower spinning nuclei by this time. Now, a large signal, so called spin echo,
which is of negative sign but with the T2∗ effects removed can be measured.
5. Multiple echos can be formed by repeating steps 2–4 as long as some
signal due to T2 decay is available. Each signal has its own echo time TE1 ,
TE2 , . . . after the 90◦ RF excitation.
The gradient echo (GRE) pulse sequence utilizes partial flips with angles be-
low 90◦ , which allow for faster acquisitions compared to SE sequences. Note
that the acquisition time for a single slice in a typical spin echo sequence is
given by TR · Ny · Nex , where Ny and Nex are the number of phase encoding
steps and excitations, respectively. As these parameters determine the result-
108 6 Magnetic Resonance Imaging
B0 z B0 z B0 z
B1 90◦ B1 B1
y y y
x x x
(a) (b) (c)
◦
90
t
B0 z B0 z B0 z
B1 180◦ B1 B1
y y y
x x x
(d) (e) (f)
◦
180 echo1
Figure 6.10: A 90◦ pulse tips the magnetization into the xy plane (a).
Some nuclei spin slightly faster (dashed) and others slower (dotted) than the
resonance frequency due to local field variations (b). This process continues
and leads to a reduction of measurable signal (green line on timeline) along
B 1 (c). A 180◦ pulse inverts the magnetization vectors at t = 12 TE and the
spins start to rephase again (d). The total magnetization builds up as the
magnetization vectors become in phase (e) and reaches its peak at t = TE
(f). This yields the first echo of the original signal, indicated by the red dot
in the timeline.
ing resolution and SNR, one wishes to reduce the scan time by selecting a
TR as small as possible that still yields enough signal. However, with a typ-
ical 90◦ pulse the longitudinal magnetization M z cannot recover sufficiently
for very short TR. Thus, the trick is to use low-flip angles which tip only
parts of the longitudinal magnetization into the transversal plane such that
enough longitudinal magnetization is available for the next repetition after
a short TR. The flip angle α of the RF pulse directly controls the resulting
magnitude of the transversal magnetization kM xy k = kM 0 k sin α, and the
residual longitudinal magnetization kM z k = kM 0 k cos α, where M 0 is the
initial magnetization.
6.4 Advanced Topics 109
Another difference to the spin echo sequence is that GRE uses no 180◦ refo-
cusing pulse which makes it more susceptible to field inhomogeneities leading
to a T2∗ weighting instead of a T2 weighting.
Also, as there is no 180◦ pulse, the echo in a GRE sequence is formed via
a negative gradient in readout or frequency encoding direction which is an
intended dephasing of the magnetization. The idea behind it is that since
some time for preparations such as the spatial encoding is needed before the
actual signal can be read out, we intentionally delay the time of the peak
signal (echo) to a more convenient time. To this end, the dephasing gradient
has the inverse sign of the readout gradient and is applied in advance for half
of the time of the readout gradient. This ensures that the maximal signal
can be obtained at the half of the readout period, since the positive readout
gradient reverses the effects of dephasing and recalls the signal during the
first half while it gradually dephases again during the second half. This is
where the name gradient recalled echo stems from.
Yet another characteristic of the GRE sequence is the formation of a steady
state. In contrast to spin echo sequences, the GRE can have such a short TR
that the signal decay due to T2∗ is incomplete and some transversal magneti-
zation remains when the next RF pulse follows. In consequence, transversal
magnetization accumulates over a few cycles which is referred to as steady
state. As the steady state may be unfavorable for some applications, the so
called spoiled GRE sequence tries to eliminate the residual transverse magne-
tization. Otherwise its effects will manifest itself in the image contrast. More
on this is subject to further reading [3].
Up to this point, the principles for morphological imaging have been intro-
duced. We will now look at some advanced topics related to speeding up the
acquisition process, suppressing signals from unwanted or enhancing signals
from desired tissue classes as well as methods for functional imaging.
Long acquisition times are a major drawback of MRI systems, with manifold
negative consequences. Patients may experience discomfort, having to spend
extended amounts of time in a narrow space. There is an impact on image
quality, as patient motion during the acquisition is inevitable. And from a
financial standpoint, the amount of MRI examinations per unit time is much
less than for other modalities.
110 6 Magnetic Resonance Imaging
Figure 6.11: Parallel imaging uses local receiver coils placed around the
patient which capture the signal from multiple positions, allowing for a re-
duction of acquisition time. The coils are embedded in the table as well as in
the flexible coil on top of the patient’s chest, in this case an 18-element body
coil where individual coil elements are arranged in 6 columns and 3 rows.
Three individual coil elements are exemplarily indicated by the red circles.
Image courtesy of Siemens Healthineers AG.
Techniques for shortening the acquisition time are, thus, of vital impor-
tance and an active field of research. Parallel imaging is an established tech-
nique which allows a reduction of the amount of data needed to be acquired
in order to reconstruct an image. The name stems from the use of multiple
local receiver coils which are placed around the patient and acquire resonance
signals in parallel. Modern MRI systems have coil elements which are em-
bedded in the table as well as flexible coil elements which can be placed on
the patient (see Fig. 6.11).
This enables many possibilities for undersampling in k-space to reduce
the acquisition time, one of which we will present here. Suppose we regularly
undersample the k-space by acquiring only every n-th line. The value n is
referred to as the undersampling factor, indicating the reduction in acquisi-
tion time. A standard reconstruction by performing a Fourier transform is
no longer possible in this case because it will introduce aliasing artifacts in
the reconstructed image (see Fig. 6.12).
A reconstruction technique called sensitivity encoding (SENSE) is able to
reconstruct an image without aliasing artifacts from such undersampled data
by employing the information collected from the multiple receiver coils. It
exploits the fact that each local receiver coil “sees” a slightly different image,
namely one in which parts closer to the coil are better represented than those
6.4 Advanced Topics 111
Figure 6.12: A visualization of a fully sampled k-space (a) with its cor-
responding reconstruction (b) as well as a regularly undersampled k-space
with every second line missing (c) with a reconstruction showing aliasing
artifacts (d).
farther away. A so-called coil sensitivity map sγ (r ρ ) describes how well a coil
γ sees a pixel at position r ρ .
In case of regular undersampling, each pixel in the aliased image (see
Fig. 6.12(d)) can be described as a weighted sum of several pixels in the
unaliased image (see Fig. 6.12(b)). In the given example, 2 pixels in the una-
liased image contribute to a pixel in the aliased image. Different local receiver
coils “see” different aliased images, where the weights for the weighted sum
of pixels are described by the coil sensitivity map of the respective coil. Geek
Box 6.3 illustrates this for an idealized 2-coil setup whereas Geek Box 6.4
discusses how to calibrate the sensitivity maps online.
112 6 Magnetic Resonance Imaging
iγ (r ρ )
sγ (r ρ ) = qP , (6.11)
C
0 i γ 0 (r ρ )2
γ =1
Fusion
qPof Coil 1 and 2 Sensitivity Map 1 Sensitivity Map 2
C
γ 0 =1 iγ 0 (r ρ )
2 s1 (r ρ ) s2 (r ρ )
The resulting coil sensitivity maps show artifacts in the regions with
little to no signal as compared to the idealized maps. Reconstructions
using these coil sensitivity maps will display noise amplification in air
regions due to these artifacts. More advanced methods exist to deal
with these issues, but they are beyond the scope of this text.
114 6 Magnetic Resonance Imaging
M xy T2 decay
arterial blood
myocardium
venous blood
time
B0 B0
z venous blood z
arterial blood
myocardium
180◦ pulse
90◦ pulse
y y
x x
(a) (b)
B0 B0
z z
180◦ pulse
−90◦ pulse
y y
x x
(c) (d)
Figure 6.14: In the initial state, the spins for myocardial tissue, venous
and arterial blood are all aligned along the main magnetic field. A 90◦ pulse
pushes them into the xy plane, where they precess (a). 180◦ refocusing pulses
are used to counteract T2∗ relaxation, so only T2 relaxation affects the spins
while they are in the xy plane (b), (c). Due to the different T2 constants,
the transversal magnetization decays faster for myocardial and venous blood
spins. After a set amount of time, spins are realigned with the main mag-
netic field by a −90◦ pulse (d). An imaging sequence following this prepara-
tion sequence will now display increased contrast between arterial blood and
myocardial tissue. Figure recreated from [6].
Up to this point, we have quietly assumed that all measured spins remain
stationary for the duration of the imaging. This assumption is, of course,
invalid if we image the human body. For many applications, we have to adapt
the acquisition protocol to minimize artifacts due to motion. For example, we
may ask the person being scanned to hold their breath to minimize respiratory
motion artifacts. Imaging sequences with a short TR can be used to “freeze”
cardiac motion. But in some cases, we can actually use non-stationary spins
to our advantage.
Angiography, the imaging of blood vessels, is commonly performed by
administering a contrast agent which increases the contrast of blood to sur-
116 6 Magnetic Resonance Imaging
saturated saturated
imaging imaging
plane plane
blood blood
vessel vessel
flow flow
direction direction
(a) Directly after saturation (b) Unsaturated spins enter the imag-
ing plane due to blood flow
Figure 6.15: For non-contrast TOF angiography, all spins in the desired
imaging plane are saturated, i. e., put in a state such that they cannot be
excited by an RF pulse, indicated by the gray shading (a). Due to blood
flow, unsaturated spins enter the imaging plane for blood vessels that are
not entirely parallel to the imaging plane (b). Subsequent imaging sequences
will show a high contrast between those vessels and surrounding stationary
tissue.
rounding tissues. In MRI, we can use the fact that blood spins move contin-
uously during the image acquisition to perform a non-contrast angiography.
In the magnetic resonance context, the TOF effect refers to the short amount
of time that flowing blood spins remain within an imaging slice.
For TOF angiography, we first saturate all spins within our imaging plane,
i. e., put them in a state where they cannot be excited by an RF pulse. For the
duration of their T1 relaxation, stationary spins within the imaging plane will
thus show little to no signal if we perform an imaging sequence. But due to
the TOF effect, unsaturated blood spins continuously flow into the imaging
plane and will appear bright in contrast to the surrounding stationary tissue
if imaged.
This imaging technique works best for blood vessels perpendicular to the
imaging plane. For vessels that lie within the imaging plane, the contrast
becomes weaker with increasing distance to the point where unsaturated spins
enter the imaging plane. Fig. 6.15 illustrates the concept of TOF angiography.
6.4 Advanced Topics 117
Further Reading
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 7
X-ray Imaging
7.1 Introduction
In this chapter, the physical principles of X-rays are introduced. We start
with a general definition of X-rays compared to other well known rays, e. g.,
the visible light. In Sec. 7.2, we will learn how X-rays can be generated and
how they can be characterized with respect to their energy. The most relevant
concept to understand how X-ray imaging works is the behavior of X-rays
when they interact with matter. This is outlined in detail in Sec. 7.3. In
Sec. 7.4, conventional X-ray imaging is described with a focus on detector
types and sources of noise. Finally, we finish this chapter with an overview
of well known application areas for X-ray imaging in Sec. 7.5.
X-rays belong to the group of electromagnetic rays, hence, they follow the
rules of electromagnetic radiation. Electromagnetic radiation transports en-
c The Author(s) 2018
A. Maier et al. (Eds.): Medical Imaging Systems, LNCS 11111, pp. 119–145, 2018.
https://doi.org/10.1007/978-3-319-96520-8_7
120 7 X-ray Imaging
Visible Light
Frequency fp [Hz]
100 102 104 106 108 1010 1012 1014 1016 1018 1020 1022 1024
108 106 104 102 100 10−2 10−4 10−6 10−8 10−10 10−12 10−14 10−16
Wavelength λ [m]
ergy, also called radiant energy, through space by waves and photons, just as
radio waves, the visible light or microwaves. It can either be represented by
photons or by a wave model. We will use both representations in the process
of this chapter. Radiation can be classified by its wavelength λp which is the
length of one period of the wave. The wavelength can also be represented by
frequency fp and the waves propagation speed, i. e., the speed of light c0 .
c0
λp = (7.1)
fp
In Eq. (7.2) the energy of photons is given, where h denotes Planck’s constant
(≈6.626 069 × 10−34 J s) and c0 is the speed of light (≈2.997 92 × 105 m s−1 ).
The energy is directly related to the photon’s wavelength λp or its frequency
fp and is given by the unit electron volt [eV]. We can easily obtain that the
photon energy is proportional to its frequency and inverse proportional to its
wavelength, that means the higher its frequency, the higher its energy.
h c0
Ep = = fp h (7.2)
λp
The energy is also used to characterize electromagnetic radiation into dif-
ferent groups, i. e., radio waves, microwaves, infrared (IR), visible light, ul-
traviolet (UV) light, X-rays and γ-rays. Fig. 7.1 shows these groups with
respect to their characteristic ranges of frequency and wavelength. Note that
the wavelength of most X-rays lies in the range of 0.01 nm up to 10 nm. This
corresponds to an energy range of 100 keV down to 100 eV.
As visible light, X-rays loose a certain amount of energy when they pass
through different materials. The energy loss depends on the absorption be-
havior of the material. For example if X-rays pass through 10cm of water,
they loose less energy than if they would pass trough 10cm of bone. The
reduction of energy is caused by absorption which is the main principle of
traditional X-ray imaging. Generally speaking, X-ray radiography measures
the amount of energy loss. Because this energy loss differs for the different
materials, we can see a certain contrast in the image. For example an X-ray
7.1 Introduction 121
image shows high intensities for soft tissue and lower intensities where the X-
rays passed through bones. Note that the absorbed energy is directly related
to the dose that is delivered to the patient during an acquisition.
Discovery of X-rays
X-rays Today
Today, X-rays are routinely used in diagnostic but also in interventional med-
ical imaging around the globe. Also in industry, X-rays are often the method
of choice, for example to test for very small cracks in metal parts in the field
of non-destructive testing. In medical imaging, a variety of applications have
been developed that go far beyond simple radiographic imaging. For example
122 7 X-ray Imaging
Röntgen donated his discovery to humanity and never made any com-
mercial profit. He also never filed a patent for his invention. The actual
commercial roll-out of the technology was performed by several small
companies:
• In 1877, Erwin Moritz Reiniger, a former employee of Friedrich-
Alexander-University Erlangen-Nuremberg, Germany, founded a
small workshop right next to the university with the aim of produc-
ing batteries and physical measurement devices. In 1886, Max Geb-
bert and Karl Friedrich Schall joined Reiniger’s workshop found-
ing the Vereinigte Physikalisch-Mechanische Werkstätten Reiniger,
Gebbert & Schall — Erlangen, New York, Stuttgart. In 1896, they
switched the focus of production to X-ray tubes. Over the years this
small company grew and is today known under the name Siemens
Healthineers AG.
• Victor Electric Company was founded in 1893 by C. F. Samms and
J. B. Wantz in a basement in Chicago, United States of America,
with the aim of producing physical measurement gear. In 1896,
they also began with the production of X-ray tubes. Also this small
company turned out to be very commercially successful and is today
known as General Electric Healhcare.
• In 1896, C. H. F. Müller developed the first commercial X-ray tube
in Hamburg, Germany, in cooperation with the University Clinic
Hamburg-Eppendorf. In 1927, the company was bought and is today
an integral part of Philips Medizin Systeme GmbH.
fluroscopy allows for real time X-ray sequences which are often inevitable
in minimally invasive interventions. Further, digital subtraction angiography
(DSA) provides an effective tool to visualize even small vessel structures. In
the 1970s, the step to CT was done which now allows to visualize the com-
plete human body in three dimensions. Another point of rapid development is
the awareness that X-rays can also be harmful. High energies emitted to the
body during an X-ray acquisition can lead to ionization. That means the ra-
diation changes the atomic structure of the tissue which can potentially lead
to an increased risk for the development of cancer. Here, deoxyribonucleic
acid (DNA) becomes damaged by the radiation. In most cases, the DNA will
be repaired by the cell itself. Yet, the repair process sometimes fails which
in some cases leads to an unregulated division of cells that might result in
cancer. X-ray-based foot scanners where still in use to measure foot sizes in
shoe stores until the 1970s. Nowadays, the majority of people is aware of the
risk posed by X-rays and the transmitted patient dose during X-ray scans
7.2 X-ray Generation 123
Figure 7.2: Wilhelm Figure 7.3: One of the Figure 7.4: Image
Conrad Röntgen first X-rays, taken from taken in 1896 showing
Wilhelm Röntgen of his Albert von Kölliker’s
wifes hand. hand.
has been significantly reduced in the past decades1 . In Fig. 7.5, we can see
an X-ray taken in mammography which typically uses low-energy X-rays to
increase soft-tissue contrast. Another example is shown in Fig. 7.6, where we
can see an X-ray image taken from the thorax, i. e., the chest, of a patient.
X-rays
Cathode
Anode
Cooling
System
Acceleration Voltage
Figure 7.7: Vacuum X-ray tube: The image on the left shows a schematic
how electrons are accelerated from the cathode to the anode to genereate
X-ray photons. The image on the right shows a historic vacuum X-ray tube.
Image provided by Science Museum, London, Wellcome Images under Cre-
ative Commons by-nc-nd 2.0 UK.
that an electron loses all its velocity and thus its energy in one step. In
this case, only one photon containing the complete energy of the electron is
created.
The production of X-rays is caused by two different processes as shown
in Fig. 7.8. First, if the electron interacts with an inner-shell electron of the
target, characteristic X-radiation can be produced. This kind of X-rays results
from a sufficiently strong interaction that ionizes the target atom by a total
removal of the inner-shell electron. The resulting “hole” in the inner-shell is
7.3 X-ray Matter Interaction 125
7.3.1 Absorption
I = I0 · e−µx . (7.3)
When the ray passes through inhomogeneous objects, factor µ(x) is the linear
attenuation coefficient at each point on the ray path
Z
I
− ln = µ(x)dx . (7.4)
I0
However, in practical setups, the emitted X-ray photons have various energies,
resulting in polyenergetic energy spectra as shown in Fig. 7.8. The measured
intensity of a polychromatic beam I on the detector can be written as the
sum of monochromatic contributions for each energy E in the X-ray spectrum
(E ∈ [0, Emax ]). The attenuation coefficient µ is also energy dependent. When
polychromatic X-rays are taken into account, we get
Z Emax Z
I= I0 (E) exp − µ(x, E)dx dE (7.5)
0
The photoelectric effect was originally described by Einstein under the es-
tablishment of the quantized nature of light. It describes a situation in which
the incident X-ray photon energy is larger than the binding energy of an
electron in the target material atom. The incident X-ray photon gives up its
entire energy to liberate an electron from an inner-shell. The ejected electron
is called photoelectron. The incident photon then ceases to exist. The photo-
electric effect often leaves a vacancy in the inner shell of the atom, where the
ejected electron was previously located. As a result, the “hole” created in the
inner-shell is filled by an outer shell electron. Since the outer shell electron
is at a higher energy state, a characteristic radiation occurs. Therefore, the
photoelectric effect produces a positive ion, a photoelectron, and a photon of
characteristic radiation. For tissue-like materials, the binding energy of the
K-shell electrons is very small. Thus, the photoelectron acquires essentially
the entire energy of the X-ray photon.
incident photon’s wavelength. The excess energy from the vibrating electron
transfers to an electromagnetic photon which has the same wavelength but
possibly a different direction than the incident photon. In this interaction,
electrons from the material’s atom are not ejected and no energy is converted
into kinetic energy, thus, no ionization occurs. Rayleigh scattered photons are
mostly emitted in a forward direction with respect to the incident photon’s
direction. For X-rays used for imaging, the contribution of Rayleigh scattering
to total attenuation is usually small compared to other contributions.
In the previous sections, the concepts of X-ray generation and also their
interaction behavior with matter has been outlined. In this section, we will
now focus on different detection methods used to convert the X-rays that have
passed the patient to an actual image. Unlike the old X-ray films, which use
X-rays directly to change the chemical properties of the X-ray film material,
the modern detection systems first convert the X-rays to light and eventually
to electrons.
X-ray image intensifiers are vacuum tubes that are used to convert X-rays into
visible light, i. e., an image. The schematic principle of this process is shown
in Fig. 7.10. First, the incoming X-ray photons are converted to light photons
using a phosphorus material called the input phosphor. The produced light is
further converted to electrons by exploiting the photoelectric effect inside a
photocathode. These electrons are then accelerated and focused towards the
output phosphor using an electron optic system. In the output phosphor, the
electrons are converted back to visible light which can then be captured by
film material or television camera tubes.
7.4 X-ray Imaging 131
Input Photo-
Evacuated
Phospor cathode
Tube
Accelerating
Anode
X-rays
Light
Electron
Optics
Output
Phosphor
Input
Window
7.4.1.1 Function
After passing through the input window, the X-rays hit the input phos-
phor used to convert X-ray photons to light photons. The generated light
photons trigger a photoelectric effect in the photocathode which then emits
(photo-)electrons. The input phosphor and the photocathode are typically
layered to one piece. Starting with the input phosphor that consists of another
aluminum plate coated with the phosphor layer, followed by an intermediate
layer and the photocathode layer.
Let us focus on the input phosphor layer in more detail. One important
property that influences the efficiency of the input phosphor layer is its thick-
ness. The thicker the phosphor layer, the higher is its absorption, thus, more
X-ray photons are absorbed and converted to light. Hence, less X-ray pho-
tons are required which reduces radiation exposure to the patient. However,
with increasing thickness also more light photons become scattered within
the phosphor layer which effectively reduces the spatial resolution.
Another property that is used to increase conversion factors is the chem-
ical composition of the input phosphor material and its resulting mass at-
tenuation coefficient. Ideally, the input phosphor’s attenuation coefficient is
adjusted to the residual incoming X-ray spectrum. Initially, zinc-cadmium
sulfide (ZnCdS) has been used as phosphoric material, which has been re-
placed by cesium iodide (CsI) in modern detector systems. The advantages
of CsI over ZnCdS are twofold. In Fig. 7.12 we illustrate the mass attenuation
coefficient of CsI (dashed, dark blue line) and ZnCdS (dotted, light blue line)
w. r. t. the photon energy. Additionally, the estimated spectral distribution of
a typical X-ray spectrum after transmission through the patient is depicted as
solid, orange line. The higher the overlapping area between attenuation char-
acteristics and residual X-ray spectrum, the better its conversion efficiency.
We can clearly see that the mass attenuation coefficient of CsI matches better
to the expected residual X-ray spectrum and is thus favorable.
Additionally, the manufacturing process of CsI allows to build the phos-
phor layer as a collection of small and local cylindrical structures as indicated
in Fig. 7.14. The cylindrical wires act as optical fibers which can steer the
emitted light to the photocathode with a high spatial accuracy. Thus, scat-
tering of the light photons within the phosphor material can be drastically
reduced. In modern detectors, the input phosphor is about 300 µm to 500 µm
thick and can absorb up to 70 % of the incoming X-ray photons. A single
60 keV X-ray photon can create up to 2600 light photons, where approxi-
mately 62 % reach the photocathode.
The photocathode layer typically consists of antimony-cesium (SbCs3 ).
Similar to the input photon layer, the incoming light should fit to the sen-
sitivity spectrum given by the photocathode. Fig. 7.13 shows the sensitivity
spectrum of an SbCs3 photocathode, together with the characteristic light
spectra emitted from a CsI as well as a ZnCdS phospor layer. We can see
that also here CsI seems to produce a light spectrum that matches better to
the photocathode, hence, leading to a higher conversion efficiency from light
photons to electrons.
7.4 X-ray Imaging 133
25 Absorption CsI
Absorption ZnCdS 0.8 ← ZnCdS
Relative units
20
0.6
15
0.4
10 SbCs3 →
0.2 ← CsI
5
X-rays
Input Phospor
Cesium Iodine
Photo-
cathode
Electrons
Figure 7.14: Cesium-iodine layer has cylindrical structure and acts as opti-
cal fibers. Thus, the scattering of the light photons is reduced significantly.
After the electrons leave the photocathode, they are accellerated by the
anode as shown in Fig. 7.11. Moreover, the accelerated electrons are focused
onto the output phosphor using electrostatic fields produced by the electron
optic. No additional electrons are induced into the system by this process,
the existing electrons are merely accelerated and deflected. The increase of
kinetic energy that originates of the acceleration process results in a higher
number of light photons that are emitted when the electrons hit the output
phosphor. Hence, the intensity or brightness of the output phosphor can
be altered by a regulation of the acceleration voltage. The output phosphor
consists typically of silver-activated zinc-cadmium sulfide (ZnCdS:Ag) and is
very thin (4 µm to 8 µm). About 2000 light photons are generated for a single
25 keV electron. Due to the fact that one electron is emitted by one light
photon in the photocathode, this also represents an increased brightness by
a factor of 2000.
134 7 X-ray Imaging
1
Relative Luminance
0.8
0.6
0.4
−20 −15 −10 −5 0 5 10 15 20
Distance from image center [cm]
Besides common limitations that all imaging systems share, e. g., spatial res-
olution and contrast ratio, image intensifier systems are most known for vi-
gnetting and distortion artifacts. Vignetting, as described in Fig. 7.15, de-
scribes a drop in brightness that occurs at the outer parts of the screen. It is
caused by light scattering that deflects light photons in the output phosphor
from the outer part of the phosphor to the inside. However, no scattering
occurs from completely outside the material to the outer regions of the phos-
phor, yielding an increased brightness at the central regions. Another common
artifact is image distortion as indicated in Fig. 7.16. It is known that the elec-
tron optics of image intensifiers is susceptible to external magnetic or electric
fields. Even the earth’s magnetic field causes considerable distortions in the
output image. To correct for distortion artifacts, regular calibration is needed
where the distortion field is estimated by measuring predefined calibration
objects. The distortion can be corrected by either adjusting the electron op-
tics accordingly or by subsequent image processing in case the images have
been digitized.
In the recent years, flat panel detector (FPD) became the state-of-the-art
in X-ray detector technology for radiography, angiography, and C-arm CT
applications. They were first introduced in the mid 1990s and their main
advantages are a direct digital readout of the X-ray image and an increased
spatial resolution. Flat panel detectors can be categorized into direct and
indirectly conversion FPDs.
7.4 X-ray Imaging 135
Similar to the image intensifier system discussed in the previous section, the
FPD still converts X-rays to light photons by using a layer of cesium iodide
(CsI). Also the tubular structure of the CsI is identical to the input layer
of an image intensifier system as shown in Fig. 7.14. The major difference
are the subsequent detection steps. Image intensifiers make use of a further
conversion of light photons to electrons which are then accelerated to increase
and control illumination. This additional conversion step is not necessary for
flat panel detectors. Instead a matrix of photodiodes is directly attached to
the CsI layer and converts the emitted light photons to an electric charge
which is then stored in capacitors for each pixel. Each pixel also contains a
thin-film transistor (TFT) which acts as small “switch” used for the readout
of the stored charges.
For both the indirect but also the direct conversion FPDs, the readout of the
pixels is done row-wise using a certain readout frequency. A row is selected
by “switching on” the TFTs of this row’s pixels, i. e., by applying a voltage to
the gate of the TFTs. The stored charges of each pixel are directed to a charge
integrating amplifier and subsequently converted to a digital representation.
These digital pixel values are serialized and transferred over a bus system to
the imaging computer. Common FPDs for medical imaging can have a side
length of up to 40 cm and a pixel size of about 100 µm to 150 µm. They
are available in quadratic but also in wide formats. The analog to digital
conversion uses a quantization of 12 to 16 Bit. To increase the signal to noise
ratio multiple pixels are often combined to a bigger pixel during the readout
process, which is also known as binning. Typical binning modes are 2x2 or 4x4
binning, reducing the image size by a factor of 2 or 4, respectively. Because
binning does not require any additional time, the frame rate increases by
the binning factor. Frame rates typically vary between 7.5 and 30 frames
136 7 X-ray Imaging
There are two types of undesirable effects in medical imaging systems: prob-
abilistic noise and artifacts. Similar to noise, artifacts are image degradations
that also find their source in physical effects during the scan. However, the
difference to noise is that when a scan is repeated using the exact same object
and scan parameters, artifacts are reproduced exactly whereas noise effects
will change based on a probabilistic scheme. Some artifacts, for example,
distortion and vignetting, have already been shown in the section Sec. 7.4.1
on image intensifier detectors. In the following, we focus on the sources and
propagation of noise in X-ray imaging.
As illustrated in Fig. 7.17, there are different states of an X-ray pho-
ton. Each step in this chain follows either a Poisson distribution (cf. Geek
Box 7.3) or a binomial distribution (cf. Geek Box 7.4). In Fig. 7.18, we show
both distributions in comparison. The X-ray photon generation process (cf.
Geek Box 7.5) follows a Poission distribution. The matter interaction and
the detection step (cf. Geek Box 7.6) follow a binomial distribution. Both
processes interact along the path of the X-ray (cf. Geek Box 7.7) resulting
in yet another Poisson distribution. As such Lambert-Beer’s law also has a
probabilistic interpretation (cf. Geek Box 7.8) and every observation on the
detector is Poisson distributed in the monochromatic case.
A common quality measure for imaging is the signal-to-noise ratio (SNR).
It is not uniquely defined over different fields of applications. In X-ray imaging
it makes sense to use the definition based on statistics, i. e.,
n̄ E(N )
SN R(N ) = =p . (7.6)
σ E((N − n̄)2 )
0.2
0.1
0.1
0
0 5 10 15 20 0
0 5 10 15 20
(a) Mass distribution functions of Pois-
son distributions with varying expecta- (b) Mass distribution functions of bino-
tion value N0 , plus their Gaussian coun- mial distributions with varying number of
terparts. trials N and fixed probability p = 0.5.
(N0 )n −N
Poisson(N0 ) = p(N = n) = e 0 , (7.7)
n!
where N0 is the expectation value of the observed event E(N ). We
now show a simple example for the usage of the Poisson distribu-
tion. Assume a local shop records its daily number of customers for
a year which results in an average of N0 = 15 customers per day.
The Poisson distribution can now be used to calculate the probability
that on a new day there will be n = 20 customers in the shop, i. e.,
(N0 )n −N 20
p(N = 20) = n! e 0 = 15 20! e
−15
≈ 0.0418. In Fig. 7.18(a), the
mass distribution function as defined in Eq. (7.7) is shown for three
different expectation values N0 . If the number of N0 becomes high,
the Poisson distribution approaches a normal distribution with mean
n̄ = N0 and standard deviation σ = N0 . This is based on the so
p
called “central limit theorem”. In Fig. 7.18(a), we have also added the
corresponding mass distribution functions for each Poisson distribu-
tion. You can clearly see that the higher N0 , the closer the discrete
Poisson distribution gets to a normal distribution.
7.5.1 Radiography
N!
N
p(N = n) = pn (1−p)N −n = pn (1−p)N −n . (7.8)
n n! (N − n)!
N0 = Ne pex
(Ne pex )n −Ne pex
P (N = n) = e , (7.9)
n!
where N0 denotes the expected value for the number of electrons that
trigger an X-ray photon, which is also known as a measure for the
radiation intensity. We use P (Nx ) as the probability that an X-ray
source produces exactly Nx X-ray photons. P (Nx ) is then given by
above equation, where n has been replaced by Nx .
140 7 X-ray Imaging
The generated X-ray photons are now traveling through space towards
the detector and interact with the matter they pass through according
to Beer’s law as introduced in Eq. (7.3). Whether the photons interact
with the matter or pass through it unaffected can be interpreted as
Bernoulli trial, i. e., an experiment with random binary outcome. If
a single photon encounters an interaction depends on the material
properties along the photons path, i. e., the attenuation µ(x). The
probability pa for the photon passing unaffected is again given by
Beer’s law: R
pa = e− µ(x)dx . (7.10)
As the individual X-ray photons are independent from each other in
terms of interacting or not, the process can be described as a binomial
distribution. Furthermore, it can be shown that when we have a Pois-
son distributed variable (Nx ) that represents the number of samples
in a binomial distribution, the outcome is again Poisson distributed.
We refer to Geek Box 7.7 for further information.
(N0 pa )ns −N pa
P (N = ns ) = e 0 , (7.11)
ns !
where ns is the number of photons that pass through the matter un-
affected. We can also determine the expectation value of this Poisson
distribution, i. e.,
R
E[ns ] = N0 pa = N0 e− µ(x)dx . (7.12)
7.5.2 Fluoroscopy
Let us focus on the distribution for the number of photons that ar-
rive at the detector, i. e., ns . The distribution is given by Eq. (7.11).
From Poisson statistics we know that the expectation value is the only
parameter of the mass distribution function, which was already com-
puted in Eq. (7.12), hence, n̄ = N0 pa . Further, it can be shown that
the variance (σ 2 ) of a Poisson distribution is equal to the expectation
value, i. e., σ 2 = N0 pa . The SNR after photon interaction with the
object can then be computed by
N pa
SN R(N ) = p 0
N 0 pa
= N 0 pa (7.13)
p
1
R
= N0 e − 2 µ(x)dx
p
.
(a) Pacemaker Fluroscopy. Image (b) Clinical Setup. Image courtesy of Siemens
taken from [5]. Healthineers AG.
Figure 7.20: Left: Image from a fluoroscopy sequence showing the placement
of two electrodes of a pacemaker. Right: Typical clinical setup of a minimally
invasive surgery. The fluoroscopy is acquired using a freely positionable C-
arm device. The image shows the X-ray source at the bottom and the FPD
is right above the patient.
Figure 7.21: Process of creating a DSA. In (a) the hand was imaged and
no contrast agent has been injected (mask image). In (b) the same hand has
been imaged but including injected contrast agent. The difference of (b) and
(a) represents the angiogram as shown in Figure (c). Images provided by
Adam Galant, Siemens Healthineers AG.
are iodine and barium, where the first is used for intravascular and the latter
for gastrointestinal examinations. Thus, iodine is injected into the blood cir-
culation whereas barium can be swallowed to investigate, e. g., the stomach
or colon.
In DSA, a fluroscopic sequence of a fixed anatomy is acquired. At the
same time contrast agent is injected in regular intervals into the vessel sys-
tem. X-ray images that have acquired the scene without contrast agent are
assumed to show the background tissue that is typically not of interest. If we
now subtract the initially acquired background image from an X-ray image
with contrast and assume that no patient motion has taken place, we can
measure the attenuation caused only by the injected contrast agent. As the
contrast agent is limited to the vessel system, it has been injected to, the
outcome of such a subtraction will be a visualization of the vessels only. In
Fig. 7.21, an example for a DSA acquisition is presented. First the contrast
agent free image, i. e., the mask image, is acquired as shown in Fig. 7.21(a).
Then contrast is injected into the vascular system and after some waiting a
further image, i. e., the fill image, is acquired. The difference of both images
is then the so called angiogram that shows only the contributions given by
the contrast agent and thus the vessels are visualized (cf. Fig. 7.21(c)).
Further Readings
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 8
Computed Tomography
8.1 Introduction
CT is doubtlessly one of the most important technologies in medical imaging
and offers us views inside the human body that are as valuable to physicians
as they are fascinating (cf. Fig. 8.1).
8.1.1 Motivation
In the previous chapter, we have seen how X-rays can be used to acquire
2-D projection images. However, a single projection image does not retain all
spatial information, as it merely shows something akin to “shadows” of the
imaged objects. An example is given in Fig. 8.3(a), which shows an X-ray
projection image of a luggage bag. Two arrows indicate objects that cannot
easily be identified. Using multiple projection images from different angles, we
c The Author(s) 2018
A. Maier et al. (Eds.): Medical Imaging Systems, LNCS 11111, pp. 147–189, 2018.
https://doi.org/10.1007/978-3-319-96520-8_8
148 8 Computed Tomography
Figure 8.3: 2-D X-ray projection image of a luggage bag (a) and a corre-
sponding 3-D reconstruction, visualized with a volume rendering technique
and (b, c) and as orthogonal cross-sectional slices (c). 1) Indicates a hidden
text revealed in (b) and 2) an apple that is virtually sliced in (c). Images
courtesy of Chris Schwemmer.
practical application for the following 50 years. The main concepts introduced
in his article will be outlined in Sec. 8.2 and used in Sec. 8.3 to explain how
CT image reconstruction works.
Only in 1971, the first CT system was built by Sir Godfrey Newbold
Hounsfield and Allan McLeod Cormack1 . Fig. 8.2 shows the result of the
first clinical scan of a patient’s head performed in the same year. For their
seminal invention, they received the Nobel prize in medicine in 1979.
A major advance in the field was the introduction of spiral (or, more
accurately, helical) CT by Willi Kalender et al. in 1990. Its name is derived
from the novel aspect of its acquisition trajectory describing a helix. Amongst
others, this geometry and its advantages will be described in Sec. 8.3.3.
In the early days of CT imaging, data acquisition was fairly slow, taking
approximately 4 minutes per rotation. The reconstruction of a single 2-D
slice with a low spatial resolution of 80 × 80 pixels and 3 bit quantization
took several hours. By 2002, rotation speed had improved drastically with
one rotation performed in only 0.4 seconds. Up to 16 slices in parallel could
be reconstructed on-the-fly, at a higher resolution of 512 × 512 pixels and a
quantization depth of 16 bit.
In recent years, this trend continued with temporal and spatial resolutions
constantly improving. In this context, the development of dual source CT
in 2005 was another significant milestone, featuring two X-ray sources and
detectors in a single scanner. In addition to offering additional information
when both employed X-ray tubes are operated at different voltages (dual
energy scan), it can also be used to speed up the acquisition significantly.
The amount of slices acquired in parallel had also increased further, covering
a field of view measuring up to 16 cm in axial direction at voxel sizes below
one millimeter. This allows imaging of complete organs such as the heart in
a single rotation, thus reducing motion artifacts.
In 2014, a modern CT system (cf. Fig. 8.4) was able to acquire up to 128
slices in parallel at a temporal resolution as low as 195 ms with a single X-ray
source.
In this section, we will first introduce the Radon transform as the underly-
ing mathematical principle of the image formation process in CT imaging.
Inverting this transform is the fundamental problem solved by image recon-
struction methods. Subsequently, we will detail the Fourier slice theorem, a
1
Both researchers were working for EMI at the time, a British music recording and
publishing company well known for housing the Beatles’ label. This has spawned a
widespread belief that the Beatles’ success contributed to financing the initial develop-
ment of CT.
150 8 Computed Tomography
property related to the Radon transform that constitutes the core idea of an
important class of reconstruction algorithms.
Radon’s key insight was that any integrable function f (x, y) can be uniquely
represented by – and therefore recovered from – all straight line integrals over
its domain,
Z +∞
p(`) = f (x(l), y(l)) dl, ∀l : (x(l), y(l))> ∈ line ` (8.1)
−∞
with θ the angle between the line’s normal vector and the x-axis and s the
orthogonal distance between line and origin (cf. Fig. 8.5). Implicitly, this line
is described by the equation x cos θ + y sin θ = s. Only those points that
satisfy it, i. e., those that fall on the line, are selected by the Dirac function δ
in Eq. (8.2), as it vanishes everywhere else (cf. Eq. (2.10)). In that way, the
integration of f (x, y) is only performed along the respective line.
The complete set of line integrals p(θ, s) can now be obtained by going
through the angles θ ∈ [0◦ , 180◦ ] and distances s ∈ [−∞, +∞]. Apart from
orientation which has no influence on the integration, any other line would be
8.2 Mathematical Principles 151
Figure 8.5: The blue line is uniquely described by its distance s to the
origin and the angle θ which defines its normal vector (cos θ, sin θ)> . This
representation immediately gives rise to the implicit line equation.
Figure 8.6: f (x, y) has a constant non-zero value inside the blue circle and
vanishes everywhere else. We see a single projection on the left and where it
fits into the whole sinogram on the right.
equivalent to one of these. For a fixed angle θ, the 1-D function pθ (s) = p(θ, s)
is called a projection. It contains all line integrals over f (x, y) with a constant
angle θ and variable distance s to the origin. Arranging all projections side-
by-side as a 2-D image yields the sinogram. It owes its name to the sinusoidal
curves emerging from the underlying geometry. We can see that every point
in 2-D except for the origin is found at different distances along the s-axis
depending on the angle θ. An example of a sinogram is given in Fig. 8.6.
Turning the function values f (x, y) into line integral values p(θ, s) is known
as the Radon transform in 2-D. The aim of CT reconstruction is the com-
putation of the original function values from measured line integral values,
i. e., the inverse Radon transform.
152 8 Computed Tomography
using the Fourier transform F (cf. Sec. 2.3 (p. 22)). As illustrated in
Fig. 8.7, P (ξ, θ) is equivalent to the part of F (u, v) that falls on a
radial line with angle θ. To see why this is the case, we start with the
1-D Fourier transform P (ξ, θ) of pθ (s),
Z +∞
P (ξ, θ) = pθ (s)e−2πiξs ds. (8.5)
−∞
Using the definition of the projection pθ (s) from the previous section,
we obtain
Z +∞ Z Z +∞
P (ξ, θ) = f (x, y)δ(x cos θ + y sin θ − s) dxdy e−2πiξs ds.
−∞ −∞
(8.6)
Rearranging the order of the integrals yields
Z Z +∞ Z +∞
P (ξ, θ) = f (x, y) δ(x cos θ + y sin θ − s)e−2πiξs dsdxdy.
−∞ −∞
(8.7)
Eliminating the delta function reads as
Z Z +∞
P (ξ, θ) = f (x, y)e−2πi(x cos θ+y sin θ)ξ dxdy. (8.8)
−∞
In effect, we can get the complete Fourier transform Fpolar (ξ, θ) of the
unknown function f (x, y) in polar coordinates (ξ, θ) by varying θ.
154 8 Computed Tomography
Figure 8.8: Graphical visualization of the Fourier slice theorem. In fact, com-
putation of the projection and correlation with a sinusoidal function (dashed
lines) is equivalent to a 2-D correlation with the respective Fourier base func-
tion (dotted lines, cf. Fig. 6.9).
Figure 8.9: A close look at the Fourier base functions reveals that they are
actually computing an integration along the wave front. As such projection
and convolution operate in orthogonal domains and can therefore be sepa-
rated into a projection and a 1-D correlation, i. e., a 1-D Fourier transform.
8.3 Image Reconstruction 155
Material / Tissue HU
Air −1000
Lung −600 to −400
Fat −100 to −60
Water 0
Muscle 10 to 40
Blood 30 to 45
Soft tissue 40 to 80
Bone 400 to 3000
Table 8.1: HUs observed in several materials and tissue classes found in the
human body. In general, denser structures exhibit larger HUs.
As described in Sec. 7.3 (p. 125), X-ray projections can be converted to line
integrals using Beer’s law, which enables us to apply Radon’s ideas to CT
reconstruction. A single slice of our imaged object corresponds to the bivariate
function f (x, y). More precisely, the function values reconstructed in CT are
the linear attenuation coefficients of the imaged material. Typically, they are
linearly transformed to the Hounsfield scale, which is normalized such that
the absorption of water equals 0 HU,
µ
µ =
∗
− 1 · 1000, (8.11)
µWater
where µ and µ∗ denote the coefficients before and after Hounsfield scaling,
respectively. Tab. 8.1 lists the HU ranges of several tissue classes found in
the human body.
Below, we will discuss the two main methods for 2-D image reconstruc-
tion from parallel-beam projections as they have been introduced above. In
conventional CT imaging, a 3-D image volume is then obtained simply by
acquiring and reconstructing multiple axial slices at slightly offset locations
such that they can be stacked on top of each other (Fig. 8.10).
where h(s) corresponds to the inverse Fourier transform of |ξ|. This amounts
to the back-projection of pθ (s) convolved with h(s). As a consequence, this
method is called filtered back-projection.
Unfiltered back-projection, i. e., just “smearing” line integrals in a projec-
tion pθ (s) back along their corresponding lines without filtering (cf. Fig. 8.12),
is equivalent to adding P (ξ, θ) to F (u, v), as suggested by the Fourier slice
theorem, without considering the factor |ξ|. In fact, this is not the inverse,
but the dual or adjoint of the Radon transform.
8.3.1.2 Filters
Due to the shape of |ξ|, the filter h(s) is typically called ramp filter. Sam-
pling in polar coordinates leads to an oversampling in the center of the Fourier
space (cf. Fig. 8.11). Using a ramp filter, this oversampling is corrected by en-
hancing the high frequency components while dampening the low frequencies
in the center of the Fourier space.
According to the sampling theorem, as described in Sec. 2.4.2 (p. 32), with
a detector spacing of ∆s, the largest frequency that can be detected in pθ (s) is
ξmax = 2∆s1
. Additionally, noise in the projections is amplified when using an
unlimited ramp filter |ξ|. Therefore, high frequencies should be limited in the
8.3 Image Reconstruction 157
From the Fourier slice theorem, we know F (ξ, θ) = P (ξ, θ), thus
Z π Z +∞
f (x, y) = P (ξ, θ)|ξ|e2πiξ(x cos θ+y sin θ) dξdθ.
0 −∞
filtered projections p̃θ (s). For this purpose, we can generalize the following
equation Z +∞
p̃θ (s) = P (ξ, θ)|ξ|e2πiξs dξ, (8.14)
−∞
In practice, various ramp-like filters are used depending on the desired im-
age characteristics, typically involving a trade-off between a smoother image
appearance and a higher spatial resolution.
One of the most widely known filters was described by Ramachandran and
Lakshminarayanan, in short known as the “Ram-Lak” filter. It corresponds
to |ξ| cut off at ξmax on both sides. In the spatial domain, this results in a
filter kernel that reads
sinc ∆s s
sinc2 2∆s
s
h(s) = − , (8.16)
2(∆s)2 4(∆s)2
In order to derive the Ram-Lak filter, we need to start with the inverse
Fourier transform of |ξ|
Z +∞
h(s) = |ξ|e2πiξs dξ.
−∞
Now the band-limited inverse Fourier transform of |ξ| takes the fol-
lowing form:
h i
h(s) =F −1 B − rect Bξ ∗ rect Bξ rect 2B
ξ
h i
=F −1 B rect 2Bξ
− F −1
rect ξ
B ∗ rect ξ
B rect ξ
2B
| {z } | {z }
support on [−B,B] =1 on [−B,B]
h i h i h i
=F −1 B rect ξ
2B − F −1 rect ξ
B · F −1 rect ξ
B
2∆s2 s = 0
1
hs = 0 s even
π 2 (s∆s)2 s odd
1
160 8 Computed Tomography
Figure 8.13: The responses in frequency domain (top row) as well as the
discretized spatial domain kernels (bottom row) of the Ram-Lak and Shepp-
Logan filters.
8.3.1.3 Discretization
becomes X
p̃θ,s = pθ,s · hs−s0 ∆s. (8.19)
s0
8.3 Image Reconstruction 161
Figure 8.14: Example for an image grid and some projection rays.
where θi denotes the ith angle. Note that the value p̃θi (s) will generally have
to be interpolated from the p̃θi ,s since s is not necessarily an integer number.
For each position (x, y), we can find f (x, y) by summing over corresponding
(interpolated) values of each filtered projection p̃θi . As a rule of thumb, it is
recommended to avoid interpolation in the output space, i. e., in our case, we
should sample f directly. This comes naturally with the formulation given in
Eq. (8.20). In contrast, back-projecting one p̃θi at a time to the whole volume
would require interpolation on grid points in the domain of f in each step.
Ax = p, (8.21)
162 8 Computed Tomography
where
are the sequentially numbered unknown pixels and measured line integrals,
respectively. Each element aij of the system matrix A describes the contri-
bution of a particular pixel to a particular ray. There are many possibilities
to model A. In the simplest case, aij is a binary value that is 1 when the ray
passed through the pixel and 0 otherwise. The length of intersection can also
be used, or even the area of intersection in case we assume the rays to have
a non-zero thickness.
Solving this system of linear equations for the solution x directly using
matrix inversion (Gaussian elimination, singular value decomposition, etc.)
is not feasible in practice as the problems are typically large, ill-conditioned
and over-determined. Instead, an iterative solution to this system of linear
equations is sought.2
The algebraic reconstruction technique (ART) aims to find such an itera-
tive solution using the Kaczmarz method. The basic idea behind this method
is that each linear equation defines a line (2-D) or, generally speaking, a hy-
per plane (higher dimensions) in the solution space, the dimensionality of
which equals the number of unknowns. All points on a hyper plane fulfill its
corresponding equation. Consequently, the point of intersection of all hyper
planes forms the correct solution to the problem. Thus, by repeatedly pro-
jecting the current estimate orthogonally onto a different equation’s plane, we
iteratively improve the solution (Fig. 8.15). A simple mathematical intuition
for the ART algorithm is given in Geek Box 8.4.
Using Kaczmarz’ method, we can now find an iterative solution for
Eq. (8.21). For each line integral measurement pi and each row a i of the
system matrix A we perform the following update step,
pi − a i x k >
x k+1 = x k + ai , (8.29)
aia>
i
In the 2-D case, consider a point x and a line n > c = d, where n is the
normal vector, c a point on the line and d the distance to the origin.
Note that n > c describes the scalar vector product. The orthogonal
projection x 0 of x on this line must be in the direction of the normal
vector n:
x 0 = x + λn (8.24)
The projected point x 0 is part of the line and therefore fulfills
n > x 0 = d. (8.25)
d − n>x
x0 = x + n. (8.28)
n>n
with X
uk,i = 1, (8.31)
i
(a) For a parallel beam geometry as introduced in the previous sections (shown
on the left), the X-ray source needs to be shifted perpendicularly (dotted line) to
the direction of projection, casting pencil beams through the object. If all beams
instead emanate from a single position for each angle, we obtain a fan of no
longer parallel rays (fan beam geometry; on the right), increasing acquisition
efficiency at the cost of a slightly more complicated reconstruction problem.
Apart from the flat shape shown here, there also exist curved detectors with an
equiangular spacing.
(b) Multiple detector arrays allow for simultaneous acquisition of multiple im-
age slices from one X-ray source position (multi-slice CT; shown on the left).
However, in this setup, the beams no longer all lie within the rotation plane.
This issue becomes much more important in the case of cone beam CT (shown
on the right): Here, the small stack of detector rows gives way to a larger de-
tector matrix, with the beams now forming a cone in 3-D.
However, there are certain geometries for which (AA> )−1 can be de-
termined analytically. For the case of parallel-beam geometries, we
know that (AA> )−1 takes the form of a convolution with the ramp
filter. A> is the adjoint operation of the projection operator. In con-
tinuous form, we introduced this operation already as back-projection
(cf. Geek Box 8.2). Thus, filtered back-projection is a direct solution
for the above system of linear equations.
Figure 8.17: A C-arm system for interventional X-ray imaging. Image cour-
tesy of Siemens Healthineers AG.
8.4 Practical Considerations 167
Figure 8.18: In spiral CT, although the X-ray source still conveniently ro-
tates in the x-y plane, the trajectory it describes in relation to the imaged
object is a helix due to the patient table being slowly moved through the
gantry. This enables the acquisition of a large object region while rotating
continuously. Projections for an ideal circular trajectory can be interpolated
along z from neighboring helical segments.
Figure 8.19: The effective sizes bF of the focus and bD of the detector in the
isocenter can be calculated from the distances to the isocenter rF and rD .
Figure 8.20: A bar phantom can be used to evaluate the spatial resolu-
tion of an imaging system. At sufficiently high spatial frequencies, individual
lines can no longer be separated after imaging, i. e., we have determined the
system’s resolution limit.
bA = cA · ∆s, (8.35)
with ∆s the sampling distance and cA a constant factor which represents the
reconstruction algorithm characteristics3 . Finally, the total blur is modeled
by q
btotal = b2F + b2D + b2M + b2A , (8.36)
whereas bacq represents the maximum spatial resolution given by the geomet-
ric setup, which could be achieved if we were to use a very fine sampling and
a reconstruction algorithm with a sharp kernel. It becomes obvious that as
a user, we only have limited influence on spatial resolution. We can decide
which convolution kernel we want to use, but the geometrical parameters are
defined by the system’s scan modes.
Spatial resolution can be measured directly and indirectly. For direct mea-
surement, a bar phantom can be used. Such phantoms consist of alternating
black and white bars of varying thickness. The resolution is determined by
evaluating whether bars of a certain thickness are still distinguishable after
acquisition and reconstruction (cf. Fig. 8.20). A more reliable and objective
evaluation is the indirect approach. For this purpose, we scan a thin wire
phantom4 , thereby obtaining the system’s so called point spread function
(PSF). The Fourier transform of the PSF yields the modulation transfer
function (MTF) (cf. Fig. 8.21). Frequency is typically measured in line pairs
per cm (lp/cm), a unit that can be intuitively understood if we recall the bar
phantoms of the direct approach mentioned before. The spatial resolution of
a system is often given by the 10% value of the MTF, which represents the
3
E. g., in filtered back-projection, a smooth convolution kernel reduces noise but also
spatial resolution, whereas a sharp kernel leads to more noise but yields a high spatial
resolution (cf. Sec. 8.3.1).
4
Essentially, this mimics a point object for each 2-D slice.
170 8 Computed Tomography
Figure 8.21: If we could scan an ideal point object, the resulting recon-
structed image would be the PSF of the system. The Fourier transform of
the PSF is the MTF, which shows the relative contrasts achievable for vary-
ing object frequencies. In practice, this measurement is typically performed
using thin wire phantoms.
Figure 8.22: The left image shows the reconstruction of a water cylinder
phantom. The noise is stronger in the center than in the peripheral regions.
On the right side, an elliptic phantom with two high intensity insets is de-
picted. In its center, streak noise emerges that is caused by the strongly
attenuating structures. For both phantom simulations, 90,000 photons per
pixel with an energy of 75 keV were used.
frequency at which the contrast has dropped to 10% of the maximum value
at 0 lp/cm.
8.4.2 Noise
From the considerations regarding noise in X-ray projections (cf. Sec. 7.4.3
(p. 136)), we know that the number of photons ns measured by our detector
8.4 Practical Considerations 171
Figure 8.24: Two regions within a pixel with different absorption coefficients
result in different measured intensities I for different projection angles.
In the first case, we observe two separate regions with the correspond-
ing absorption equations
I1 · e−µ1 ∆x , (8.38)
I2 · e−µ2 ∆x , (8.39)
Metal artifacts are among the most common image artifacts in CT imag-
ing. This term covers many different types of artifacts that we already dis-
cussed. There are various reasons why metal artifacts can occur. Metal causes
beam hardening and scatter, which results in dark streaks between the metal
objects. Additionally, its very large attenuation coefficient leads to photon
starvation behind the metal object; as most photons are absorbed, only an
insufficient number of them can be measured, leading to noisy projections.
The noise is amplified in the reconstruction and will lead to streak artifacts
in these regions, cf. Fig. 8.25.
Metal artifacts can be reduced by increasing the X-ray tube current or with
automatic tube current modulation. Alternatively, there are metal artifact
reduction algorithms that try to solve this problem without additional dose.
Some algorithms aim to remove the metal objects in the reconstructed image
and iteratively interpolate the holes in the forward-projected images.
Truncation occurs when a scanned object is larger than the detector area or
X-ray beams are intentionally collimated to a diagnostic region of interest
for saving dose. Both cases will result in laterally truncated data. Due to the
non-local property of the ramp filter, filtered back-projection reconstruction
requires information of the whole projections for each point in the object. This
176 8 Computed Tomography
Imono (E0 )
Z
qmono (E0 ) = −ln = µ (s, E0 ) ds. (8.47)
I0 (E0 )
8.5 X-ray Attenuation with Polychromatic Attenuation 177
Bone
Soft Tissue
attenuation coefficients
1
10
0
10
1 2
10 10
keV
Thus, for the monochromatic case, there is a linear relationship between the
monochromatic attenuation qmono (E0 ) and the intersection length of X-ray
and the object.
However, the X-ray sources in typical clinical scanners are polychromatic
sources. In addition, monochromatic measurements cannot provide real quan-
titative information as HUs are energy-dependent, i. e., different spectra and
filters will result in different HUs. Although there exist physical ways to cre-
ate monochromatic X-rays at sufficient intensity to perform X-ray CT, e. g.,
using a monochromator or inserting thick absorption filters to narrow the
spectrum, these methods are very expensive and the usage is restricted to
research experiments at few institutions. As detailed in Geek Box 8.7, in con-
trast to the monochromatic X-ray situation, there is no linear relationship
between the polychromatic attenuation qpoly and the intersection length of
X-ray and the object.
Fig. 8.30 depicts the relationship between the intersection length p and
the attenuation q when a polychromatic X-ray beam, which is emitted at a
tube voltage of 110 kV, penetrates a homogenous aluminium object, and the
relationship between the intersection length p and the attenuation q when a
monochromatic X-ray beam, which is emitted at the effective energy 45.56
keV of the aforementioned polychromatic X-ray beam, traverses the same
object.
178 8 Computed Tomography
where
S (E) D (E)
N (E) = (8.50)
I0
refers to the normalized energy spectrum with the effective detected
intensity (system weighting function) I0 defined by
Z Emax
I0 = S (E 0 ) D (E 0 ) dE 0 . (8.51)
0
4
x 10
15
2
Number of photons per mm
10
0
0 50 100 150
Energy E in keV
25 Monochromatic
Polychromatic
20
Attenuation
15
10
0 5 10 15 20
Intersection length [mm]
tities of these algorithms differ with respect to the diagnostic demands: they
range from energy-normalized attenuation values over physically motivated
quantities like density and effective atomic number to spatial distributions of
whole attenuation spectra. The most popular current diagnostic applications
are bone removal, PET/SPECT attenuation correction, lung perfusion diag-
nosis, or quantification of contrast agent concentrations, for instance in the
myocardium.
4
x=10
No=wedge=filter=(Eeff ==56.57=keV)
2
Number=of=photons=per=mm
12 0.5=mm=Al=(Eeff ==57.34== keV)
1.5=mm=Al=(Eeff ==58.69=keV)
10 10.5=mm=Al=(E
eff
==66.12=keV)
30.0=mm=Al=(Eeff ==74.18=keV)
8
0
20 40 60 80 100 120 140
Energy=in=keV
(a) (b)
Figure 8.33: Beam hardening example: (a) A simple phantom set-up con-
sisting of water (gray), bone (white) and air (black). (b) Reconstruction of
the phantom with visible beam hardening artifacts.
darker than the outer part, and a corresponding cupping appears in the
reconstruction. Streak artifacts appear as dark bands or streaks in the recon-
structed image. The occurrence of such artifacts is due to the fact that X-rays,
which pass through only one dense object, are less hardened than those pass-
ing through both dense objects. Fig. 8.33 shows a simulated example of an
elliptical water phantom with two dense bone insets.
Various beam hardening correction algorithms exist. For a soft-tissue cal-
ibration, projection measurements through soft-tissue like materials of vari-
able known thicknesses are performed. For these, the equivalent water thick-
nesses are known. A simple function is fit through the pairs of measured and
expected values. This function is inverted and then used as a look-up table:
182 8 Computed Tomography
for each measured attenuation, the equivalent water thickness is looked up,
which then replaces the measured attenuation. If bone-induced beam hard-
ening is also corrected, a separation into bone and water components can be
performed. For dual energy CT, special quantitative correction methods exist.
These take advantage of the two measurements at different energy weightings
and special properties of the attenuation functions of body tissues.
8.6 Spectral CT
The easiest method for producing spectral measurements is called Dual kVp.
For this method, two subsequent CT scans are performed at different tube
voltages, e. g., 80 kVp and 140 kVp; see Fig. 8.34(a) for spectra of these two
voltages. As mentioned before, it is desirable to use two spectra with as
little overlap as possible in order to ensure the maximal spectral separation
between the two acquired datasets. To this end, usually a special filter can
be used to attenuate the lower energy components in the high energy tube
spectrum (140 kVp); see Fig. 8.34(b).
The main advantage is that no special equipment is needed for this method.
In medical CT, this method is prone to motion artifacts as the alignment of
the two datasets cannot be ensured due to patient motion in between the two
scans. However, this is a valid method for evaluating spectral CT algorithms
on phantom data.
8.6 Spectral CT 183
0.07 0.035
Normallizedy80kVyspectrum
Normalizedp80kVpspectrum
Norm.y140kVyspectrumy(noytinyfilter)
Normaliizedpintensityp(a.pu.)
Normalizedyintensityy(a.yu.)
0.04 0.02
0.03 0.015
0.02 0.01
0.01 0.005
0 0
20 40 60 80 100 120 140 20 40 60 80 100 120 140 160
EnergyyinykeV EnergypinpkeV
(a) (b)
tube A
gantry
detector B
tube B detector A
8.6.1.2 KVp-switching
scintillator material
reflector / septa
photodiode layer
substrate / wiring
Dual Source CT is similar to dual kVp with the two CT scans being performed
simultaneously by a special CT system. In this system, the gantry houses
two tube-detector pairs A and B with a fixed angular offset (see Fig. 8.35).
The two X-ray tubes are operated at different tube voltages. More recent
systems offer an optional tin filter on one tube to increase spectral separation
whereas the two detectors are usually identical in terms of spectral sensitivity.
Most available systems, however, use differently sized detectors due to space
restrictions within the gantry. So the measurements of the smaller detector
offer a limited field of view (FOV). The data from the larger detector can be
used to compensate for truncation artifacts in the reconstruction but Dual
Energy data is only available for the smaller FOV. Since the two tube-detector
pairs are operated simultaneously, scatter radiation from tube A impairs
the signal of detector B and vice versa. This is a major drawback of this
technology, as this property decreases signal quality and leads to an increased
patient dose.
4
xv10
7
Countsvforv5vtov60vkeV
6 Countedvphotonsvabovev60vkeV
Overlap
Countedvphotons 5
0
40 60 80 100 120 140
EnergyvinvkeV
Two methods to recover cW (r) and cB (r) are presented in Geek Boxes 8.8
and 8.9.
Further Reading
[1] M. Balda. Quantitative Computed Tomography. PhD thesis, Friedrich-
Alexander-Universität Erlangen-Nürnberg, 2011.
[2] Martin Berger et al. “Marker-free motion correction in weight-bearing
cone-beam CT of the knee joint”. In: Medical Physics 43.3 (2016),
pp. 1235–1248. doi: 10.1118/1.4941012.
[3] T. M. Buzug. Computed Tomography: From Photon Statistics to Mod-
ern Cone-Beam CT. Springer, 2008. isbn: 9783540394082.
[4] Frank Dennerlein and Andreas Maier. “Approximate truncation robust
computed tomography - ATRACT”. In: Physics in Medicine and Bi-
oloagy 58.17 (2013), pp. 6133–6148. doi: 10.1088/0031-9155/58/17/
6133.
8.6 Spectral CT 187
Inserting Equation (8.53) into the line integral of the spectral atten-
uation law yields:
Z Z Z
µ (r, E) ds = fW (E) cW (r) ds + fB (E) cB (r) ds. (8.54)
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 9
9.1 Introduction
Modern medical imaging is achieved with state-of-the-art devices, that use
cutting-edge technology from different fields of engineering. Oftentimes,
progress is driven by discoveries in a at first sight seemingly unrelated field
of research. This enables the construction of new devices that was thought to
be impossible before. In this chapter, we introduce a new imaging modality
that has the potential to develop into a future medical imaging technology:
X-ray phase-contrast imaging. At its current state, its medical use still has
to be demonstrated. Yet, several early experimental results indicate that it
has some potential for clinical applications.
Conventional X-ray imaging measures the attenuation of X-rays. X-ray
attenuation happens due to different interactions between X-ray photons
c The Author(s) 2018
A. Maier et al. (Eds.): Medical Imaging Systems, LNCS 11111, pp. 191–205, 2018.
https://doi.org/10.1007/978-3-319-96520-8_9
192 9 X-ray Phase Contrast: Research on a Future Imaging Modality
ncomplex = 1 − δ + iβ , (9.1)
and matter (see Chapter 7 for details). However, attenuation does not fully
describe X-ray interaction with matter. In the early 1930s, phase contrast
microscopy, described in Chapter 5, was introduced. It measures refraction
of visible light, which provides an alternative to standard transmission mi-
croscopy for mostly translucent objects, such as biological cells. Since visible
light and X-ray are both electromagnetic waves, it is theoretically possible
to transfer this principle to X-ray imaging. To measure X-ray refraction has
two motivations: First, it may allow visualizing materials whose attenuation
properties differ only slightly. Second, it was assumed that refraction infor-
mation would deliver improved contrast over attenuation (see Geek Box 9.1
for details).
However, since the wavelength of X-rays is several orders of magnitude
smaller than the wavelength of visible light, it took several decades until
manufacturing technologies were developed to realize X-ray phase-contrast
imaging.
Several systems have been designed for phase contrast imaging. While all
of these approaches are highly interesting from a physics point of view, we
limit the presentation in this chapter to the Talbot-Lau Interferometer (TLI).
A short description of other approaches can be found in the Geek Box 9.2.
Most of these other approaches are more sensitive than TLI, but TLI’s relative
9.1 Introduction 193
mild system requirements make it currently the most attractive system for
implementation in a clinical setup. In particular, it only requires additional
gratings to be mounted between a regular medical X-ray tube and detector to
be operated. Note that all of the methods presented in this chapter are current
research topics and none of them are clinically used at present. However, it is
expected that the presented methods will have clinical impact in the future.
194 9 X-ray Phase Contrast: Research on a Future Imaging Modality
∆d = d sin(θ) , (9.2)
where d is the distance between the two origins. Here, we assume that the
distance D between the slits and the screen is large, such that ϑ is close to
zero and θ ≈ ϑ. Constructive interference shows as a bright spot. It occurs
when the waves arrive “in phase”. This is the case if the path difference is
zero or an integral multiple of the wavelength,
θ ϑ
d
X-ray
Source ∆d
D
Interference
Pattern
Detector
y Object
Analyzer
Phase Grating G
X-ray Source Grating G1
2
z
Source
Grating G0
x
grating bar and a grating slit must now add up to the imprinted phase shift for
constructive interference. Due to the Talbot effect, these interference patterns
cleanly overlay at the so-called Talbot distances. Fig. 9.3 shows a simulation
of the interference pattern behind grating G1 . The wave is assumed to travel
from left to right. The detector is placed at a location where constructive
and destructive interference form a clearly distinguishable black-and-white
pattern. However, the pattern periodically repeats, which is why a system
designer can in theory choose from infinitely many positions to place the de-
tector. In practice, however, the pattern quickly washes out, such that the
detector has to be located at one of the first replications of the pattern.
Gratings G0 and G2 are used to solve several engineering problems such
that the interferometer can be built within compact medical X-ray setups.
The interference pattern at the detector is normally much smaller than a
single detector pixel. To resolve the pattern, the analyzer grating G2 is used.
G2 is an absorption grating such that only a part of the interference pattern
passes through the slits onto the detector. This makes it possible to sample
the interference pattern by taking X-ray images while moving G2 along the
pattern (which is further discussed in Sec. 9.2.2).
Grating G0 addresses another practical problem, namely the size of the
focal spot. Interference effects can only be observed on coherent waves. In the
sketch on Young’s double-slit experiment in Fig. 9.1, coherence is obtained
automatically by the small slit behind the X-ray source, which effectively acts
as a small focal spot. There exist so-called microfocus X-ray sources that do
provide such a small focal spot but they produce only very few X-rays per
time. For practical applications, the imaging times are typically much too
long. Medical X-ray tubes produce orders of magnitude more X-rays, which
allows to take an X-ray image within a fraction of a second. However, such
X-ray tubes can only be built with a focal spot that is much larger, typically
between half a millimeter and a millimeter. To overcome this issue, grating G0
is used, which can be seen as a long array of micrometer-sized slits. Each slit
acts as a microfocus spot and creates an interference pattern at the detector.
The distances between the slits are now chosen in a way that the pattern of
each of the slits exactly overlays with the pattern of the other slits. If the
setup parameters are chosen correctly (cf. Geek Box 9.3), all these periodic
structures at the detector are aligned and add up. This enables imaging using
conventional X-ray sources.
9.2 Talbot-Lau Interferometer 197
One quality measure of the TLI can be derived from the system setup
directly, the so-called sensitivity s,
dist(G1 , G2 )
s= , (9.5)
2πp2
defined by the distance between G1 and G2 and p2 which is the period of the
analyzer grating G2 . The sensitivity can be interpreted as an “amplification
factor” of the refractive angle.
Phase stepping denotes the process of shifting one of the gratings (typically
G2 ) by a fraction of its period to sample the interference pattern. This is
198 9 X-ray Phase Contrast: Research on a Future Imaging Modality
Analyzer
Grating G2
0.6 Measurements
Intensity value
Stepping Fitted Sine
0.5
direction
0.4
0.3
0.2
0.1
Stepping Stepping 2 4 6 8
position 2 position 6 Phase steps
Figure 9.4: Left: phase stepping. The Talbot carpet is sampled at different
positions of grating G2 . Right: the intensities at different stepping positions
form the sinusoidal phase stepping curve. From this curve, attenuation, dif-
ferential phase and dark-field can be calculated.
shown in Fig. 9.4 (left): the Talbot carpet forms the interference pattern. At
equidistant stepping positions, the detector records an image. The intensity
values for a pixel at different stepping positions is called phase stepping curve.
This curve can be considered as a convolution of the profile of G2 with the
intensity pattern of the G1 pattern. Convolution of two rectangular functions
of G1 and G2 leads ideally to a triangular signal. In reality, blurring from the
slits of G0 leads to a sinusoidal curve that is fitted to the measured steps.
These steps are shown in Fig. 9.4 (right). After fitting a sine function to
the phase stepping curve, three quantities can be calculated: attenuation,
differential phase, and dark-field. Attenuation is the offset of the intensities,
which can be computed as the average of all intensities of the phase stepping
curve. Differential phase is the phase offset of the sine. Dark-field is one minus
the ratio between amplitude of the curve and two times attenuation.
In practice, these three quantities cannot be calculated directly. Instead,
it is necessary to acquire two scans: a reference scan and an object scan.
The reference scan is acquired without an object in the beam path such that
it only shows the background. It captures inhomogeneities in the setup, for
example, in the grating bars. The object scan is acquired with an object
before or behind G1 . The reference scan is used to normalize the object
scan after calculating attenuation, differential phase, and dark-field for both
scans. Several works address the further suppression of imaging artifacts in
software [9, 8, 11]. More details on the calculation of attenuation, phase, and
dark-field can be found in Geek Box 9.4.
A common metric for the quality of an interferometer is visibility. Visibility
is a measure of contrast of the intensity modulation and is given by the sine
amplitude divided by its offset. Thus, the dark-field signal corresponds to the
reduction in interferometer visibility, for example, due to micro scattering.
9.3 Applications 199
max(R) − min(R)
VR = ,
max(R) + min(R)
9.3 Applications
An example for the three resulting signals is shown is Fig. 9.5. The shown
gummi bears are modified with artificial defects, namely powder, a needle,
200 9 X-ray Phase Contrast: Research on a Future Imaging Modality
Figure 9.5: Top left: photograph of gummi bears with artificial defects. At-
tenuation (top right), differential phase (bottom left) and dark-field (bottom
right) visualize different modifications in the gummi bears. Differential phase
is particularly sensitive to high-frequency variations in the signal. Dark-field
is particularly sensitive to micrometer-sized structural variations, such as
powders or fibers. Pictures courtesy of ECAP Erlangen.
and a toothpick. In the top right, the absorption image is shown, which clearly
shows the metal needle. On the bottom left, the differential phase is shown,
which shows a large amount of high-frequency details, including the center of
the needle head and the powder structure. In the bottom right, the dark-field
signal is shown, which is particularly sensitive to the fine-grained structural
variations in the powder and the toothpick.
An example scan on biological data is shown in Fig. 9.6. From left to
right, the images show attenuation, differential phase, and dark-field of a
female breast [10]. Particularly interesting in this visualization is the dark-
field image, which is particularly sensitive to microcalicifications in the breast,
a common indication for breast cancer.
Overall, X-ray attenuation visualizes both variations in density and atomic
number. Thus, for example, it excels at the visualization of bones. Bones are
more dense than the surrounding tissue and also contain a substantial amount
of calcium.
Phase information on the other hand is sensitive to variations in electron
density. Thus, it is expected to deliver increased contrast over attenuation
when there is a similar elemental composition between two structures. One
example is imaging of soft tissues. Talbot-Lau interferometers can only obtain
differential phase information. Hence, its main advantage is in imaging high
frequency details, such as edges that lie perpendicular to the grating bars.
Conversely, it is less effective for imaging low-frequencies information.
Dark-field imaging provides two interesting properties that set it apart
from absorption and differential phase: First, dark-field is sensitive to den-
9.3 Applications 201
sity variation at nano- and micrometer scale. This is typically below the
resolution of the detector, and as such too small to be resolved in attenua-
tion imaging. Second, when scanning ordered structures such as fibers, the
dark-field intensity varies with the angle between the fibers and the gratings.
Both properties together can be used to deduce the orientation of fibers below
the resolution limit of the detector [7, 12, 1]: when performing tomography
on a plane in which micrometer-scaled fibers are located, the dark-field signal
oscillates. From this oscillation, the direction of the fibers can be deduced,
although the individual fibers are too small to be resolved by the detector.
An example is shown in Fig. 9.7. On the left, a wooden block with different
layers of wooden fibers is shown. This block is scanned in a tomographic
setup, and the fiber orientations are deduced from the signal oscillations [1].
On the right of Fig. 9.7, the reconstructed orientations are shown, in different
colors per layer.
Overall, phase-contrast and dark-field signals offer several interesting prop-
erties. It depends on the imaging task to decide whether the offerings of phase
and dark-field signals or conventional attenuation is advantageous. The list
below enumerates applications where Talbot-Lau imaging can potentially of-
fer an advantage over traditional absorption imaging.
• Mammography. Mammography relies on imaging soft tissue structures
in the breast, as well as on the detection of micro-calcifications. Imag-
ing of soft tissues could benefit from the phase signal since it provides a
strong signal to noise ratio at high frequencies. It has also been shown
that the dark-field signal can reveal micro-calcifications that are invisible
in the attenuation image since their porous structure creates dark-field
signals [13].
202 9 X-ray Phase Contrast: Research on a Future Imaging Modality
Z
X
Figure 9.7: Layered wooden block (left) and example dark-field reconstruc-
tion of the dominant fiber orientations in each layer (right).
• Lung imaging. The human lung relies on millions of small alveoli to per-
form gas exchange. Lung diseases such as chronic obstructive pulmonary
disease (COPD) and pulmonary fibrosis lead to a change in the structure
of the alveoli. However, due to their small size, they cannot be resolved
individually. The dark-field signal is able to detect abnormalities in the
alveoli microstructure and may thus provide a benefit when diagnosing
these diseases [23].
• Bone imaging. The directionality of the dark-field signal can possibly be
used to detect osteoporosis, which can lead to less aligned structures in the
bone [21]. Furthermore, the phase signal can visualize low contrast struc-
tures such as cartilage and tendons which may not be visible in attenuation
imaging.
• Micro-CT. CT scanners which provide high resolutions (voxel sizes in
the size of micrometers) are called Micro-CT systems. They are used for
analyzing small samples and animals. At high resolutions, phase-contrast
CT delivers a higher image quality than conventional CT [16]. This can
be explained by the fact that the recorded phase signal is differential. A
first Talbot-Lau Micro-CT system is commercially available [20].
• Industrial applications. Beyond medical imaging, dark-field imaging
has been applied to non-destructive testing [18, 17]. For example, it can
detect defects in carbon fibers or foreign bodies in food that are unde-
tectable using attenuation imaging.
9.4 Research Challenges 203
edges. This makes its tomographic reconstruction challenging and has led
to the development of dedicated algorithms to solve this problem [1].
Further Readings
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 10
Emission Tomography
10.1 Introduction
In contrast to the structural imaging used to visualize tissues in the body,
functional imaging is used to observe biological processes. In the field of
nuclear medicine, functional imaging relies on radioisotopes that are tagged
to tracers whose biochemical properties cause them to congregate at regions
of diagnostic interest in the body. As opposed to transmission tomography
with X-ray CT, where the source of imaging radiation is a part of the imaging
device, the radiation source in this case is located within the patient. For
this reason, functional imaging methods in the field of nuclear medicine –
also known as molecular imaging – belong to a family of modalities called
emission tomography, whose differing physical properties make them quite
distinct from the transmission case.
The process begins with radioactive decay, which results when an unstable
isotope ejects particles from its nucleus while transitioning to a stable state.
Although a very complicated process, two modes are of interest to molecular
imaging: γ and β. In the former case, gamma rays are ejected directly from
c The Author(s) 2018
A. Maier et al. (Eds.): Medical Imaging Systems, LNCS 11111, pp. 207–236, 2018.
https://doi.org/10.1007/978-3-319-96520-8_10
208 10 Emission Tomography
the nucleus that can be imaged with a so-called gamma camera. 3-D images
can then be reconstructed from 2-D projections in a process called SPECT. In
the second case, a positron is emitted which travels a small distance until an
electron (its antiparticle) is encountered. The ensuing annihilation produces
a pair of 511 keV photons traveling in opposite directions that, when detected
simultaneously, yield lines of response that can be reconstructed into an image
in a process known as positron emission tomography (PET).
Although X-rays produced by bombarding targets with electrons had been
in use since their discovery by Röntgen in 1895, the use of naturally decay-
ing radioisotopes for medical imaging did not occur until 1935, when George
de Hevesy investigated rats injected with radioactive 32 P. Using a Geiger
counter, de Hevesy investigated the relative amount of radioactivity in differ-
ent organs after dissection and found that the skeleton had a disproportion-
ately high level of uptake. In doing so, he not only settled once and for all the
ongoing medical question of whether or not bones have an active metabolism
(they do, otherwise they would not have taken up the 32 P atoms), but he was
also the first to use radioisotopes and imaging equipment to investigate the
body’s biochemistry. Thus, the so-called tracer principle was born. For his
work in the field of radiotracers, de Hevesy was awarded the Nobel Prize for
Chemistry in 1943, and a variant of his original bone-imaging methodology
based on phosphates is still in wide use today. In the decades following de
Hevesy’s discovery, research from the field of radiochemistry and molecular bi-
ology have yielded a plethora of tracers with desirable uptake characteristics.
Complimentary technical advances have provided imaging devices capable of
aiding physicians answer a range of diagnostic questions.
(a) (b)
Figure 10.2: Exponential decay curve for a 100 MBq sample of a radioiso-
tope having a half-life of six hours (the same as 99m Tc).
isotope’s half-life is six hours, which is the same as that of 99m Tc, the most
commonly used isotope for SPECT imaging.
Although Eq. (10.1) represents a sample’s aggregate decay properties, the
emission of individual photons (or photon pairs for β decay) within a partic-
ular time window is a discrete process and follows a Poisson distribution with
a mean ν proportional to the amount of radioactivity present. Note that we
can assume independence between the voxels and describe an entire image
in a vectorized format using for a single voxel and ν for an entire image.
Similarly, the number of photons counted in a particular observation of this
process, such as a pixel of a SPECT projection, is a Poisson-distributed ran-
dom variable as well, provided that the image formation chain is linear.1 If
1
In practical scanners, this is not strictly the case, but it is assumed to be here.
210 10 Emission Tomography
we represent the projection pixels and image voxels as vectors, the distribu-
tion of photon counts on the detector D is related to the activity distribution
being imaged ν via the following relation:
d̄ d̄ p
SN R = = √ = d̄. (10.3)
σd d̄
The SN R is thus simply the square root of the mean number of photon counts
in the object and increases monotonically, albeit with plateauing benefits,
with the number of counts in the image.
In X-ray CT imaging, where the radiation source is located outside the
body and can be easily controlled by the system, large numbers of photons
are easily attainable, as the patient can be irradiated with a high flux for a
short period of time. However, in molecular imaging the radioactive source is
10.2 Physics of Emission Tomography 211
(a) t = 0
(b) t = 2t1/2
(c) t = 4t1/2
(d) t = 6t1/2
Figure 10.3: Simulated images (left) and central horizontal profiles (right)
from an object filled with the activity described in Fig. 10.2. The images were
simulated after zero, two, four, and six half-lives (a, b, c, and d, respectively).
The mean value of the object is shown with a dashed red line through each
profile. Note how the images become noisier as the mean decreases.
212 10 Emission Tomography
located within the body and will continue to bombard tissue with potentially
harmful ionizing radiation until it either decays or is excreted by the body.
Therefore, to limit patient dose, relatively small amounts of activity are
usually injected, typically ranging from 100 to 1,000 MBq. The activity is then
distributed throughout the body, leading to low numbers of emitted photons
at any given area. The imaging task is thus similar to taking a photograph
in a very dark room. A long exposure time can yield a better SN R, but
comes with problems of motion blur and patient discomfort. A typical SPECT
projection lasts 15 seconds, resulting in a total scan duration of 30 minutes
for the usual 120 projections! Despite this effort, projections typically have
only about 20 or fewer useful photons per pixel in diagnostically interesting
areas. A representative projection from a skeletal SPECT scan in shown in
Fig. 10.4. The mean pixel value is a measly 0.6, and maximum is only 17,
significantly less than even the noisiest simulation in Fig. 10.3(d). Due higher
scanner sensitivities, PET statistics are slightly better, with roughly a factor
of 10 more counts per pixel at typical scan durations of 4–6 minutes for an
equivalent field of view.
The fundamental challenge in emission tomography is therefore to pro-
duce reconstructed images of the activity distribution ν with acceptable im-
age quality from noisy acquired data. The following sections describe other
physical issues encountered as well.
Aside from the fundamental problem of noisy data, the second most im-
portant physical factor affecting emission tomography is photon attenuation.
Photons traveling through a medium may interact with atoms and eventually
be absorbed, resulting in a detected flux I less than that originally emitted.
In Chapter 8, we learned how to describe this principle using Beer’s law and
10.2 Physics of Emission Tomography 213
exploit it for imaging. For transmission tomography like X-ray CT, this phe-
nomenon is imaged directly to yield reconstructed images of the medium’s
(i. e. patient’s) structure. This is possible because the location and current
of the emitted flux I0 is known. In emission tomography, however, I0 is de-
termined by the activity distribution in the body ν, which is unknown. At-
tenuation is therefore a hindrance that leads to errors if it is not accounted
for.
Amongst the photon-matter-interactions, Compton scatter is very impor-
tant for emission tomography (cf. Sec. 7.3). In this interaction, the photon
is not absorbed as in attenuation, but merely deflected. The relationship be-
tween deflection angle θ and pre- and post-collision energies E0 and Escat is
described by the Klein-Nishina formula:
E0
Escat = . (10.4)
1+ ( 511 )(1
E0
− cos θ)
Scatter is an important component of emission tomography due to its role
in the degradation of image quality. Specifically, deflected photons may be
erroneously assumed to come from locations in the image volume along their
scattered trajectories, rather than their original paths. This process is illus-
trated in Fig. 10.5, where a photon originating in the image is scattered and
counted at a detector pixel corresponding to a trajectory other than its orig-
inal (vertical) path. This has the effect of reducing resolution, contributing
to image noise, and reducing contrast.
214 10 Emission Tomography
10.3.1 SPECT
Db (Lb + z)
rPSF ≈ , (10.5)
Lb
where Db is the bore diameter, Lb its length, and z the distance between the
source plane and the face of the collimator. From Eq. (10.5), it can be seen
that the resolution is depth-dependent and becomes wider with increasing z
for given collimator dimensions.
An image of a point source showing a true PSF is shown in Fig. 10.8.
The image is saturated to highlight the complex structure. In the bright
central area outlined in red, primarily geometric photons are present. In the
216 10 Emission Tomography
Figure 10.8: Measured PSF from a 99m Tc point source imaged at a distance
of 10 cm, shown saturated to emphasize low-intensity regions. The geomet-
ric bright geometric is outlined in red, and most extra-geometric counts lie
between the red and blue hexagons, where a single partial septal wall is pen-
etrated.
and resolution. The former directly impacts the quality of counting statis-
tics, and therefore noise, in an acquired image. The latter is related to the
accuracy with which the detector can localize them and properly reproduce
small features such as edges. A third consideration appears via the septal
thickness which, when increased, limits the star artifacts shown in Fig. 10.8
at the expense of smaller ρ.
Once a photon has passed through the collimator, it impacts the system’s
scintillator (typically composed of NaI), releasing several lower energy pho-
tons in the visible range. These photons then travel to the PMTs, where
they initiate an electron avalanche that is detected as a current signal at the
PMT output. To determine the 2-D location of a photon, a type of centroid
is computed by the output electronics of the PMT array in a process known
as Anger Logic, after its inventor. In the 1-D case, the estimated location of
the photon detection x̂ can be calculated as
P
q x q Gq
x̂ = P , (10.7)
q Gq
where Gq and xq are the signal strength at and location of the q-th PMT.
Applied in this fashion directly, images will suffer from nonuniformities and
pincushion distortions. These are removed by replacing Gq with some non-
linear function thereof. Even after this correction, the method is not exact,
and the resulting finite resolution rDET addsp in quadrature with that of the
PSF to yield a total system resolution rSYS = rPSF2 +P 2
rDET . Another impor-
tant property of the PMT output is that the value of q Gq is proportional
to the energy of the initial photon. This allows SPECT cameras to be energy
resolving as well, allowing the effects of Compton scatter to be mitigated.
10.3.2 PET
As shown in Fig. 10.1, the β decay that forms the basis of PET produces
two photons that travel in opposing directions away from each other. This
is exploited for imaging purposes by using a ring detector and looking for
coincidences in the observed data. This coincidence detection principle is
illustrated in Fig. 10.9, where a PET ring composed of many small detector
blocks is shown. Extremely high speed electronics monitor each detector’s
output signal and record a detection event when two impulses are detected
simultaneously. The detector blocks themselves are traditionally composed of
a scintillator crystal mated to a small PMT array as with the Anger camera.
However, no collimator is needed to restrict the scintillator’s acceptance angle
in this case because the photon’s incidence angle is implicitly provided by the
detector block at the opposite side of the ring. Nevertheless, inaccuracies in
the scintillator blocks and PMTs still induce a finite PSF in PET whose
218 10 Emission Tomography
Figure 10.9: PET ring detector and coincidence detection principle. The
detector electronics simultaneously monitor signals from each detector block
and record counts when an impulse is detected from two blocks at the same
time.
(a) (b)
Figure 10.10: 2-D (left) and 3-D (right) detection configurations for PET.
The latter offers better sensitivity at the expense of more scatter events.
TOF PET systems with 3-D detection thus typically offer superior res-
olution and noise characteristics compared to SPECT, but this comes at a
price. 18 F, the most common isotope used in PET, has a half life of only 110
minutes and is more difficult to produce than 99m Tc, requiring a complex
logistical network to minimize the time between production and injection.
Furthermore, the higher energy photons imaged in PET require costly ex-
otic scintillator materials. This, combined with highly specialized detector
electronics, makes PET systems more expensive to procure and operate than
their SPECT counterparts.
10.4 Reconstruction
age and their projected pixels at the detector. This ignores all of the other
physical factors, such as attenuation, scatter, and the PSF, that play a vital
role in the emission tomography image formation chain. This oversight leads
to artifacts in reconstructed images that greatly degrade image quality. For
these reasons, filtered back-projection is generally no longer used in clinical
situations.
The relationship described in Geek Box 10.1 is quite complex, and it is not
immediately clear how to maximize the likelihood. However, a seminal paper
by Shepp and Vardi in 1982 showed that this can be accomplished via the Ex-
pectation Maximization (EM) algorithm, whose general framework involves
the estimation of the “complete” information, given a set of observations and
hidden, “latent”, information. Although a detailed description is outside the
scope of this text, it is worth outlining that for emission tomography, the
complete information is the actual emission distribution ν, and the observa-
tions are the counts in the projections d. The latent information is comprised
of all of the photons originating in the image that escape detection.
As shown by Shepp and Vardi, EM’s methodology of alternatingly forming
a conditional expectation via marginalization over a particular variable and
then maximizing the resulting likelihood provides a convenient framework for
attacking Eq. (10.8) as encountered in emission tomography. This expecta-
tion/maximization cycle is repeated until a suitable image is obtained, and
each one of these repetitions is referred to as one iteration k. The algorithm
begins by initializing some estimate of the activity distribution ν̂ 0 . The pro-
cess proceeds at each iteration by forward projecting the current estimate ν̂ k ,
comparing it to the measured data, backprojecting the result, and applying
10.4 Reconstruction 221
νd
P (D = d) = e−ν ,
d!
which is known as the likelihood of the observation.
b) Moving one step further, where an array of observations d if formed
by photons emitted from a vector of voxels with means ν. This is the
same scenario we examined in the example in Fig. 10.3. Here, the
system matrix is equivalent to the identity matrix A = I , and each
voxel contributes to a single detector element. As each observation is
independent of the others, we can multiply them together to obtain
our likelihood:
Y ν di
P (D = d) = exp (−νi ) i ,
i
di !
where i represents the index of the detector and image elements, which
are equivalent in this case.
c) In a true imaging scenario, A 6= I , and multiple image voxels
contribute to a single detector element. To account for this, we must
subdivide the total detected counts
P in each pixel dm into contributions
from each image voxel: dm = n dm (n). The probabilities contained
in the system matrix must also be included. The total likelihood is
therefore the sum over each of these possible scenarios:
d (n)
Y (νm am,n ) m
P (D = d) = exp (−νm am,n )
m,n
dm (n)!
The dual product has the effect of incorporating the contribution from
each voxel to each pixel.
222 10 Emission Tomography
a weight to the estimate to create a new ν̂ k+1 .The update mechanism for the
algorithm can be expressed using the following equation:
ν̂nk X dm
ν̂nk+1 = P an,m P k . (10.9)
am0 ,n m ν̂n0 am,n0
m0 n0
X
ν̂ ∗ = argminkd − Aν̂k2w = argmin w m (d m − [Aν̂]m )2 , (10.10)
ν̂ ν̂ m
where [Aν̂]m is the m-th pixel of the forward projected estimate and w is
a vector of weights. The weights are often used to take noise into account
by, for example, setting each element of w equal to an estimate of the vari-
ance at the corresponding detector pixel. This has the effect of adjusting
each pixel’s contribution to the objective function depending on its noise
properties. The weighted least squares objective function is convex and can
be solved with gradient-based optimization techniques such as the conjugate
gradient method.
not only useful for individual patient management, but enables larger, multi-
center clinical studies as well.
Assuming an accurate system model is available, the cornerstone of a quan-
titative imaging system is the calibration. This anchors the counts observed
during an acquisition to a physical amount of radioactivity in the detector’s
field of view. A common way of doing this is to perform an acquisition on
a homogeneous phantom with a known activity concentration expressed as
ml . A volume of interest may then be defined in the reconstructed image,
kBq
and a count density in units of counts per ml may be determined. Time must
then be taken into account by correcting for decay and normalizing by the
acquisition duration. After these steps, a volumetric sensitivity factor αVOL
may be defined relative to its units as follows:
counts
αVOL = minute · ml . (10.11)
kBq
ml
With this factor in hand, subsequent acquisitions may be quantified, provided
they are acquired with the same isotope and reconstructed in the same way.
The procedure for this is straightforward and consists of obtaining the count
rate density from a volume of interest in units of min·ml
counts
and dividing by
αVOL , thus producing the desired absolute units of ml .
kBq
Figure 10.11: 1-D signal (green) convolved with gaussian to yield “observed
data” that is reconstructed (i. e. deconvolved) using the MLEM algorithm
(blue curve). In this case, the blurring function is not modeled, and the
reconstruction cannot improve upon the observed data (the two curves are
equal). Figure courtesy of Siemens Molecular Imaging Inc., USA.
Figure 10.12: Reconstruction after six MLEM iterations where the blur-
ring function is modeled in the system matrix. Edge resolution is improved
over Fig. 10.11, but ringing artifacts also become visible. Figure courtesy of
Siemens Molecular Imaging Inc, USA.
Figure 10.13: MLEM reconstruction after 300 iterations. The edges have
become sharper, and artifacts amplified relative to Fig. 10.12. Figure courtesy
of Siemens Molecular Imaging Inc., USA.
Figure 10.14: Six MLEM iterations with on data perturbed with Poisson
noise. The result is slightly irregular but comparable to Fig. 10.12. Figure
courtesy of Siemens Molecular Imaging Inc., USA.
Figure 10.15: 300 MLEM iterations on noisy data. The interior of the large
object is highly irregular, and quality is noticably worse than Fig. 10.13.
Figure courtesy of Siemens Molecular Imaging Inc., USA.
some resolution and thus entails making the same type of compromise. In
practice, the choice of many reconstruction parameters is therefore another
example of the bias/variance trade-off already discussed above with respect
to SPECT collimator design.
Another complication for iterative reconstruction in emission tomography
is that there are source-dependent factors such as attenuation and scatter in
the system matrix. This implies that the properties of the reconstruction will
vary from patient to patient, even if the same acquisition and reconstruction
parameters are used. Furthermore, depth- and position-dependent PSFs in
SPECT and PET lead to shift-variant properties within a given image as
well. These factors should be taken into account when reconstructing and
interpreting images.
10.5.1 Diagnostics
As the most common use for emission tomography, diagnostics is also the
most diverse. In the field of neurology, both PET and SPECT offer perfusion
tracers that give physicians insight into the amount of blood flow in the
brain during the scan, which is proportional to brain activity. An example
of a SPECT brain perfusion procedure using 99m Tc-Ethylcysteinat-Dimer
(ECD) is shown in Fig. 10.16(a). An epileptic patient is scanned immediately
following a seizure and during a neutral state. The reconstructed images are
subtracted and fused with an MR of the patient’s brain to localize the focus
of the seizure. More specialized applications including imaging of amyloid
plaques linked to Alzheimer’s disease (PET) and dopamine receptor imaging
(SPECT) are also available.
For oncology, 18 F, the most commonly used PET isotope, may be bonded
to a molecule in the glucose family resulting in so-called fludeoxyglucose
(FDG). Using these FDG-PET scans, doctors can search for areas with
high glucose metabolism – a sign of rapidly growing metastatic tumors.
Fig. 10.16(b) shows an FDG-PET scan of a patient with melanoma. Ma-
lignant metastases are visible below the liver and beside the heart. A com-
mon oncological use for SPECT is skeletal imaging with 99m Tc bonded to
phosphorous compounds. High uptake of these tracers is often indicative of
secondary lesions from e. g. prostate or breast cancer.
228 10 Emission Tomography
(a) (b)
Figure 10.17: Same breast cancer patient imaged on three different dates
with 99m Tc-DPD. The calculated SUVmax from the volume of interest at the
posterior right area of the skull is also shown. A decrease in uptake was noted
between the second and third scans, indicating a response to therapy. Images
courtesy of the University Hospital Erlangen, Clinic of Nuclear Medicine.
begun. In the final scan shown in Fig. 10.17(c), SUVmax decreased, indicating
a response to therapy.
230 10 Emission Tomography
10.5.2 Therapy
SPECT and PET offer excellent sensitivity for the detection of disease due
to the functional information they provide. However, pathological regions of
an image may be difficult to localize in the body in the absence of structural
information.
Take, for example, a hypothetical surgeon who is planning a biopsy and
needs to find the specific Sentinel Lymph Nodes (SLNs) draining a tumor in a
breast cancer patient. Prior to surgery, a SPECT scan has clearly shown the
presence of an SLN with high uptake in the underarm area, but it is known
that there are multiple possible lymph nodes here. This stand-alone SPECT
might appear similar to the left pane of Fig. 10.18, where only a single bright
spot is visible. How will the surgeon proceed?
Historically, during planar acquisitions, a technologist might trace the out-
side edge of a patient’s body with a radioactive “pen” to provide a rough
anatomical point of reference in the image. The advantages of this method
are limited, and it is, in any case, not possible for SPECT, where attempts
may be made to register a previously acquired CT to the current SLN SPECT
study. However, the human anatomy is non-rigid, and shifts in posture and
time between scans may lead to errors. Our surgeon would therefore be left
with the option to operate in the general area of the suspected SLN and rely
on tedious scanning with gamma counting probes to find the exact node.
10.6 Hybrid Imaging 231
Figure 10.18: SPECT data after labeling of a Sentinel Lymph Node (SLN)
with 99m Tc-Nanocoll. The corresponding structural information from a com-
plimentary X-ray CT scan helps provide proper localization of the activ-
ity. Images courtesy of the University Hospital Erlangen, Clinic of Nuclear
Medicine.
P (d|ν)P (ν)
ν̂ ∗ = argmax P (ν|d) = argmax , (10.13)
ν̂ ν̂ P (d)
own, we could work this prior knowledge from MR into the objective function
of an iterative reconstruction algorithm.
The family of maximum a posteriori (MAP) algorithms is capable of doing
exactly this by building upon the maximum likelihood method with a term
representing some prior information known about the object. As the name
implies, the MAP method seeks to maximize the posterior probability of the
observed data given the distribution being imaged. We explain this principle
in Geek Box 10.2.
Another example of this higher level of integration, although one not re-
lying on the MAP principle, is the xSPECT Bone algorithm from Siemens,
which is currently used for reconstructing SPECT skeletal scans. This method
works by segmenting the CT into several different tissue classes and forward
projecting them separately at each iteration. In addition to voxel-wise image
updates, the classes themselves are also allowed to be scaled independently
while optimizing the objective function. This scaling allows the SPECT re-
construction to have very sharp edges at the boundaries between tissue classes
(e. g. if a cortical bone class has much more uptake than a neighboring lung
region), while maintaining a typical SPECT-like resolution within each class.
An example of this method is shown in Fig. 10.19 below. Note how the
edges of the vertebrae in the xSPECT image (center) are much sharper than
234 10 Emission Tomography
the standard MLEM SPECT reconstruction (left) due to the boundary in-
formation provided by the CT (right). However, the bladder appears very
similar in the two SPECT reconstructions, as there is little additional CT
boundary information here.
Despite the advantages of methods such as MAP and xSPECT Bone, there
are risks as well. For example, a MAP method may assume that bone den-
sity is always positively correlated to tracer uptake and enforce this behavior
to improve quantitative accuracy. This is indeed generally the case, but in
the early stages of a bone infarction, there is little or no blood supply to
the bone and, hence, no tracer uptake, despite a normal CT. Our example
MAP algorithm would then try to allocate activity here during the recon-
struction and potentially provide the physician with a false negative. One
should therefore be very careful when designing priors, as they are generally
based on assumptions about anatomy or biochemistry that may not be true
in all cases.
Further Reading
[1] Adam Alessio and Paul Kinahan. “PET Image Reconstruction”. In:
Nuclear Medicine 2 (2006).
[2] Adam Alessio and Paul Kinahan. “Quantitative Accuracy of Clinical
99mTc SPECT/CT Using Ordered-Subset Expectation Maximization
10.6 Hybrid Imaging 235
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 11
Ultrasound
11.1 Introduction
Acoustic waves with frequencies ξ between 16 Hz and 20 kHz can be sensed
by the human hearing and are thus called audible waves or audible sound. If
ξ > 20 kHz, one speaks of ultrasound (Tab. 11.1). Some animal species such
as bats can perceive ultrasound and use it for echolocation: by measuring
the time between sending and receiving (after partial reflection on a sur-
face) ultrasonic waves, the distance of an object (e. g., a wall or prey) to the
sender (bat) can be computed accurately, assuming that the sound velocity
is known. In the previous century, modern technology started to make use of
this technique with applications ranging from marine distance measurement
(1920: SONAR) to medicine (1958: first ultrasound device in clinical use). A
typical system is shown in Fig. 11.1.
Today, medical ultrasound often is the first-resort clinical imaging modal-
ity due to its cost-effectiveness and lack of ionizing radiation. Typical medical
ultrasound frequencies are between 2 MHz < ξ < 40 Mhz. Traditionally, med-
ical ultrasound is mainly put to use in diagnostic applications, however, more
therapeutic applications are emerging.
f Examples
Infrasound 0 . . . 16 Hz Seismic waves
Music
Audible sound 16 Hz . . . 20 kHz
Human Speech
Bat, Dolphin, and Whale Sounds
Ultrasound > 20 kHz Acoustic Microscopy
Ultrasound Imaging
The human body contains various kinds of boundaries between different ma-
terials, for instance, at borders between organs and liquids or other tissues.
At such boundaries between two media, sound waves are partially reflected
and partially transmitted.
11.2.2.1 Reflection
The well known law of reflection states that the angle of incidence equals the
angle of reflection. This also holds for reflection of sound waves (cf. Fig. 11.2(b)).
For perpendicular incidence (cf. Fig. 11.2(a)), the reflection R and transmis-
sion T coefficients write:
2
Z2 − Z1
Jr
R= = (11.2)
J0 Z2 + Z1
Jt 4Z1 Z2
T = = , (11.3)
J0 (Z1 + Z2 )2
240 11 Ultrasound
Z1 Z2
Transducer α2
Jr α1
Jt
J0
Z1 Z2
(a)
(b)
Transducer
Z1 Z2
(a) (b)
11.2.2.2 Scattering
11.2.2.3 Diffraction
When sound waves pass barriers, obstacles, or openings on their path, they
get diffracted. Diffracion involves a change in direction of the sound wave. In-
creasing wavelength λ yields an increased amount of diffraction (sharpness of
bending), and vice versa. If λ is smaller than the size of the barrier, obstacle,
or opening, the occurring diffraction becomes negligible.
11.2.2.4 Refraction
v1 sin α1
= , (11.5)
v2 sin α2
where α1 and α2 denote the angle of refraction in two different media, also
applies to sound waves. However, since sound velocities (v1 , v2 ) in human soft
tissue differ only marginally (see Tab. 11.2), the little effects of refraction in
medical ultrasound are negligible and therefore not considered further in this
chapter.
11.2.3 Attenuation
Voltage
al
teri
ma
c t ric
l e
zoe Electrode
Pie
11.3.1 Transducers
axial
lateral
Pulse sufficient
Transducer d d
Pulse too long
Axial resolution concerns structures lying behind each other w. r. t. the direc-
tion of the ultrasound waves. The better the axial resolution, the smaller the
distance between two structures can be such that they can be distinguished by
the transducer. Axial resolution is highly dependent on the ultrasound wave
frequency f . The illustration in Fig. 11.6 explains that dependency based on
a simple example, where an ultrasonic pulse generated by the transducer con-
sists in a single wave only (shortest possible pulse). The distance d between
the structures needs to be d ≥ λ/2 in order to be able distinguish between
them.
11.3 Image Acquisition for Diagnostics 245
Ultrasound offers a large variety of different imaging modes. The most com-
mon ones include A-mode, B-mode, and M-mode. A- and M-mode generate
one-dimensional (1-D) images (signals), whereas B-mode can be used to ac-
quire 2-D or even 3-D images (cf. Geek Box 11.1). Doppler (cf. Geek Box 11.2)
can be acquired in 1-D and 2-D, and with the most recent generations of
transducers also in 3-D.
A-mode is the simplest scanning mode. The height of the amplitude of the
reflected ultrasound is displayed over the sonic runtime in the sonic ray direc-
tion. Extractable measurements are: frequency, modulated frequency, height
of the impulse/amplitude, runtime, wave phase, phase shift, and attenuation.
246 11 Ultrasound
However, the major disadvantage is that only very localized information (one
single line through the body) is acquired.
The backward scattered ultrasound intensity along a single ray is called A-
mode. From a continuously running high-frequency generator, a wave packet
is cut out with a “gate” and is passed to the transducer. The returning echo is
given through a duplexer to a time-dependent amplifier (Time Gain Compen-
sation). Later arriving echoes, which are weaker because of the absorption,
are more amplified than the signals from the surface. Signals of high depth
(15 cm) are raised up to 120 dB. The signal height of an interface reflected sig-
nal is independent of the penetration depth from which the echo comes. The
signal-to-noise ratio becomes worse with increasing depth. The next sonic
impulse will be emitted if all echoes of the preliminary sonic impulse are de-
cayed. The repeat rate depends on the penetration depth and therewith on
the used frequency.
B-mode is the most common ultrasound mode. B-mode images are generated
by systematically combining a multitude of A-mode (1-D) scans into a single
2-D image, where the intensity of a pixel is defined by the amplitude of the
corresponding ultrasonic ray. In brief, in order to acquire 2-D images of the
inner body, the ultrasound device has to sample not only on a 1-D ray (as in
A-mode), but on a 2-D plane in 3-D space. Hence, various rays are sent in
different directions. To achieve this, two techniques are commonly used: the
mechanical and the electronic method.
Mechanical scanners The transducer librates in front of the patient, with-
out any external movement of the gaging head. Thus, a slice of the human
body is represented in the form of a circle segment. The intensity of the echo
is transformed into gray scales and is inserted into an image matrix (B-Mode).
An image consists of a fan of typically 100 lines.
Electronic scanners (linear/curved arrays) Here, many (60 to 100) and
very small (0.5 mm to 1 mm) transducers are used, which are arranged in a
row (“array”). A group of transducers is activated simultaneously. For scan-
ning, the whole group of elements is shifted. With a curved arrangement of
transducers, an image detail can be represented as a circle segment.
Electronic scanners (phased arrays) Every transducer element of an
array can be accessed for both sending and receiving with an individual ad-
justable delay.
11.4 Safety Aspects 247
In motion mode, ultrasonic pulses are emitted from the transducer in quick
succession without movement of the transducer. Either an A-mode or a B-
mode image is acquired each time. This allows for time-dependent measure-
ment of organ movement relative to the probe. Thus, the velocity of specific
organ structures can be obtained. This can be useful, for instance, when the
movement of the cardiac wall (myocardium) is to be analyzed (echocardio-
graphy).
Body surface
Blood vessel
In Doppler blood flow imaging, the source are the moving blood cells,
at which the waves scatter. The observer is the ultrasound transducer.
The smaller the angle between the direction of blood flow in the vessel
and the ultrasound wave direction, the better the Doppler effect can
be exploited.
11.4 Safety Aspects 249
Further Reading
[1] Olaf Dössel. Bildgebende Verfahren in der Medizin: Von der Technik zur
medizinischen Anwendung. Springer, 1999. isbn: 978-3-540-66014-9.
[2] G Goretzki. Medizinische Strahlenkunde: physikalisch-technische Grund-
lagen. Elsevier, Urban und Fischer, 2004. isbn: 978-3-437-47200-8.
[3] Paul Suetens. Fundamentals of Medical Imaging. Cambridge University
Press, 2009. isbn: 978-0-521-51915-1.
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Chapter 12
Optical Coherence
Tomography
Figure 12.1: Patient being imaged with a commercial OCT device. Image
courtesy of Carl Zeiss Meditec AG.
macula
choroid
Figure 12.2: OCT B-scan of the retina. Brighter pixels indicate tissue which
reflects more light. The upper portion of the figure shows the vitreous humor
which has very low reflectivity. The small pit in the center is the macula, the
center of vision with the highest resolution. The lowest horizontal bright band
corresponds to the retinal pigment epithelium and the cloud-like structure
below depicts the choroid, a blood vessel network supplying the retina with
nutrients and oxygen.
• Light exhibits properties of particles and waves of which only the latter
are relevant for this chapter. Light’s electromagnetic wave properties form
the basis for OCT.
• Coherence: two waves (or their sources) are described as being coherent
with each other, when they have matching wavelengths and the same shift
in phase.
• Interference: coherent waves superpose with each other (superposition
principle) and can cancel each other out (destructive interference) or re-
inforce each other (constructive interference).
• Bandwidth describes the width of the spectrum that a light source emits.
In contrast, a light source which is monochromatic, only emits light with
one wavelength. Such a light source has a bandwidth of 0.
12.1 Working Principle of OCT 253
mirror M1
mirror M2
light source
semi-transparent mirror
detector
Figure 12.3: Michelson interferometer. Half of the light from the light source
travels to mirrors M1 and M2 each, before arriving at the detector. Differences
in path lengths lead to interference.
2 ln 2 λ20
lc = (12.2)
π ∆λ
with λ0 being the central wavelength of the light source and ∆λ its
bandwidth. As can be seen, a higher bandwidth leads to a smaller
coherence length.
reflectivity
distance of reflector
intensity
coherence length
The upper half of the plot shows two reflectors with different reflectiv-
ities at different distances (as Dirac impulses). If the reference mirror
is moved to match the path length of the reflectors, the measured in-
tensity becomes maximal. Lower coherence lengths also increase res-
olution.
to the bandwidth (see Geek Box 12.1 for more details on coherence length).
Now, if the Michelson interferometer uses a low-coherence light source (a
light source which emits a spectrum), the coherence length can be used to
determine the distance of a reflector in one of the interferometer’s arms by
gradually moving the mirror in the other arm. Fig. 12.4 illustrates this, where
by moving mirror M1 to match the distances zM1 and zM2 , will generate an
intensity peak in the detector. The plot in Geek Box 12.1 shows how the
intensity peaks when zM1 and zM2 are matched.
semi-transparent mirror
detector
moving mirror
semi-transparent mirror
detector
Figure 12.5: Setup of a time-domain OCT system, one mirror has been
replaced with a sample. The other mirror can move to acquire an A-scan.
The mirror is located in the reference arm, the sample in the sample arm.
Fig. 12.5). The remaining mirror M1 forms part of the reference arm, whereas
the sample becomes part of the sample arm. The sample has to be translucent
enough to permit light to travel through it and to reflect back from different
layers. Thus, movement of the mirror over time results in a depth profile of
intensities of reflection at one position of the sample. This is called an A-scan.
Directing the beam along a line across the sample, while acquiring A-scans
at regular intervals, yields a two-dimensional image which is called a B-scan.
Creating a raster scan of B-scans yields a volume. Every pixel column in
Fig. 12.2 is an A-scan. The moving mirror is a disadvantage though, since it
limits the maximum sampling speed of the OCT device.
256 12 Optical Coherence Tomography
Figure 12.6: The OCT beam raster-scans the surface of the retina. Moving
the beam along a line results in a B-scan (2-D image). Every column in a B-
scan image is an A-scan. After each B-scan, the beam travels to the beginning
of the next one.
1mm
1mm
and platelets of 2–3 µm. This size is just about right to induce high speckle
noise in the image. In OCT angiography, this effect is exploited to create
a visualization of vessels without the need of contrast agent. The idea is to
scan the same area of the retina multiple times to generate a map of variance.
This map will have a high response in areas that contain vessels. Using the
structural OCT image (cf. Fig. 12.2), the retinal layers are then segmented
and used to create projections of each layer. Fig. 12.7 shows such projections
for the superficial and deep vascular plexi as well as the choroid. In Geek
Box 12.2, we detail measures for OCT angiography reconstruction. Note that
comparison of scans that were acquired in rapid sequence also allows for the
estimation of blood flow speed. This topic is scope of current OCT research.
12.5 Applications
OCT is predominantly used for imaging the eye. However, its application is
also quite common in other body regions. In the following, we summarize
shortly OCT’s fields of application.
• Ophthalmic Imaging: Retinal imaging is currently the major application
for OCT. Both the retina and anterior eye can be imaged for diagnos-
tic purposes completely non-invasively in 3-D. Furthermore, as described
above, the vessel structure can also be investigated in 3-D without the use
258 12 Optical Coherence Tomography
This concept
q is very similar to inter-frame variance, however, a local
scaling of In−1
2 + In2 is introduced for every amplitude difference.
Doing so, amplitude decorrelation is always scaled between 0 and 1
and therefore can be interpreted as an “inverse correlation” where 0 is
obtained for correlated observations and 1 for independent measure-
ments.
of contrast agent. As such, OCT has become the standard of care for the
diagnosis of eye diseases. Fig. 12.8 shows a volume of the anterior eye and
part of the retina.
• Cardiovascular Imaging: OCT can be used to diagnose cardiovascular dis-
eases. In order to do so, optical fibers are embedded into a catheter that is
inserted minimally invasively into the vessel system. Doing so, the vessel
wall can be imaged and areas of concern can be investigated. These are
typically calcifications and plaques that are attached to the vessel wall.
12.5 Applications 259
Figure 12.8: OCT volume showing the structure of the cornea, lens, and
iris of the anterior eye. The disc in the background is part of the retina which
is visible through the lens. These volumes are used in the visualization and
diagnosis of corneal pathologies and glaucoma.
calcified plaque
Figure 12.9: B-scan from a blood vessel. The small circle in the middle is
the OCT probe within the dark lumen. The bright ring around the lumen is
the vessel’s endothelium (inner surface). The gap on the right side is caused
by constructional properties of the probe. A calcified plaque is visible in the
top right quadrant of the endothelium.
Further Reading
[1] Bernhard Baumann et al. “Total retinal blood flow measurement with
ultrahigh speed swept source/Fourier domain OCT”. In: Biomedical
Optics Express 2.6 (2011), pp. 1539–1552. doi: 10 . 1364 / BOE . 2 .
001539.
[2] Mark E Brezinski. Optical coherence tomography: principles and appli-
cations. Academic press, 2006.
[3] Emily Cole et al. “The definition, rationale, and effects of threshold-
ing in OCT angiography”. In: Ophthalmology Retina 1/2017.5 (2017),
pp. 435–447. doi: 10.1016/j.oret.2017.01.019.
[4] Wolfgang Drexler and James G. Fujimoto. Optical coherence tomogra-
phy: technology and applications. Springer, 2008.
[5] David Huang et al. “Optical Coherence Tomography”. In: Science
254.5035 (Nov. 1991), pp. 1178–1181. doi: 10.1126/science.1957169.
[6] Martin Kraus et al. “Quantitative 3D-OCT motion correction with tilt
and illumination correction, robust similarity measure and regulariza-
tion”. In: Biomedical Optics Express 5.8 (2014), pp. 2591–2613.
[7] Jonathan J. Liu et al. “In vivo imaging of the rodent eye with swept
source/Fourier domain OCT”. In: Biomedical Optics Express 4.2 (2013),
pp. 351–363. doi: 10.1364/BOE.4.000351.
[8] Markus Mayer et al. “Retinal Nerve Fiber Layer Segmentation on FD-
OCT Scans of Normal Subjects and Glaucoma Patients”. In: Biomedical
Optics Express 1.5 (2010), pp. 1358–1383.
[9] Stefan Ploner et al. “A Joint Probabilistic Model for Speckle Variance,
Amplitude Decorrelation and Interframe Variance (IFV) Optical Co-
herence Tomography Angiography”. In: Bildverarbeitung für die Medi-
zin 2018. Ed. by Andreas Maier et al. Informatik aktuell. Erlangen,
2018, pp. 98–102. isbn: 3662565374. doi: 10.1007/978-3-662-56537-
7.
[10] Stefan Ploner et al. “Toward Quantitative Optical Coherence Tomogra-
phy Angiography: Visualizing Blood Flow Speeds in Ocular Pathology
Using Variable Interscan Time Analysis”. In: Retina 32 (2016). doi:
10.1097/IAE.0000000000001328.
[11] Carl Rebhun et al. “Analyzing relative blood flow speeds in choroidal
neovascularization using variable interscan time analysis OCT angiog-
raphy”. In: Ophthalmology Retina 2.4 (2018), pp. 306–319. doi: 10 .
1016/j.oret.2017.08.013.
[12] Franziska Schirrmacher et al. “QuaSI: Quantile Sparse Image Prior for
Spatio-Temporal Denoising of Retinal OCT Data”. In: Medical Image
Computing and Computer-Assisted Intervention - MICCAI 2017, Pro-
ceedings, Part II. Ed. by Maxime Descoteaux et al. Quebec City, QC,
Canada, 2017, pp. 83–91.
Open Access This chapter is licensed under the terms of the Creative Commons
Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/),
which permits use, sharing, adaptation, distribution and reproduction in any medium
or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were
made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended
use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder.
Acronyms
IR infrared. 120
PET positron emission tomography. 208, 212, 217–219, 222, 227, 228, 230–
233
PSF point spread function. 169, 170, 214–217, 220, 222, 224, 227
US ultrasound. 12
UV ultraviolet. 120
Author Index
Aubreville, Marc 69 Maier, Andreas 7, 13, 57, 69, 119, 147, 251
Mualla, Firas 69
Balda, Michael 147
Berger, Martin 119, 147 Neumann, Dominik 237
Bernecker, David 37
Bögel, Marco 147 Ploner, Stefan 251
Bopp, Johannes 191
Riess, Christian 191
Felsner, Lina 191
Fischer, Peter 13 Sanders, James 207
Sembritzki, Klaus 13
Haase, Sven 57
Hu, Shiyang 191 Taubmann, Oliver 147
Husvogt, Lennart 251
Wetzl, Jens 91
Kaeppler, Sebastian 191
Kollorz, Eva 237 Xia, Yan 147