Automated selection of hyper-parameters in diffuse

optical tomographic image reconstruction

A Thesis
Submitted For the Degree of
Master of Science (Engineering)
in the Faculty of Engineering

by

Jayaprakash

Supercomputer Education and Research Centre

Indian Institute of Science
BANGALORE – 560 012

FEBRUARY 2013
 1

c Jayaprakash
FEBRUARY 2013
All rights reserved
Acknowledgements

Working on this dissertation has made me grow intellectually and personally. The time
spent during this dissertation was enjoyable and it gives me immense pleasure to ac-
knowledge many people without whom this would not have been possible.
At the start I would like to thank my advisor Dr. Phaneendra K. Yalavarthy, for
encouraging me and helping me to view a problem from several dimensions, how to
connect various problems and then draw critical conclusions in the course of completion
of projects. I have acquired this kind of ability from my advisor about how a problem
can be solved. I have to acknowledge my advisor from whom I have learnt how to be
systematic in life, the importance of being dedicated to do a particular task, hard work
always results in success, and also the value of time in a person’s life.
The IISc campus presents a wonderful and informal atmosphere for intellectual growth
of a person. Me along with my friends Shiladitya Banerjee and Harikrishna were dis-
cussing about a simple idea on solar energy at Prakruti, a faculty had listened to our
discussion and informed us of a talk happening on the campus. I should acknowledge
this kind ambience of IISc campus, without which the completion of dissertation was not
possible. I thank the faculty of the Supercomputer Education and Research Centre and
Department of Electrical Engineering for teaching and exposing me to their respective
topics inside and outside of the class. I would also like to extend my thanks for the
chairman of Supercomputer Education and Research Centre.
I have to acknowledge Ravi Prasad K. J. for having fruitful discussions on solving
specific problems. I thank my friends Yamuna N. Swamy, Calvin B. Shaw and Kalyan
Ram A. without whom this dissertation work would not have been complete. I would

i
 ii

like to thank all my friends at Supercomputer Education and Research Centre and IISc
for helping me out in solving problems, and encouraging my ideas. I would also like to
thank Dr. Nick Todd, Dr. Ravi Namani, Dr. Suhesh Kumar Singh M., Deepak Karkala
and Akshay Ravikumar for the discussions that used to happen both inside and outside
the lab. I will always cherish their friendship. My interaction with them has taught me
many principles of research. I should appreciate my roommate Nikhil, who made my
stay joyful and created a nice environment while staying on campus. I would like to
thank my friends Karthik, Pawan, Praveen, Tarun, Venkit, and Akash for helping me to
relax during my course.
Last but not the least, I would like to thank my parents and my brother for their
continuous support and help in handling the pressure during completion and writing of
this dissertation work. I would like to express my sincere gratitude to all those who have
directly or indirectly helped in making this happen.
Abstract

Diffuse optical tomography is a promising imaging modality that provides functional in-
formation of the soft biological tissues, with prime imaging applications including breast
and brain tissue in-vivo. This modality uses near infrared light (600 nm - 900 nm) as
the probing media, giving an advantage of being non-ionizing imaging modality.

The image reconstruction problem in diffuse optical tomography is typically posed as

a least-squares problem that minimizes the difference between experimental and mod-
eled data with respect to optical properties. This problem is non-linear and ill-posed,
due to multiple scattering of the near infrared light in the biological tissues, leading
to infinitely many possible solutions. The traditional methods employ a regularization
term to constrain the solution space as well as stabilize the solution, with Tikhonov type
regularization being the most popular one. The choice of this regularization parameter,
also known as hyper parameter, dictates the reconstructed optical image quality and is
typically chosen empirically or based on prior experience.

In this thesis, a simple back projection type image reconstruction algorithm is taken
up, as they are known to provide computationally efficient solution compared to reg-
ularized solutions. In these algorithms, the hyper parameter becomes equivalent to
filter factor and choice of which is typically dependent on the sampling interval used
for acquiring data in each projection and the angle of projection. Determining these
parameters for diffuse optical tomography is not so straightforward and requires usage
of advanced computational models. In this thesis, a computationally efficient simplex

iii
 iv

method based optimization scheme for automatically finding this filter factor is proposed
and its performances is evaluated through numerical and experimental phantom data. As
back projection type algorithms are approximations to traditional methods, the absolute
quantitative accuracy of the reconstructed optical properties is poor. In scenarios, like
dynamic imaging, where the emphasis is on recovering relative difference in the optical
properties, these algorithms are effective in comparison to traditional methods, with an
added advantage being highly computationally efficient.

In the second part of this thesis, this hyper parameter choice for traditional Tikhonov
type regularization is attempted with the help of Least-Squares QR-decompisition (LSQR)
method. The established techniques that enable the automated choice of hyper pa-
rameters include Generalized Cross-Validation(GCV) and regularized Minimal Residual
Method (MRM), where both of them come with higher overhead of computation time,
making it prohibitive to be used in the real-time. The proposed LSQR algorithm uses
bidiagonalization of the system matrix to result in less computational cost. The proposed
LSQR-based algorithm for automated choice of hyper parameter is compared with MRM
methods and is proven to be computationally optimal technique through numerical and
experimental phantom cases.
Keywords

Medical imaging, computational methods in medical imaging, biomedical optical imag-

ing, diffuse optical tomography, dynamic diffuse optical imaging, inverse problems, regu-
larization parameter, hyper parameters, filter factor, LSQR based reconstruction, Back-
projection based reconstruction.

v
Publications based on this Thesis

1. Jaya Prakash and Phaneendra K. Yalavarthy, “A LSQR-type method provides a

computationally efficient automated optimal choice of regularization parameter in
diffuse optical tomography,” Medical Physics, 40(3), 2013 (in press).

vi
Contents

Acknowledgements i

Abstract iii

Keywords v

Publications based on this Thesis vi

1 Introduction to Diffuse Optical Tomographic Imaging 1

1.1 Introduction to Optical Imaging Applications . . . . . . . . . . . . . . . 1
1.2 Soft Tissue Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 X-ray Mammography . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Magnetic Resonance Imaging (MRI) . . . . . . . . . . . . . . . . 4
1.2.3 Ultrasound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Advantages and Limitations of Diffuse Optical Tomography . . . . . . . 6
1.4 Computational Challenges in Diffuse Optical Tomography . . . . . . . . 7
1.5 Contribution of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Organization of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Diffuse Optical Tomographic Imaging 11

2.1 Forward Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Analytic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.2 Monte-Carlo Methods . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 13

vii
CONTENTS viii

2.2 Inverse Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2.1 Levenberg-Marquardt (LM) minimization . . . . . . . . . . . . . . 16
2.2.2 Tikhonov minimization . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 Computational Complexity . . . . . . . . . . . . . . . . . . . . . . 19

3 Back-Projection Type Image Reconstruction 22

3.1 Back-projection Type Method . . . . . . . . . . . . . . . . . . . . . . . . 26
4 LSQR based Optimal Regularization Estimation 31
4.1 Regularized Minimal Residual Method (MRM) . . . . . . . . . . . . . . . 33
4.2 Generalized Cross-Validation (GCV) . . . . . . . . . . . . . . . . . . . . 34
4.3 LSQR-Type Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.4 Estimation of Optimal λ using a LSQR-Type Method . . . . . . . . . . . 38

5 Results and Discussion 42

5.1 Numerical and Experimental Evaluations of the proposed Backprojection-
type Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.1.1 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . 42
5.1.2 Gelatin Phantom Experiment . . . . . . . . . . . . . . . . . . . . 47
5.1.3 Dynamic Phantom Experiment . . . . . . . . . . . . . . . . . . . 48
5.2 Results based on Backprojection Algorithm . . . . . . . . . . . . . . . . 49
5.3 Discussion of results obtained using Backprojection-type Algorithm . . . 53
5.4 Numerical and Experimental Evaluation for the proposed LSQR algorithm 56
5.4.1 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . 56
5.4.2 Gelatin Phantom Experiment . . . . . . . . . . . . . . . . . . . . 57
5.5 Results for the proposed LSQR-type algorithm . . . . . . . . . . . . . . . 58
5.6 Discussions of results obtained using proposed LSQR-type algorithm . . . 61

6 Conclusion 63

A List of Packages/Programs 65
A.1 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
CONTENTS ix

A.2 Developed Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

Bibliography 69
List of Figures

1.1 Plot showing the variation of absorption coefficient of the absorption chro-
mophore’s (Oxyhaemoglobin, Deoxyhaemoglobin and water) as a function
of wavelength of light [27]. . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Comparison of resolution and intrinsic contrast characteristics of tradi-
tional imaging modalities with Near Infrared (NIR) diffuse optical imaging
Courtesy: B. W. Pogue. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.1 A cartoon showing the forward and inverse problems in diffuse optical
tomography. Courtesy Calvin Shaw B. . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Flowchart presenting the important steps in reconstruction of optical prop-
erties in diffuse optical tomography. . . . . . . . . . . . . . . . . . . . . . 20
2.3 Reconstruction of the target (shown in the top-left corner) using different
regularization parameter (λ) with Tikhonov minimization scheme and 1%
noise added to numerically generated data, with target representing the
expected distraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.1 Flowchart presenting the important steps in reconstruction of optical prop-

erties in back-projection type diffuse optical tomography. . . . . . . . . . 29

4.1 Comparison of the spatial resolution of MRM and GCV based optimal
regularization parameter estimation with “perfect” data and 1% noisy
data numerically generated using target [45]. . . . . . . . . . . . . . . . . 36

x
LIST OF FIGURES xi

4.2 Flowchart presenting the important steps in reconstruction of optical prop-

erties in diffuse optical tomography using LSQR-based optimal estimation
of regularization parameter. . . . . . . . . . . . . . . . . . . . . . . . . . 41

5.1 Comparison of reconstruction performance in 3D cylindrical mesh using

the standard and proposed backprojection-type optimization work. Here
two spherical anomalies are separated along the diagonal whose target
is shown on top. The reconstruction of the image using 1% noisy data
is shown for both proposed backprojection-type and standard technique.
Corresponding line profile is shown through the plot x=-y line the center
(z=0). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Comparison of reconstruction performance in 3D cylindrical mesh using
the standard and proposed backprojection-type optimization work. Here
two cylindrical anomalies are separated along the diagonal whose target
is shown on top. The reconstruction of the image using 1% noisy data
is shown for both proposed backprojection-type and standard technique.
Corresponding line profile is shown through the plot x=-y line the center
(z=0). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3 Comparison of reconstructions using the standard technique and pro-
posed backprojection-type optimized case for the 3D cylindrical mesh
case. We have shown the target µa and reconstruction with proposed
backprojection-type and standard techniques. The reconstructions are
performed for 1% and 3% noise cases. . . . . . . . . . . . . . . . . . . . . 45
5.4 Profile plot showing the target µa , standard and proposed backprojection-
type reconstruction for 1% and 3% noise case shown in Fig. 5.3. The
profile was drawn across line z=0 in this case. . . . . . . . . . . . . . . . 46
5.5 Comparison of reconstruction performance in breast mesh using the stan-
dard and proposed optimization work. Here we have a single anomaly
whose target along with reconstruction of the image using 1% noisy data
is shown for both proposed backprojection-type and standard technique. 47
LIST OF FIGURES xii

5.6 Comparison of proposed backprojection-type method with standard image

reconstruction for the stationary target with data have noise level of 1%.
The corresponding frame numbers are given on top of first row. . . . . . 48
5.7 Comparison of reconstructions using the standard technique and proposed
backprojection-type optimized case for the experimental data case. We
have shown the target µa and reconstruction with proposed backprojection-
type and standard techniques. Also shown is the profile plot across the
line z=0 in this case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.8 Reconstruction distribution of µa in case of experimental phantom data
using proposed backprojection-type and standard methods. The time cor-
responding to each frame is given in the top row. Maximum recovered
contrast corresponding to the results is presented in the graph. . . . . . . 50
5.9 Gelatin phantom used in the experimental studies [111] to verify the pro-
posed LSQR-type algorithm, the two-dimensional cross-sectional geome-
try is as shown in the top left of Fig. 5.12. . . . . . . . . . . . . . . . . . 57
5.10 The target µa distribution is shown on the top-left corner. The recon-
struction method along with the data noise level are given on top of each
reconstructed µa distribution. The one-dimensional cross-sectional plot
of target and reconstructed µa distributions along the line shown in the
target image are given in the bottom-right corner. . . . . . . . . . . . . . 58
5.11 Similar effort as Fig. 5.10 except the target µa distribution has two anoma-
lies and the noise level in the data is kept at 1% . . . . . . . . . . . . . . 59
5.12 Performance comparison of MRM and the proposed LSQR method for the
case of gelatin phantom data . . . . . . . . . . . . . . . . . . . . . . . . . 61
List of Tables

5.1 Comparison of total computational time (in seconds) for the three cases
considered in this thesis. The last row gives the proposed backprojection-
type method timing results. First two rows corresponding to Eqs. 2.22
and 2.23 gives the timing for standard methods. The total number of
iterations taken to converge are given in the parenthesis. . . . . . . . . . 52
5.2 Comparison of regularization parameters obtained for the results pre-
sented in this work. The λ values are decreasing with increasing number
of iterations for all cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Comparison of total computational time (in seconds) including the over-
head time, for the three cases considered in this work. The last column
indicates the proposed method timing results. The second column indi-
cates the image reconstruction using regularized MRM method. The total
number of iterations taken to converge are given in the parenthesis. The
corresponding regularization parameter values are given in Table-5.2. . . 60

xiii
Chapter 1

Introduction to Diffuse Optical

Tomographic Imaging

Diffuse optical tomography helps in non-invasively acquiring the functional information

of the tissue under investigation. This imaging modality uses near infrared (NIR) light as
the probing media and typically the measurements are acquired at the tissue boundary
non-invasively and in-vivo. Since, the propagating medium is turbid in nature, these
NIR photons does not travel ballistically through this medium. Rather than the light
experiencing dominance of absorption property, they would experience many scattering
events. This modality provides a novel approach of differentiation of cancer cells with a
capacity of measuring the oxygen concentration in these tissues. Diffuse optical tomog-
raphy also provides high specificity due to high intrinsic contrast provided by tissue at
NIR wavelengths compared to other imaging modalities. Since this modality provides
high sensitivity and is non-invasive, it can help in decreasing the number of unnecessary
breast biopsies with effectively screening breast cancer.

1.1 Introduction to Optical Imaging Applications

Breast cancer is affecting approximately one in nine women and is found to be approxi-
mately 29% fatal [1]. Usage of light for detecting the tumor was first proposed by Cutler

1
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING2

in 1931 [2]. Cutler tried to distinguish between tumors and cysts in breast. Later with
advancements in 1980’s it was shown that there is a window that exists in the near in-
frared range of wavelengths(600 nm to 1000 nm) which can help in imaging large thick
tissues and also provide a mechanism for estimating tissue absorption in-vivo [3–13].
The major drawback observed by usage of this light was that the resulted images were
found to be blurry due to high degree of breast scattering [14]. The recent advancements
in modeling light transport and technological advances in tissue optics has made diffuse
optical imaging to be feasible for estimation of tumor properties with high rates of speci-
ficity and resolution better than the earlier reported works.

Figure 1.1: Plot showing the variation of absorption coefficient of the absorption chro-
mophore’s (Oxyhaemoglobin, Deoxyhaemoglobin and water) as a function of wavelength
of light [27].

The most important advancement that lead to the widespread acceptance of diffuse
optical imaging to characterize breast tumor, is the modeling of light transport over long
distances in thick tissues using a diffusive process [15–17]. This model helps for separa-
tion of tissue scattering from tissue absorption quantitatively [18–22]. This enables the
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING3

estimation of the major absorption chromophores like haemoglobin, oxyhaemoglobin and

water by using different wavelengths (Fig. 1.1). The scattering chromophore like scatter-
ing amplitude and scattering power could also be estimated by adding more wavelengths
to the data acquisition set-up. Diffuse optical imaging could be used in many applica-
tions which requires the estimation of the oxygen concentrations in the biological tissues.

The major applications of of diffuse optical imaging is towards in vivo soft tissue
imaging like in brain haemodynamics and metabolism [23, 24], in detecting and char-
acterizing breast tumors [25], and in tumor biology [26]. Near Infrared Diffuse Optical
tomography has a potential to monitor fetal cerebral oxygenation [25] and haemody-
namics non-invasively in utero [28–30], as the radiation used is non-ionizing enabling
prolonged and continuous monitoring. This is required as the current non-invasive tech-
nology namely antepartum screening is found to have high false positive which leads
to Cesarean sections and premature delivery [31–33]. Diffuse optical tomography could
also be used similar to pulse oximetry that is used for monitoring a patient’s blood
haemoglobin as absorption chromophores can be estimated [29].

1.2 Soft Tissue Imaging

There are many existing soft tissue imaging modalities at the clinical level namely X-ray
mammography, magnetic resonance imaging (MRI), and ultrasound. This subsection
discusses the strengths and limitations of each modality, motivating the development of
alternative imaging modalities, including diffuse optical imaging.

1.2.1 X-ray Mammography

The most traditional method of detecting the tumors in breast is by using low energy X-
ray photons. X-ray mammography passes ionizing x-ray radiations through compressed
breast and these are then detected by a radiographic film/digital detectors to create
images. These images are then analyzed by radiologists to conclude whether there are
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING4

any abnormalities like tumors present. This kind of breast cancer screening performs
poorly with dense breast tissues as it results in large false-negatives [34–36]. Also the
radiation exposure due to usage of low-energy X-ray source is very high, leading to
high dosage, hence repeated screening in the same year is not recommended. Since this
mechanism has low specificity [37–39], the healthy patient might be recommended to
undergo unnecessary biopsies (which in turn will become a invasive surgical procedure).

1.2.2 Magnetic Resonance Imaging (MRI)

Magnetic resonance imaging (MRI) can also be used for soft tissue imaging, here a
strong magnetic field is applied for imaging. The strength of the magnetic field is of the
order of 0.5 to 3 Tesla. The signal in MRI is due to absorption and emission of radio-
frequency energy by the atoms (primarily water) which posses an unpaired spin. This
modality is very helpful in soft tissue imaging, as the signal strength is proportional to the
concentration of the hydrogen nuclei. One of the drawbacks of MRI is the low intrinsic
contrast for distinguishing the abnormal tissue with normal tissue as compared to optical
imaging (Fig. 1.2). Contrast-enhanced MRI can be used for studying the functionality
of blood flow more effectively. MRI is known to have very high sensitivity [40] but
low specificity (false positive) [36,40]. The major limitations of the MRI is its prolonged
procedure time, making it more vulnerable to motion artifacts [41]. The cost effectiveness
of MRI is also low, making it one of the expensive procedures in cancer screening. MRI
gives a very good structural information of the imaging tissue, it can be combined with
the functional information provided by diffuse optical tomography, which is an active
area of research.

1.2.3 Ultrasound

Ultrasound imaging is a technique that uses same principles as radar imaging, except
the reflected signal is an indicative of mismatch in the acoustic impedance between
different tissue types. The mechanical wave of the frequency 1-15 MHz is typically used
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING5

for probing the tissue. This frequency range is not in the audible range of humans.
This kind of imaging has high sensitivity and specificity depending on the technician
proficiency [40–42]. This imaging can be used routinely for prolonged examination of
tissue as the radiation is non-ionizing in nature. The acquired images from ultrasound
should be read and interpreted by an experienced sonographer. The major limitation
of this modality is its inability in effective visualization of deep lesions as the detected
signal strength is inversely proportional to square of depth. Also these ultrasound signal
does not pass through bone, which makes it difficult to image inner body organs.

Figure 1.2: Comparison of resolution and intrinsic contrast characteristics of traditional

imaging modalities with Near Infrared (NIR) diffuse optical imaging Courtesy: B. W. Pogue.

As each of these traditional imaging modalities has limitations in terms of imaging

younger patients, visualizing deep lesions, and prolonged procedural time, it is highly
desirable to develop an alternate imaging modalities. Moreover, it will be advantageous
if the imaging mechanism is non-invasive, non-ionizing, and provides higher intrinsic
contrast. The Near Infrared (NIR) imaging (also known as diffuse optical imaging)
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING6

has this ability and can overcome the limitations of the traditional soft-tissue imaging
modalities. The intrinsic contrast that is provided by the NIR imaging is atleast 10-fold
higher compared to their counterparts (depicted in Fig. 1.2). The next section describes
the diffuse optical tomographic imaging limitation in the light of the discussion in this
section.

1.3 Advantages and Limitations of Diffuse Optical

Tomography
As described earlier, diffuse optical tomographic imaging uses NIR light and is known
to be non-ionizing and hence provides the flexibility for repeated usage for examining
the tissue under investigation. The intrinsic contrast of diffuse optical imaging is very
high compared to the traditional methods which can be seen in Fig. 1.2. The contrast
of the traditional methods could be improved by using external contrast agent, making
the imaging procedure invasive in nature. The contrast information in diffuse optical to-
mography is based on the optical properties (absorption and scattering) of haemoglobin,
oxyhaemoglobin, and water. Another major advantage of diffuse optical tomography is
that this modality helps in visualizing deep lesions effectively when compared to ultra-
sound as infrared rays can penetrate deeper tissues.

The major limitation of diffuse optical tomography is its poor resolution characteris-
tics compared to traditional imaging modalities (Fig. 1.2). The modern diffuse optical
imaging systems are multi-model in nature, where the limitation of spatial resolution is
overcome by combining diffuse optical imaging system with traditional imaging modal-
ities. Examples being MRI combined with NIR imaging, X-ray combined with diffuse
optical imaging. These multi-modal imaging systems inherits the high intrinsic contrast
provided by NIR imaging combining with high spatial resolution provided by traditional
imaging modalities, which have been shown to be more effective compared to traditional
diffuse optical tomographic systems.
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING7

1.4 Computational Challenges in Diffuse Optical To-

mography
As the NIR light propagation in tissue is dominated by scattering, the computational
model that is used for describing this process tends to be computationally complex in
nature due to the usage advanced numerical model, as tissue could be irregular in shape.
Obtaining images in real-time is one of the main bottlenecks in deploying diffuse optical
imaging in real-time as it requires repeated usage of these advanced numerical models.
The iterative technique used in getting diffuse optical images relies on computing the
sensitivity matrix (also known as Jacobian) and modeled data at every iteration [43].
Depending on the model used the computation of the Jacobian can go up to several
hours, especially in three-dimensional cases. The most preferred numerical model that
can effectively handle the irregular shaped tissues is the finite element method. The
computation of Jacobian matrix is performed by discretizing the imaging domain into
linear triangular elements (for two dimensions) or linear tetrahedral elements (for three
dimensions) connected using NN finite element vertex nodes. The computationally ex-
pensive step in the diffuse optical tomographic imaging is inversion of Hessian or its
equivalent, which is O(NN3 ), for real-valued Jacobian.

Typical reconstruction time for three-dimensional imaging could be as high as 3

hours per iteration, making the computational complexity associated with image recon-
struction as one of the main bottlenecks for diffuse optical imaging to be deployed in
real-time. Moreover, the image reconstruction scheme requires regularization parameter
(also known as hyper parameter) to stabilize the inverse solutions. This regularization
parameter dictates the reconstructed image quality and is known to bias the results,
making the choice of it as most important parameter in the reconstruction scheme.
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING8

1.5 Contribution of Thesis

The traditional methods employ a hyper-parameter to constrain the solution space as
well as stabilize the solution, with Tikhonov type regularization being the most popular
one. The choice of this hyper parameter, dictates the reconstructed optical image quality
and is typically chosen empirically or based on prior experience.

In this thesis, algorithms have been developed that can enable the automated choice
of these hyper-parameter, which does not require any prior information. Also the pro-
posed schemes have been shown to have an added advantage of solving these inverse
problem efficiently compared to the existing methods. In the initial part of the thesis,
a back-projection type algorithm is explored and the filter factor (hyper-parameter in
this scheme) present in this back-projection type algorithm was automatically selected
based on an optimization scheme. Also, it was shown that even-though back-projection
algorithms does not provide absolute quantitation, the relative quantitation (which is
important in case of dynamic diffuse optical tomography) of this algorithm is compara-
ble with the traditional methods.

In the later part of this thesis, in Tikhonov type regularization scheme automated
selection of hyper-parameter using an LSQR-type algorithm was explored. It was previ-
ously shown that Minimal Residual Method (MRM) based estimation of optimal regular-
ization parameter gives the better spatial resolution than Generalized Cross-Validation
(GCV) based or L-curve based methods in diffuse optical tomography [44,45]. Hence the
proposed algorithm is evaluated with the existing MRM-based automated selection of
regularization parameter using numerical simulation (for spatial resolution) and experi-
mental phantom case. It was observed that the proposed algorithms have a significant
computational advantage than the existing standard schemes.
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING9

1.6 Organization of Thesis

The rest of thesis is organized as follows.
The Chapter-2 presents the forward and inverse problems of diffuse optical tomog-
raphy and existing computational methods to solve these problems. This chapter also de-
scribes the reconstruction technique based on Levenberg-Marquardt minimization scheme
and its variants, it also brings out the importance of selection of the regularization pa-
rameter (λ) and how this λ affects the quality of reconstructed images. It also gives a
brief overview of the computational complexities associated with the traditional diffuse
optical tomographic image reconstruction methods.

The Chapter-3 presents the proposed computationally efficient back-projection type

image reconstruction algorithm. This chapter also presents the proposed method for
automated selection of filter factor (hyper parameter) using a simplex method based
optimization scheme.

The Chapter-4 presents a brief overview of the Lanczos bidiagonalization algorithm

(decomposition of matrix into its bidiagonal form). This chapter also explains how this
algorithm was adopted in the diffuse optical tomography framework for automatically
selecting an optimal regularization parameter (hyper parameter). Existing methods for
automated selection of the regularization parameter namely regularized Minimal Resid-
ual Method (MRM) and Generalized Cross-Validation (GCV) were also explained along
with their performance in the context of diffuse optical tomography based on the litera-
ture.

The Chapter-5 presents the results using proposed methods discussed in Chapter 3
and Chapter 4 and is compared with the traditional methods discussed through a sys-
tematic series of simulations and experimental-phantom data. It also brings out the
computational advantage of the proposed back-projection type technique using three-
dimensional imaging domain by comparing the time taken for reconstruction. Even
CHAPTER 1. INTRODUCTION TO DIFFUSE OPTICAL TOMOGRAPHIC IMAGING10

though back-projection type algorithm gives poor absolute quantitative accuracy, it was
shown that back-projection type algorithms can be used for dynamic diffuse optical to-
mography case where the emphasis is on the relative difference in optical properties. This
chapter also shows that the proposed LSQR-type algorithm’s performance is similar to
that of the traditional MRM based methods, with having an added advantage of being
computationally efficient.

Finally, the chapter-6 gives the conclusions drawn based on the results presented in
this thesis and outlines the future work.
Chapter 2

Diffuse Optical Tomographic

Imaging

Diffuse optical tomographic imaging involves solving the forward and inverse problems
which is schematically shown in the Fig. 2.1. Forward problem is a process of generating
the boundary data (y) for a given absorption and reduced scattering coefficient of the
tissue (µa and µ0s ) using a model for the light propagation in the tissue. Inverse problem
is a procedure where the optical properties (µa and µ0s ) is reconstructed with a given
boundary data (y). Solving of this inverse problem is computationally expensive due to
the under-determined nature [43] (as the number of optical properties to be estimated
>> the number of boundary measurements). Also this inverse problem tends to be ill-
posed as scattering is dominated over the absorption [43]. This inverse problem is solved
iteratively as it is a non-linear problem. [43].

2.1 Forward Problem

The continuous-wave diffuse optical tomographic image reconstruction is performed by
matching the modeled data G(µa ) (where µa represents the optical absorption coeffi-
cient and G is the non-linear model) with the experimental data (y) in the least-square
sense. In here, G represents the numerical-model used to solve the steady-state diffusion

11
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 12

equation (DE), where the DE is given by [43, 46, 47]

Forward Problem Inverse Problem

Fibers y?? Fibers

y
μa, μ '
s
?? μa, μ's??
??

Figure 2.1: A cartoon showing the forward and inverse problems in diffuse optical to-
mography. Courtesy Calvin Shaw B.

−∇.D(r)∇Φ(r) + µa (r)Φ(r) = Qo (r) (2.1)

where Φ(r) represents the photon density (real values) at position r. The isotropic light
source is given by Qo (r) at r. The diffusion coefficient D(r), defined as

1
D(r) = (2.2)
3[µa (r) + µ0s (r)]

with µ0s (r) and µa (r) representing the reduced scattering coefficient and absorption co-
efficient respectively. Solving the DOT forward problem is classified into three types.

2.1.1 Analytic Methods

Analytical solutions are traditionally solved with the help of Green’s function, hence
are also called as Green’s function approach. A Green’s-function is the solution when
the source is considered as a delta-function. Solutions for the extended sources can be
derived using convolution procedure of these Green’s function with the current source
distribution, hence is easy to compute. Major drawback of these methods are that the
solutions can be derived/exist only for simple homogenous media [48], and solutions for
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 13

irregular geometries are not straightforward to derive. Analytical solutions for the diffuse
equation exist for most homogenous regular geometries [49]. The main advantage of
analytical solutions lies in its simplicity in both computing and behavior of the solutions,
making them as one of the favorite techniques for estimating the bulk optical properties.
Moreover, these solutions are also typically deployed in calibrating the experimental
data. Cases like layered slab media, the analytical solution are derived for time-domain
DE [49].

2.1.2 Monte-Carlo Methods

Monte-Carlo methods are probabilistic in nature. These are also stochastic methods for
modeling of transport of light photons through tissue, such as Markov-chain models [50]
or random-walk [51]. In Monte-Carlo methods, a large number of photon packets are
launched in the site under investigation, then these photons are traced by probabilistic
means of absorption and scattering process until they exit or are completely absorbed.
Statistically meaningful information on amplitude and phase can be extracted only by
usage of large number of photons (typically 6-10 millions). Major usage of these methods
are in solving forward model, especially in small-imaging domains. There are few investi-
gations in the literature [52,53] for regular imaging geometries. Major limitation of these
methods are that these methods are computationally costlier compared to conventional
schemes. As these Monte-Carlo methods are know to be equivalent of solving radiative
transport equation (RE), it could be used for any imaging domain.

2.1.3 Numerical Methods

The various numerical schemes used for solving the forward problem in diffuse optical to-
mography include Finite Volume Method (FVM) [54], Finite Difference Method (FDM)
[44], Finite Element Method (FEM) [55], and Boundary Element Method (BEM) [56].
Among all, FEM is the most preferred choice for solving the forward model [55] as it en-
ables easy handling of irregular geometries. In FEM based schemes, the imaging domain
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 14

is discretized (divided into discrete finite elements). The diffusion equation is solved on
this discretized domain via usage of linear finite elements.

In this thesis, a finite-element framework is used to numerically solve diffusion equa-

tion (forward model) and to generate modeled boundary data (G(µa )) for a given distri-
bution of absorption coefficients (µa ) [46,57,58]. Refractive-index mismatch at the tissue
boundary is taken care by Robin (type-III) boundary condition [58]. If F (non-linear
function) is the given forward model solved by FEM which gives the fluence (Φ) at every
point,
Φ = F (µa ) (2.3)

The modeled data (G(µa )) is obtained by sampling of Φ(r) at the measurements po-
sitions, given as G(µa ) = S{Φ(r)}, where S represents a sampling matrix (containing
source/detector positions) [59]. Under the Rytov approximation, the experimental data
becomes y = ln(A), where ln(A) is the natural logarithm of amplitude (A). Hence the
forward model can be summarized as

y = G(µa ) (2.4)

Then the inverse problem should be given as

µa = G−1 (y) (2.5)

But this is not that straightforward, as the system is non-linear and the problem is
typically under-determined in nature. Hence inversion is performed by linearizing the
problem, Taylor expansion of Eq. 2.4 around a close initial estimate (µa0 ) gives [60]

G(µa ) = G(µa0 ) + J4µa + (4µa )T H4µa + ... (2.6)

∂G(µa )
Here J = ∂µa
is the jacobian matrix having a dimension of NMxNN with NM being
∂ 2 G(µa )
the number of boundary measurements, H = ∂ 2 µa
is the hessian and µa is the actual
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 15

absorption coefficient. Linearizing the above equation and ignoring the higher-order
terms, gives us
4y = J4µa (2.7)

where 4µa = (µa − µa0 ) and 4y = y − G(µa0 ). Inverting the above equation is
not straightforward and requires additional help of regularization parameter as J is
ill-conditioned.

2.2 Inverse Problem

The image reconstruction (Inverse problem) is a process of estimating the optical prop-
erties (µa ) from the boundary measurements as shown in Fig. 2.1. Hence this inverse
problem requires solving the forward problem and calculation of Jacobian (J) at every
iteration [46]. The inverse problem in diffuse optical imaging is solved by minimizing
the objective function (Ω) in a least-square sense. There are many schemes available
for solving this kind of least-square problem. Gradient optimization based schemes do
not require explicit hessian matrix inversion, hence are found to be computationally
efficient [61, 62]. These methods require optimization scheme for choosing a step-size,
hence are not straightforward as direct matrix inversion. Total variation methods were
used for obtaining edge-preserved reconstruction results [63]. Solving this least square
could also be seen as a Bayesian approach for obtaining maximum a posterior (MAP)
estimates [64, 65]. The traditional and most popular method of performing image recon-
struction is based on Full-Newton approaches which involves direct inversion [43], the
same has been used as reference in this entire thesis. But since the inverse problem is
ill-posed, a regularization term is added for stabilizing the solution [43,57]. The following
section will discuss these minimization technique and the mathematical framework for
these techniques.
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 16

2.2.1 Levenberg-Marquardt (LM) minimization

The Levenberg-Marquardt optimization scheme is the most popular and widely used in
diffuse optical image reconstruction. In this the objective function is defined as [66, 67],

Ω = ky − G(µa )k2 (2.8)

where y is the obtained experimental data and G(µa ) is the modeled data. This equation
is minimized by taking the first order condition and equating it to zero, which is given
by
∂Ω
= JT δ = 0 (2.9)
∂µa
where δ = y − G(µa ) is the data-model misfit, J is the Jacobian (of dimension NMxNN,
with NM being the number of measurements and NN being the number of finite element
nodes in the mesh) and T represents the matrix transpose operator. The solution is
typically obtained using linearization of the original problem and taking many such
linear steps. In this scheme, we find the iterative update equation by expanding the
Taylor series on G(µia ) around µi−1
a given by

1  i T 0  i 
G(µia ) = G(µi−1 i
a ) + J4µa + 4µa J 4µa + . . . (2.10)
2

where 4µia = µia − µi−1

a . Linearizing the above equation and rewriting using δ gives

δ i = y − G(µia ) = y − G(µai−1 ) − J4µia = δ i−1 − J4µia (2.11)

Substituting the above equation in Eq. 2.9 by considering the ith iteration gives,

J T δ i = J T (δ i−1 − J4µia ) = 0 (2.12)

Rewriting the above equation leads to

J T J 4µia = J T δ i−1
 
(2.13)
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 17

Since the symmetric matrix in the above equation (JT J) is ill-conditioned, a diagonal
term is added for stabilizing the solutions. Hence the update equation becomes:

−1 T i−1

4µia = J T J + λI

J δ (2.14)

where 4µia is the update for the parameter at the ith step. In practice [43, 47], the
regularization parameter (λ) decreases with iterations (always > 0). There are many
limitations of using this scheme, which are discussed thoroughly in Ref. [60, 67].

2.2.2 Tikhonov minimization

The other popular method for obtaining a solution for ill-posed problems is through
Tikhonov type minimization scheme. In this scheme, the minimization function is given
by [68, 69]
Ω = ky − G(µa )k2 + λkL(µa − µa0 )k2 (2.15)

where λ is the regularization parameter, L is a dimensionless regularization matrix, and

µa0 is the initial guess. This regularization parameter depends on the variance in data
and variation in the optical properties. This can be shown as follows

ky − G(µa )k2 kL(µa − µa0 )k2

 
Ω= + (2.16)
(σδ )2 (σµa −µa0 )2

where σδ2 is the variance in the data-model misfit and σµ2 a −µa0 is the variance in the
optical properties. Now variance in the data-model misfit (δ = y − G(µa )) can be
written as [59, 60]
σδ2 = σy2 + σG(µ
2
a)
(2.17)

2
with σG(µ a)
= 0 implying that model is perfect, leads to σδ2 = σy2 , where σy2 is variance in
the estimated data. Now comparing Eq. 2.15 and Eq. 2.16 (multiplying this equation
by σy2 ) we get
σy2
λ= (2.18)
σµ2 a −µa0
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 18

which shows that the regularization parameter is the ratio of variance of data-model
misfit to the variance in the optical parameters. The minimization is achieved by taking
first derivate of Eq. 2.15 and equating it to zero. As σµ2 a −µa0 is unknown, this makes the
determination of λ largely heuristic. Taking the first order condition of Eq. 2.15 we get,

∂Ω
= J T δ − λLT L(µa − µa0 ) = 0 (2.19)
∂µa

and the first order condition in an iterative form (explained extensively in [60]) result in
the update equation of the form

J T J + λLT L 4µia = J T δi−1 − λLT L(µi−1

 
a − µa0 ) (2.20)

Note that this L is supposed to be positive definite and is symmetric [69]. In the absence
of spatial priors (L is of the form of identity matrix), hence the objective function turns
out to be
Ω = ky − G(µa )k2 + λkµa − µa0 k2 (2.21)

and the update equation turns out to be

−1 T
4µa = J T J + λIn

J (y − G(µa )) (2.22)

where 4µa is the update [57]. The identity matrix of dimension NNxNN is represented
by In .

The choice of this λ is known to effect quality of reconstructed images. The choice of
the regularization parameter has been reviewed thoroughly in literature [57,59,65,70–74].
Typically λ is empirically chosen based on the prior experience of the user [57,59,73]. In
our work we considered the objective function and the update equation without spatial
priors (Eq. 2.21 and Eq. 2.22). This is an iterative process, where at each iteration J is
recomputed along with G(µa ) based on the current optical property values. The same is
shown as a flow chart in Fig 2.2.
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 19

2.2.3 Computational Complexity

As it is evident from Eq. 2.22, obtaining 4µa requires O(NN3 ) operations excluding the
computation of J and G(µa ). Typically for under-determined problems (where NM <<
NN), an alternate update equation, which is known to be computationally efficient [75],
is used (equivalent to Eq. 2.22) and is given by [75]

−1
4µa = J T JJ T + λIm

(y − G(µa )) (2.23)

In here, Im represents the identity matrix of dimension NMxNM. The dominant order
of computation in this case is O(NN*NM2 ) (computation of JJT ). In either of Eqs.
2.22 or 2.23, the choice of regularization parameter is critical and is known to bias the
reconstruction results [57, 59].

The importance of selection of this regularization is shown in Fig 2.3. Here it can be
seen that the reconstruction results becomes more blurry as the value of regularization
parameter (λ) increases and the structural details are lost. But if the value of λ decreases
the noise component dominates and the reconstruction results will have some noisy
artifacts, but the reconstruction will be able to retain the structure. As the value of
λ increases the convergence becomes slower (will take more number of iterations to
converge), especially in LM minimization scheme.
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 20

Start  

  Ini)al  Guess,  μa0  

μ  =  μa0  

Solve  Forward  Model  G(μa)  

and  Calculate  Jacobian  J  

Experimental  Data  (y)  

μa  =  μa  +  Δμa  
yes  
                           <  2   Stop  

no  

Figure 2.2: Flowchart presenting the important steps in reconstruction of optical prop-
erties in diffuse optical tomography.
CHAPTER 2. DIFFUSE OPTICAL TOMOGRAPHIC IMAGING 21

Figure 2.3: Reconstruction of the target (shown in the top-left corner) using different
regularization parameter (λ) with Tikhonov minimization scheme and 1% noise added
to numerically generated data, with target representing the expected distraction.
Chapter 3

Back-Projection Type Image

Reconstruction

The traditional image reconstruction algorithms in diffusion optical tomography uses

Gauss-Newton based minimization method and requires a computationally expensive
step of inverting hessian with additional help of regularization parameter. This regular-
ization plays an important role in terms of reconstructed image characteristics as shown
in Fig 2.3 and is typically chosen empirically based on the prior experience. Choice of
this regularization parameter depends on the resolution characteristics that are needed.
There is always trade-off between geometric/spatial resolution and contrast resolution.
As diffuse optical imaging is known to have the advantage of providing high contrast res-
olution (discriminating normal from abnormal tissue), optimal choice of regularization
parameter should be based on expected contrast resolution.

In this thesis it is shown that one can eliminate the empirical choice of regularization
parameter and computationally expensive step of inversion of the regularized hessian
matrix by using an approximation to this inversion by a scalar (filter-factor or hyper-
parameter). Note that the filter-factor/hyper-parameter here refers to a scalar value and
in traditional imaging modalities like CT, they refer to a vector. An approximation of
this kind leads to a back-projection type algorithm, which is computationally efficient,

22
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 23

and will be shown that such method gives reasonably close results compared to tra-
ditional methods (Full-Newton reconstruction). The motivation for such an algorithm
arises from the geophysical inversion theory, where this type of approximation was shown
to be effective in solving the reconstruction problem [76].

In the geophysical studies this approach was used and approximate solutions were
found to be reasonably accurate and computationally efficient. The forward model in
those studies were given to be [76]
δu = Gδc (3.1)

where δu is the wave field, G is the corresponding linear operator, and δc is the model
vector representing the change in velocity (similar to Eq. 2.7). The forward model in
those studies (Eq. 3.1 is similar to the forward model in diffuse optical tomography).
In the geophysical inverse model, they would be searching for c = co + δc [76]. The
constraints for the search would be that the model (c) is close to initial guess (co ) and Cc
is defined as the covariance in that model and the other constraint is that the observed
residue δu must be close to δuo with Cu defined as the covariance in the residue. In
this thesis we consider all these covariance models to be equivalent to an identity matrix
(I). Then in this inverse model they have tried to minimize a quadratic function S in a
generalized least-square (GLS) sum sense given by [76]

1
(δuo − Gδc)∗ Cu−1 (δuo − Gδc) + δc∗ Cc−1 δc

S= (3.2)
2

Since the Eq. 3.2 is quadratic, the solution always exist and will be unique, as the
function defined in Eq. 3.2 is a convex function. Differentiating this with respect to δc
(γ = −grad(S)) we get [76],

γ = G∗ Cu−1 (δuo − Gδc) − Cc−1 δc (3.3)

CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 24

The derivative should be 0 at the minima, hence from first order condition we get [76],

(G∗ Cu−1 G + Cc−1 )δc = G∗ Cu−1 δuo (3.4)

If we consider the covariance matrices to be identity, this update is similar to Eq. 2.22
in diffuse optical tomography. Now in Ref. [76] an alternate kind of update equation was
proposed where the update equation (Eq. 3.4) was discretized and solved as a matrix
systems. If this was not discretized it could considered as continuous system and a
gradient based method could be applied to find a minima for S. In that work conjugate
gradient method was used and an alternate algorithm was proposed given as [76]

δck+1 = δck + αk 4ck (3.5)

4ck = Cc G∗ Cu−1 (δuo − Gδck ) − δck (3.6)

where αk (filter factor) is given as [76]

4c∗k Cc−1 4ck

αk = (3.7)
(G4ck )∗ Cu−1 (G4ck ) + 4c∗k Cc−1 4ck

Hence an alternate update equation by approximation of the inversion of hessian to a

scalar is given as [76]
δc = αk Cc G∗ Cu−1 δu (3.8)

this factor of αk was selected empirically in that work [76], this kind of framework was
used in diffuse optically tomography. In this thesis, we have proposed a methodology
to select this filter factor automatically. Since there is always a tradeoff between con-
trast resolution and geometric resolution, where both of them depend on the filter factor
choice. The automated selection of filter factor will ensure that reconstructed image is
more close to the target (expected) image via choosing a proper trade-off between both
resolutions. The covariance matrices (Cc and Cu ) are considered to be identity matrix
in diffuse optical tomography. Also in diffuse optical imaging, earlier works have shown
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 25

that back-projection type approach can be used as an alternative to Gauss-Newton based

image reconstruction method with a reasonable accuracy for solving the inverse prob-
lem [77]. The filter factor used in X-ray CT was depending on the sampling interval
used for acquiring data in each projection and the angle of projection [78, 79]. Also in
later years error analysis for Fourier filter methods were explored in great detail for X-ray
CT [80–82], but these cannot be used here as the NIR light propagation is not in straight
lines.

Choice of the filter factor, similar to regularization parameter, dictates the recon-
structed image quality [77–79], the heuristic choice of this filter factor results in a recon-
structed image, which may not have optimal resolution characteristics. A poor choice
of filter factor induces unwanted bias in the solutions making the reconstructed image
quality and quantitation subjective to such a choice. To remove such subjectivity, in
this thesis, we propose a method that chooses the filter factor automatically in the back-
projection type algorithms based on an optimization scheme that uses a simplex method.
It is also shown later that this approach of finding the filter factor including the recon-
struction of optical properties turns out to be computationally efficient when compared
with Full-Newton reconstruction.

As the proposed algorithm is an approximation to traditional reconstruction meth-

ods, it has a limitation in terms of recovering absolute optical properties for the tissue
under investigation. But in scenarios like rapid dynamic diffuse optical imaging [83–86],
where the aim is to recover the relative difference in the optical properties of the tumor
vasculature, the proposed method can be as effective as the traditional methods with an
added advantage being computationally efficient. The same is shown using both numer-
ical and experimental phantom cases.

There are techniques available for the automated choice of regularization parame-
ter in diffuse optical tomography for the Full-Newton methods, including L-Curve [87],
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 26

Generalized Cross-Validation (GCV) [23], and Minimal Residual Method (MRM) [45].
These methods provide an optimal reconstruction via appropriate choice of regularization
parameter, where an optimal choice provides the balance between spatial and contrast
resolution. The techniques for the automated choice of filter factor in back-projection
type methods (equivalent to regularization parameter) is not in existence till now, mak-
ing this kind of algorithm first of its kind in diffuse optical tomography. The discussion is
limited to continuous-wave(CW) case, where the unknown in the reconstruction problem
is the absorption coefficient(µa ).

3.1 Back-projection Type Method

The aim of this work is to propose a method that removes the bias induced by reg-
ularization parameter (λ), resulting in best compromise between geometric resolution
and contrast resolution. To achieve this, we make an approximation to inversion of ei-
ther JT J + λIn or JJT + λIm by a scalar, which can immensely reduce the required
computation time for calculation of update (4µa ). This approximation is motivated by
geophysical inversion schemes, where the aim is to predict the seismic hazard at a deeper
location using the data obtained at the periphery of the ground as discussed earlier [76].
Note that the aim here is to obtain the sharp changes in material properties in seismic
hazard situation, which is similar to detection of abnormal tissue. Writing the equation
explicitly results in [76]

−1 −1
J T J + λIn = αIn or JJ T + λIm
 
= αIm (3.9)

with α representing the filter factor (scalar value). To prove equivalence of above relation,
 −1  −1
we assume that optimal scalar value equivalent to J T J + λIn and JJ T + λIm is
α1 and α2 respectively, then
 Im
JJ T + λIm =

(3.10)
α1
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 27

and
 In
J T J + λIn =

(3.11)
α2
Now by left and right multiplying Eq. 3.11 by J and J T respectively leads to

 T T T
 JJ T
JJ JJ + λJJ = (3.12)
α2

Taking JJ T common on both sides in Eq. 3.12 gives rise to

 JJ T
JJ T JJ T + λIm =

(3.13)
α2

Substituting Eq. 3.10 in Eq. 3.13 results in,

JJ T JJ T
= (3.14)
α1 α2
 −1
For the above equation to be true, α1 = α2 . In other words, If J T J + λIn = αIn
 T −1
then JJ + λIm = αIm .

As α is a scalar value, the update equation becomes (using Eq. 3.9 in either Eq. 2.22
or 2.23)
4µa = αJ T (y − G(µa )) (3.15)

which is a back-projection type algorithm. The reconstruction parameter in this inver-

sion scheme is α, which can be seen equivalent to λ in the original update equations (Eq.
2.22 and 2.23). As stated earlier, in this work an optimization procedure is adopted to
find the optimal α.

As the aim of the inverse problem is to match y with G(µa ), one can Taylor expand
the G(µa ) around µa0 (which could be a guess or approximation to original µa ), leading
to [73]
G(µa ) = G(µa0 ) + J4µa + (4µa )T H4µa + ... (3.16)
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 28

where J and H are the Jacobian and Hessian evaluated at µa0 respectively and 4µa =
µa - µa0 . Linearizing the above equation and assuming δ = y − G(µa0 ) produces a new
objective function [73]
Ω̃ = kδ − J4µa k22 (3.17)

The optimal choice of α leads to a minimum value of Ω̃, as computing 4µa is dependent
on α (Eq. 3.15). So finding α is posed as an optimization problem to minimize Ω̃ with
respect to α. This proposed scheme to find the filter factor (α) efficiently is one of the
main contribution of this thesis. As the optimization procedure is only for a scalar (filter
factor), a derivative/gradient free simplex method [88] is applied for finding the α and
the corresponding matlab code is made available as an open source [89] for the interested
users. The complexity of simplex algorithm is known to be O(p*NN*NM), where p is
the number of function evaluations in finding the optimal α [88]. Once an optimal α is
found, the same is used in Eq. 3.15 to compute the update of µa .

Flow chart shown in Fig 3.1, gives a good overview of the changes made in the pro-
posed algorithm when compared to the traditional diffuse optical tomographic image
reconstruction. It could be seen clearly that the update equation in flow chart (Fig. 3.1)
is changed when compared to the original image reconstruction algorithm (Fig. 2.2) (Eq.
2.22 is replaced by Eq. 3.15 in the proposed algorithm). Also in the proposed scheme
the filter factor is chosen automatically, but in traditional methods the regularization
parameter was selected empirically.
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 29

Start  

  Ini)al  Guess,  μa0  

μ  =  μa0  

Solve  Forward  Model  G(μa)  

and  Calculate  Jacobian  J  

Experimental  Data  (y)  

μa  =  μa  +  Δμa  
yes  
                           <  1   Stop  

no  

Automated  Selec)on    
of  Filter  Factor  (α)  

Figure 3.1: Flowchart presenting the important steps in reconstruction of optical prop-
erties in back-projection type diffuse optical tomography.

Note that the update (4µa ) obtained by using the Eqs. 2.22, 2.23, and 3.15 is added
to the current µa and re-computation of J and G(µa ) is performed using the updated
µa in the next iterations. The process is stopped when the L2-norm difference in the δ
between successive iterations is less than 1%, as the typical noise level in the data is 1%.
Note that the sensitivity analysis using different stopping criterion (2% and above) has
been performed in Refs. [90,91]. The λ (in either Eq. 2.22 or Eq. 2.23) was kept constant
CHAPTER 3. BACK-PROJECTION TYPE IMAGE RECONSTRUCTION 30

through out iterations, but α was computed using the optimization procedure at every
iteration. All computations were carried out using open-source matlab-based NIRFAST
package [92], which uses finite element based numerical model for light propagation in
tissue. A Linux workstation with dual six-core Intel Xeon processor 2.66GHz with 64
GB RAM is used in obtaining the results presented in this thesis. Note that only single
thread computation, equivalent of using only one processor, is used in this thesis for fair
comparison of computation time.
Chapter 4

LSQR based Optimal Regularization

Parameter Estimation

As discussed in chapter-2, the NIR image reconstruction procedure, also called as the
inverse problem, is known to be a non-linear, ill-posed, and under-determined problem,
due to the dominance of scattering in the light propagation and limited boundary data
available [43]. This procedure is iterative in nature, requiring regularization due to the
ill-posed nature of the problem. Note that the regularization scheme act as a constrain in
the solution space to result in an unique solution. This regularization plays an important
role in determining the quality of reconstructed images as shown in Fig. 2.3. Typical
selection of regularization parameter is either through the empirical choice or prior expe-
rience of the user. Recent works have shown that this regularization parameter could be
automatically estimated using techniques like Generalized Cross Validation(GCV) and
Minimal Residual Method (MRM) in diffuse optical tomography [45, 93].

In diffuse optical tomography usage of MRM for automated selection of regulariza-

tion parameter is shown to improve both spatial and contrast resolution when compared
to traditional GCV based method [45]. The limitation of that work was MRM based
algorithms are computationally expensive due to repeated solving of forward problem to

31
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 32

estimate the optimal regularization parameter [45]. This makes the usage of the MRM-
based automated choice of regularization in real-time prohibitive. It is highly desirable
to develop a computationally efficient methodology, in principle that uses the MRM-
based technique, which can be deployed in real-time. One of the aims of this thesis
is to show that conjugate gradient type LSQR [94–96] (least squares QR) can provide
the computational efficiency needed for automated selection of regularization parame-
ter. Usage of LSQR method for image reconstruction has been widely prevalent [96, 97],
specifically it was used in electrical impedance tomography to perform a computation-
ally efficient reconstruction procedure [96]. Electrical impedance tomography aims at
reconstructing the conductance/impedance of the tissue under investigation using the
measurements made with the help of electrodes placed on the boundary. The interrogat-
ing radiation is an electromagnetic in nature, where either the current or voltage patterns
are sent through the tissue. The under-lying physics is explained via an wave equation,
where as for diffuse optical tomography it is through diffusion equation. Note that both
reconstruction problems are non-linear and ill-posed. In Ref. [96], the regularization pa-
rameter was automatically estimated using L-curve based approach. In this thesis work,
keeping the computational efficiency of LSQR-based reconstruction intact, we have also
deployed it to find the optimal regularization parameter. Note that in reconstruction
problems, like diffuse optical tomographic imaging, automated way of finding the reg-
ularization parameter based on L-curve method was shown to provide overly-smooth
solutions [44]. As LSQR algorithms typically require only matrix-vector computations
(as oppose to matrix-matrix computations), it provides better computational efficiency
as opposed to traditional methods that solve the ill-conditioned inverse problems [94,95].

The LSQR method is analytically equivalent to conjugate-gradient least squares

(CGLS), requiring lesser number of operations compared to the traditional CGLS [94,95].
Moreover, the LSQR method exhibits better numerical properties compared to conju-
gate gradient type procedures when the problem is ill-conditioned (which is the case for
diffuse optical tomographic image reconstruction) [94, 95]. The CGLS method (or its
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 33

variant) has been proven to be effective for solving the inverse problem of NIR tomogra-
phy in terms of work required at every iteration [61,62]. These CGLS methods are not as
straightforward as the direct inversion (also known as full-Newton methods) due to the
ambiguity in choosing the step size [59,62], typically requiring an optimization procedure
in choosing the same. For LSQR-type methods, the choice of number of iterations be-
come critical, as it is known to have the same effect as regularization parameter in typical
ill-conditioned inverse problem [94]. In this work, we have obtained the required optimal
number of iterations based on an optimization procedure. The same optimal number of
iterations have been used for estimating the optimal regularization parameter via a sim-
plex method based optimization procedure [88]. Despite of deploying two optimization
procedures, one for finding the optimal number of iterations and another for optimal
regularization parameter, this method is still computationally efficient compared to the
MRM method (shown later). It will be shown that the obtained reconstructed images
using LSQR-method is similar to the one obtained using MRM in both numerical and
experimental cases. Similar to chapter-3 and earlier work in Ref. [45], the discussion here
is limited to two dimensional (2D) continuous wave (CW) diffuse optical tomography,
where the unknown imaging parameter is the optical absorption coefficient(µa ).

4.1 Regularized Minimal Residual Method (MRM)

The minimization problem given in Eq. 2.21 could be solved using minimal residual
method [98]. The regularization parameter can be obtained optimally using the mini-
mal residual algorithm (described in Algorithm 1) repeatedly, which is extensively dis-
cussed in Ref. [45] (the value of 0 is kept at 10−6 ). In that work the data-model misfit
(ky − G(µa )k2 ) is minimized as a function of regularization parameter using MRM type
algorithm based on the simplex method [88].

The same was compared with the Generalized Cross-Validation (GCV) (explained
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 34

Reconstruction of µa corresponding to frame n

INPUT: J|µn−1 and 4yn ; OUTPUT: 4µn .

Initialize 4µ0n (initial guess), α.

for i = 0, 1, · · · (representing inner iteration number)

1. ri = J|µn−1 4µn i − 4yn

2. lαi = lα (4µin ) = {J|µn−1 }T ri + α4µin

klα (4µin )k2
3. kαi = kJ|µn−1 lα (4µin )k2 +αklα (4µin )k2

4. Update equation: 4µi+1

n = 4µin − kαi lαi

5. The iterative process, steps: 1-4, is terminated when the misfit reaches the given
stopping criterion(o ):kri k2 ≤ o

Algorithm 1: Regularized Minimal Residual (MinRes) Method [99]

later) type of estimation of optimal λ. This is possible as the update equation is a

function of regularization parameter as given by Eq. 2.22. The optimal regularization
parameter (λM RM ) provides the minimum value for ky −G(µa )k2 resulting in an unbiased
(towards λ) estimate of update (4µa ) [45].

4.2 Generalized Cross-Validation (GCV)

The most popular method for estimation of regularization parameter (λ) in ill-posed
problem without any data noise information is Generalized Cross-Validation(GCV) [93,
100, 101]. Estimation of omitted data point could be easily done using a regularized
solution that is obtained by a reduced data set. The estimate of the regularization is
achieved by minimizing a objective function G(λ) defined as [45]

||(JJ T + λI)−1 δ||2

G(λ) = (4.1)
(trace(JJ T + λI)−1 )2
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 35

G(λ) could be efficiently evaluated using the singular value decomposition(SVD) of the
Jacobian and then performing some algebraic simplification. Considering the SVD of the
Jacobian matrix(J) as J = U ΣV , then the above objective function can be simplified
to [45]
Prank(J)  2
uTi δ
i=1 σi2 +λ2
G(λ) = P 2 (4.2)
rank(J) 1
i=1 σi2 +λ2

where σi are the ith singular values of the Jacobian and ui is the ith column of the
orthogonal matrix U . The function (G(λ)) is minimized with respect to λ using the
same scheme as described in the previous section and the minimizer thus obtained is the
required regularization(λGCV ) for the inverse problem [45]. In Ref. [45], it was shown
that the spatial resolution using the MRM method is better than the GCV. Fig 4.1 shows
that MRM is able to perform better than the widely accept GCV based methods, as in
GCV based schemes the regularization parameter is not reducing with iterations and
hence can lead to overly-smooth solutions [44].
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 36

Figure 4.1: Comparison of the spatial resolution of MRM and GCV based optimal reg-
ularization parameter estimation with “perfect” data and 1% noisy data numerically
generated using target [45].

4.3 LSQR-Type Method

Conjugate gradient type LSQR algorithm was used previously in diffuse optical tomog-
raphy with early photons [97] for solving the linearized version of the objective function
(Eq. 2.21). The LSQR type algorithm deployment for estimating the optimal regular-
ization parameter in diffuse optical tomography is one of the main contribution of this
thesis. This is achieved by using Lanczos bidiagonalization of J as given in Ref. [94].
The left and right Lanczos matrices and the bidiagonal matrix is related to Jacobian
matrix (J) are shown below [96, 102]

Uk+1 (β0 e1 ) = δ (4.3)

JVk = Uk+1 Bk (4.4)

CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 37

J T Uk+1 = Vk BkT + αk+1 vk+1 eTk+1 (4.5)

Here B represents the lower bidiagonal matrix, U and V represent the left and right or-
thogonal Lanczos matrices respectively. The unit vector of dimension kx1 is represented
by ek (=1 at the kth row and 0 elsewhere). The dimensions of Uk and Vk are (NMxk) and
(NNxk), with k representing the number of iterations the bidiagonalization is performed.
The δ is the data-model misfit (= y − G(µa )) and ui , vi represents the left and right
Lanczos vectors. The structure of U, V are given by [96, 102]

Uk = [u1 , u2 , ..., uk ]; Vk = [v1 , v2 , ..., vk ] (4.6)

and Bk is a bidiagonal matrix having α1 , ...αk in the main diagonal and β1 , ...βk is the
lower sub-diagonal of the matrix having a dimension of ((k+1)xk) [96, 102].

As the aim of the least square problem is to match y with G(µa ), one can Taylor
expand the G(µa ) around µa0 (which could be a guess or approximation to original µa ),
leading to [73]
G(µa ) = G(µa0 ) + J4µa + (4µa )T H4µa + ... (4.7)

where H and J are the Hessian and Jacobian evaluated at µa0 respectively and 4µa =
µa - µa0 . Linearizing the above equation and assuming δ = y − G(µa0 ) produces a new
objective function (linearized inversion) [73]

Ω̃ = kδ − J4µa k2 (4.8)

using Eqs. 4.3, 4.4 and 4.5 in the argument of Eq. 4.8, resulting in [96, 102]

δ − J4µa = Uk+1 (β0 e1 − Bk x(k) ) (4.9)

CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 38

(k)
where 4µa = Vk x(k) . Substitution of Eq. 4.9 in Eq. 4.8 results in,

Ω̃ = kβ0 e1 − Bk x(k) k2 (4.10)

Considering the first order condition of the Eq. 4.10, we get a new update equation
as [96, 102]
x(k) = (BkT Bk + λI)−1 β0 BkT e1 (4.11)

(k)
where β0 is the L2-norm of the data-model misfit (δ). Once x(k) is calculated, 4µa can
(k)
be evaluated using 4µa = Vk x(k) . Note that the number of operations for obtaining
(k)
4µa in this approach is O(NN2 ), where as traditional one (Eq. 2.22) requires O(NN3 )
operations. As stated earlier, the number of iterations (k) plays an important role in
determining the reconstructed image quality.

4.4 Estimation of Optimal λ using a LSQR-Type

Method
The LSQR type algorithm could be used for calculating the update (Eq. 4.11) and
then multiplying it by the right Lanczos matrix (V). This kind of evaluation of update
turns to be computationally more efficient when compared to traditional way of finding
update using Eq. 2.22. In here taking the advantage of computational efficiency, we
propose a new method of estimating the optimal regularization parameter (λ) using the
simplex method based optimization scheme [88]. The objective function in Eq. 4.8 is
minimized with respect to the regularization parameter λ. This is feasible as the update
4µa = Vk x(k) , where x(k) is a function of λ as given in Eq. 4.11.

The major role in this entire optimization scheme is played by k (number of itera-
tions of the Lanczos bidiagonalization). This determines the size of the bidiagonal matrix
(Bk having dimension of (k+1)xk) and is an important factor in estimation of optimal
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 39

regularization parameter. Increasing the number of Lanczos iterations is inversely pro-

portional to the estimated optimal regularization parameter (increase in k, decreases the
value of λ). The algorithm for determining the optimal number of iterations is given
in Algorithm-2. The first 50 iterations of Lanczos iterations are only considered as for
k=50 the value of λ turns out to be O(10−6 ), which is in the single precision limit.

Input: Lanczos Bidiagonal Matrix Bk ; Vk (k = 1,2, . . . , 50); δ, β0 ; J, µa , λlim .

Output: Optimal number of Lanczos iterations: k opt and Optimal regularization

parameter: λopt

Initialize e1 .

for k = 1,2, . . . , 50

1. Estimate the optimal λ for the given k (λopt

k ).

– Use simplex method to find λopt

k in the range of [0 λlim .] which minimizes Eq.
4.8 with 4µa = Vk ∗ x , where x(k) is found using Eq. 4.11.
(k)

2. Compute 4µa with λ = λopt

k via Eq. 4.11.

3. Make µka = µa +4µa and compute G(µka ) (modeled data).

4. Estimate Ωk = ky − G(µka )k22 .

end

k opt = index of minimal value of Ωk and λopt = λopt

kopt

Algorithm 2: Algorithm for determining the optimal number of Lanczos Iterations (k opt )
and optimal regularization parameter λopt

It is important to note that the optimal regularization parameter (λopt ) is estimated

at every Gauss-Newton iteration in both MRM and LSQR based methods, wherein the
optimal λ is searched within the bound of 0 and previous Gauss-Newton iterations opti-
mal λ (similar to the procedure adapted in Ref. [45]). For the first iteration the bound
(λlim ) for the optimal λ estimation is 1000. The optimal regularization parameter esti-
mated in both these methods reduces with each iteration.
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 40

Flow chart shown in Fig. 4.2, gives a good overview of changes made in the proposed
LSQR-type algorithm when compared to the traditional diffuse optical tomographic im-
age reconstruction. This LSQR type image reconstruction is stopped when L2-norm
difference in the δ between successive iterations is lesser than 2%, as opposed to 1% in
back-projection type reconstruction case, due to the added numerical error of Lanczos
bidiagonalization process. The LSQR method based estimation of optimal regularization
parameter is computationally more efficient than MRM based estimation of regulariza-
tion parameter, because it performs repeated computation of update using a sparse
bidiagonal matrix (Eq. 4.11). MRM kind of estimation requires O(P*NN2 ) (where P
is the number of inner iterations) computation while LSQR type of estimation of opti-
mal regularization parameter requires O(2*Q*k2 ) (where Q is the number of function
evaluations). Since the value of k is very small, it can be intuitively seen that LSQR
kind of approach is more efficient compared to MRM. The same has been verified using
numerical experiments and experimental phantom data. All computations were carried
out using open-source matlab-based NIRFAST package [92], which uses finite element
based numerical model for light propagation in tissue. The algorithm along with nec-
essary MATLAB code is provided as an open-source for enthusiastic users to use this
algorithm [89]. The Lanczos bidiagonalization was performed using the matlab based
regularization tools, which is also an open source [103]. Like other computations a Linux
workstation with dual six-core Intel Xeon processor 2.66GHz with 64 GB RAM has been
used.
CHAPTER 4. LSQR BASED OPTIMAL REGULARIZATION ESTIMATION 41

Start  

  Ini)al  Guess,  μa0  

μ  =  μa0  

Solve  Forward  Model  G(μa)  

and  Calculate  Jacobian  J  

Experimental  Data  (y)  

yes  
                           <  2   Stop  
μa  =  μa  +  Δμa  

no  
Automated  Selec)on    
[Uk  Bk  Vk]  =  LSQR(J)  
of  Size  of  Matrix  (k)  

Automated  Selec)on  of    

Regulariza)on  Parameter  (λ)  

Figure 4.2: Flowchart presenting the important steps in reconstruction of optical proper-
ties in diffuse optical tomography using LSQR-based optimal estimation of regularization
parameter.
Chapter 5

Results and Discussion

5.1 Numerical and Experimental Evaluations of the

proposed Backprojection-type Algorithm

5.1.1 Numerical Experiments

Initially, to asses the effectiveness of the proposed backprojection type method (chapter-
3), a cylindrical mesh (regular geometry) and a breast mesh (irregular geometry) with
background optical properties as µa = 0.01 mm−1 , µ0s = 1 mm−1 and uniform refractive
index of 1.33 is considered. The diameter of cylindrical mesh is 86 mm and height was
110 mm centered around origin. It had two spherical targets mimicking tumors located
diagonally opposite to each other ((-15,15,0) and (15,-15,0)) with a radius of 7.5 mm
and optical properties being µa = 0.02 mm−1 , µ0s = 1 mm−1 . Another case having two
cylindrical targets mimicking tumors were placed diagonally opposite to each other ((-
15,15,0) and (15,-15,0)) with a radius of 7.5 mm and optical properties being µa = 0.02
mm−1 , µ0s = 1 mm−1 .

The two-dimensional cross sections of the target 3D volume is displayed in increments

of 10 mm spanning from z =-55 mm to z=55 mm from left-hand side to right-hand side,
are shown in top row of Fig. 5.1 and Fig. 5.2 for cylindrical mesh having spherical target

42
CHAPTER 5. RESULTS AND DISCUSSION 43

Figure 5.1: Comparison of reconstruction performance in 3D cylindrical mesh using

the standard and proposed backprojection-type optimization work. Here two spherical
anomalies are separated along the diagonal whose target is shown on top. The reconstruc-
tion of the image using 1% noisy data is shown for both proposed backprojection-type
and standard technique. Corresponding line profile is shown through the plot x=-y line
the center (z=0).

and cylindrical target respectively. To generate numerical experimental data, a fine mesh
having 1,01,106 nodes (corresponding to 5,03,023 linear tetrahedral elements) is used.
This data was added with 1% normally distributed gaussian noise to result in experi-
mental mimicking data. The reconstructions were carried on a FEM mesh consisting of
31,446 nodes (corresponding to 1,52,925 linear tetrahedra elements). The data-collection
system had 16 fibers arranged in a circular fashion at z=0 mm, where when one fiber
acts as a source, rest act as detectors. This set up results 240 number of measurements
(NM). Each source was modeled as Gaussian source having full width at half maximum
of 3 mm to mimic the experimental case [104] and is placed at one mean transport length
inside the boundary.

The backprojection type method was also evaluated for robustness to noise, the same
cylindrical mesh was used as in the previous case. It had a spherical target mimicking
CHAPTER 5. RESULTS AND DISCUSSION 44

Figure 5.2: Comparison of reconstruction performance in 3D cylindrical mesh using

the standard and proposed backprojection-type optimization work. Here two cylindrical
anomalies are separated along the diagonal whose target is shown on top. The reconstruc-
tion of the image using 1% noisy data is shown for both proposed backprojection-type
and standard technique. Corresponding line profile is shown through the plot x=-y line
the center (z=0).

tumor located at (15,-15,0) with a radius of 7.5 mm and optical properties being µa =
0.02 mm−1 , µ0s = 1 mm−1 . The two-dimensional cross sections of the target 3D volume
is displayed in increments of 10 mm spanning from z =-55 mm to z=55 mm, from left-
hand side to right-hand side, are shown in top row of Fig. 5.3. The same meshes were
used to generate the data (fine mesh) and performing the reconstruction (coarse mesh).
The generated numerical data was added with 1% and 3% normally distributed gaussian
noises to test the effectiveness of the proposed backprojection-type method. The data-
collection strategy was similar to the previous case. The one-dimensional cross-sectional
plot for the results present in Fig. 5.3 at z=0mm, across the tumor region is given in
Fig. 5.4.

The proposed backprojection method was tested for irregular geometry (breast mesh),
CHAPTER 5. RESULTS AND DISCUSSION 45

Figure 5.3: Comparison of reconstructions using the standard technique and proposed
backprojection-type optimized case for the 3D cylindrical mesh case. We have shown
the target µa and reconstruction with proposed backprojection-type and standard tech-
niques. The reconstructions are performed for 1% and 3% noise cases.

the background optical properties were chosen same as earlier case, here only one spher-
ical target is placed at the centre of imaging domain with a radius of 7.5 mm and optical
properties similar to earlier case. The target distribution is shown in Fig. 5.5 as the
first column. The numerical experimental data was generated on the finite element mesh
having 96,088 nodes (corresponding to 5,95,162 linear tetrahedral elements), like earlier
case 1% noise was added to this data. A coarser mesh having 28,640 nodes (correspond-
ing to 1,32,110 linear tetrahedral elements) was used in the reconstruction. In the both
case, noisy data was calibrated using the coarser mesh to remove any source/detector
biases and modeling errors.

Next, we consider a case of varying contrast in the target mimicking the case of rapid
dynamic diffuse optical imaging, where the region of interest might have pulsatile ab-
sorption changes [86,99]. For this, we have used a circular mesh with background optical
CHAPTER 5. RESULTS AND DISCUSSION 46

0.02
Target
Standard with 1 % noise
Proposed with 1 % noise
0.018 Standard with 3 % noise
Proposed with 3 % noise

0.016
a

0.014
µ

0.012

0.01

0.008
−50 −40 −30 −20 −10 0 10 20 30 40 50
X

Figure 5.4: Profile plot showing the target µa , standard and proposed backprojection-
type reconstruction for 1% and 3% noise case shown in Fig. 5.3. The profile was drawn
across line z=0 in this case.

properties as µa = 0.01 mm−1 , µ0s = 1 mm−1 and a uniform refractive index of 1.33.
The radius of this circular mesh was 43 mm. A circular target mimicking the tumor was
centered at the origin with a radius of 7.5 mm. The optical properties for the tumor was
kept at µa = 0.02 mm−1 and µ0s = 1 mm−1 in the initial frame and then the absorption
coefficient is increased by a factor of 0.005 mm−1 per frame till 0.04 mm−1 and then
reduced by the same factor to 0.02 mm−1 as shown in top row of Fig. 5.6. Experimental
data was generated on a fine mesh having 10,249 nodes (corresponding to 20,160 linear
triangular elements). This data was added with 1% normally distributed gaussian noise
to mimic the experimental data. The data collection strategy was similar to cylindrical
case explained. A coarser mesh having 1,933 FEM nodes (corresponding to 3,726 linear
triangular elements) was used for performing the reconstruction. In both cases, noisy
data was calibrated using the coarser mesh to remove any source/detector biases and
modeling errors [105]. The calibrated data was used as the input to the standard and
proposed reconstruction methods.
CHAPTER 5. RESULTS AND DISCUSSION 47

Target Standard Proposed

Figure 5.5: Comparison of reconstruction performance in breast mesh using the stan-
dard and proposed optimization work. Here we have a single anomaly whose target
along with reconstruction of the image using 1% noisy data is shown for both proposed
backprojection-type and standard technique.

5.1.2 Gelatin Phantom Experiment

To effectively evaluate the proposed back-projection type method in comparison to stan-

dard method, the experimental data collected around the boundary of a gelatin phantom
was used with standard and proposed method case. This gelatin phantom was cylindrical
in shape and had a diameter of 86 mm, height 60 mm was fabricated following the pro-
cedure as explained in [106] and mixtures of 80% of deionised water and 20% of gelatin
(G2625, Sigma Inc) along with India ink for absorption and Titanium oxide(T iO2 ) for
scattering was used to build the background with the optical properties µa = 0.008
mm−1 , µ0s = 0.9 mm−1 at 785 nm wavelength.

To mimic the tumor characteristics, along the Z direction a cylindrical hole of dimen-
sion radius of 8 mm and height of 59 mm was made and filled with intra-lipid mixed with
India ink to act as an absorptive target having the optical properties µa = 0.02 mm−1 , µ0s
= 1.0 mm−1 . A cylindrical mesh consisting 36,425 nodes corresponding to 1,77,662 linear
CHAPTER 5. RESULTS AND DISCUSSION 48

Figure 5.6: Comparison of proposed backprojection-type method with standard image

reconstruction for the stationary target with data have noise level of 1%. The corre-
sponding frame numbers are given on top of first row.

tetrahedral elements was used and the experimental data were also calibrated using a
reference homogenous phantom data. The 2D distributions of expected 3D volume in
steps of 5.45 mm spanning from z = -30 mm to z = 30 mm (left to right) is given in the
top row of Fig. 5.7. The data-collection geometry was similar to numerical case and the
was collected at wavelength of 785 nm.

5.1.3 Dynamic Phantom Experiment

To evaluate the efficacy of the proposed back-projection type method in case of dynamic
diffuse optical imaging, we have used an experimental phantom dataset that was ac-
quired using a video-rate NIR tomography system [107]. This system had an imaging
array of 27 mm in diameter, mimicking the typical small-animal imaging domain size.
The NIR data was collected using 16 equi-spaced channels having a wavelength of 785
nm. A solid phantom having a diameter of 27 mm was having a hole of diameter 6.35
mm (for holding intra-lipid solution). The optical properties at 785 nm wavelength for
CHAPTER 5. RESULTS AND DISCUSSION 49

Figure 5.7: Comparison of reconstructions using the standard technique and proposed
backprojection-type optimized case for the experimental data case. We have shown
the target µa and reconstruction with proposed backprojection-type and standard tech-
niques. Also shown is the profile plot across the line z=0 in this case.

this phantom were µa = 0.0002 mm−1 and µ0s = 1.45 mm−1 . The intra-lipid was having
the same optical properties as that of solid phantom and the undiluted India ink was
injected shortly after the data acquisition. Data acquisition was done at 35 frames per
seconds and then India ink was injected using a pipet. The data was calibrated similar
to numerical data to remove the bias and match the model-response [105].

5.2 Results based on Backprojection Algorithm

The reconstruction results corresponding numerical evaluation of the method are given
in Fig. 5.1, 5.2, 5.3, 5.4, and 5.5 for cylindrical mesh and breast mesh cases. The recon-
struction result for Fig. 5.1 was performed using 1% noisy data and similar trends were
observed with perfect data. Reconstruction distribution for 1% noisy data is shown in
Fig. 5.5. As 3% noisy case was showing similar trend and not adding any value to the
reported outcome/trends, it was omitted here. The results indicate that the contrast
recovery in these 3D cases is much poorer and results from proposed method looks more
smoothened compared to standard reconstruction (Eq. 2.22) in the multiple targets case
CHAPTER 5. RESULTS AND DISCUSSION 50

(cylindrical mesh, Fig. 5.2). The experimental results obtained using gelatin phantom
data are given in Fig. 5.7 along with one-dimensional cross-sectional profiles on the
right hand side of the same figure. The proposed method contrast recovery was higher
than the standard method in this case and resulted qualitatively similar reconstruction
distributions visually. For the standard case, λ = 0.1, 1, and 0.5 for cylindrical, breast,
and phantom cases was used, which gave the best estimates of optical properties.

Figure 5.8: Reconstruction distribution of µa in case of experimental phantom data using

proposed backprojection-type and standard methods. The time corresponding to each
frame is given in the top row. Maximum recovered contrast corresponding to the results
is presented in the graph.

The total computation time along with number of iterations taken for convergence to
the solution was recorded and same is reported in Table-5.1 for the results presented in
Figs. 5.1, 5.5, and 5.7 as test cases. The computation results indicate that the proposed
algorithm results in computationally efficient scheme compared to others.
CHAPTER 5. RESULTS AND DISCUSSION 51

The reconstruction results in the case of cylindrical imaging domain (numerical ex-
periment) are given in Fig. 5.3 using the standard and proposed backprojection-type
method. In the case of standard method, a regularization parameter of 0.1 is used,
which resulted in the best estimates of optical properties. The number of iterations that
were needed to converge to the solution for the case of 1% and 3% cases are 11 and 10
respectively. The resulting values of α, with increasing iteration number, using the pro-
posed method for 1% noisy data case were 1.4475, 1.3891, 0.7034, 2.2651, 0.6010, 5.8290,
0.6201. For 3% noisy data case, the computed α values were, 1.5207, 1.52, 0.7026, 2.3175,
0.6045, 6.0183, 0.5656.

The one-dimensional profile along the tumor at z=0 mm is plotted in Fig. 5.4 for
effective comparison of the quantitative accuracy of the recovered absorption coefficient
using both standard and proposed backprojection-type methods. It is evident from Figs.
5.3 and 5.4 that the quantitative accuracy of the reconstructed optical properties is poor
in the proposed backprojection-type method, as it is an approximate solution of the
standard reconstruction result. The reconstructed image quality is similar in both 1%
and 3% noisy data cases, asserting that the proposed backprojection-type method has
similar behavior as the standard method, capable of handling noisy data. In terms of
the computation time, the total time that was needed for converging to final solution
using Eq. 2.22, 2.23, and 3.15 were 15,624, 2,795 and 1,856 seconds respectively for 1%
noise case, total time for 3% noise case using Eq. 2.22, 2.23, and 3.15 were 14,203, 2,541
and 1,859 seconds respectively.

For the case of mimicking rapid dynamic diffuse optical tomographic imaging, the
recovered two-dimensional cross-sections were shown in Fig. 5.6 using standard and
proposed backprojection-type methods. Since in case of dynamic imaging, the relative
contrast recovery gives more information about the tumor only relative difference is plot-
ted in Fig. 5.6. Even though the obtained reconstructed absorption distributions looks
more diffusive in the case of proposed backprojection-type method, the relative difference
CHAPTER 5. RESULTS AND DISCUSSION 52

Update Equation Cylindrical Mesh Breast Mesh Phantom Mesh

(Fig. 5.1) (Fig. 5.5) (Fig. 5.7)
Eq. 2.22 (Standard) 16,257(11) 16,314(15) 28,168(20)
Eq. 2.23 (Standard) 2,799(11) 3,005(15) 5,116(20)
Eq. 3.15 (Proposed) 1290(5) 1,012(5) 1,781(7)

Table 5.1: Comparison of total computational time (in seconds) for the three cases
considered in this thesis. The last row gives the proposed backprojection-type method
timing results. First two rows corresponding to Eqs. 2.22 and 2.23 gives the timing for
standard methods. The total number of iterations taken to converge are given in the
parenthesis.

in the µa of the target between any two successive reconstructed frames is similar for
both reconstructed methods. The same is plotted in the last-row of Fig. 5.6, warranting
that even though the proposed method has a limitation in terms of recovering absolute
µa values, it is capable of capturing the relative difference in µa between the successive
frames. Note that for this case λ of 1 was used for the standard cases for reconstruct-
ing each frame and typical iterations that were required to converge were 39. For the
proposed backprojection-type method, the number of iterations were 34 and typical α
values were found to be between 0.4 to 0.8 in the process of convergence.

The reconstruction results obtained using experimental phantom data were shown in
Fig. 5.8. Since the data was obtained by injecting undiluted India Ink, the reconstruction
should result in infinite contrast. The exact target is not known, hence a plot of the
maximum value in the distribution is shown in Fig. 5.8. (as maximum value should
be at the site of the tumor). The time 0 seconds indicates the start of the experiment.
The standard reconstruction method used a λ of 1606.7 and the computed α value
for the proposed backprojection-type method was 3.27e-04. Note that the regularization
parameter for the standard scheme was based on heuristic choice that resulted in visually
optimal reconstructed image. Note that in here, only a linear method (single iteration)
was deployed as the aim was to prove the effectiveness of the proposed method. The
results indicate that there is relatively more blur in the reconstructed image using the
CHAPTER 5. RESULTS AND DISCUSSION 53

proposed method and the recovery of the target shape is much poorer using the proposed
method. The recovered contrast (ratio of background to the tumor region) as a function
of time is plotted in the last row of Fig. 5.8 for both standard and proposed methods. It
could be seen that the recovered contrast for the proposed backprojection-type method
matches within 10% of the one obtained using standard method. Note that the expected
contrast should be infinity for this case as undiluted ink was injected into the target
region.

5.3 Discussion of results obtained using Backprojection-

type Algorithm
Back-projection type algorithms that were proposed earlier for diffuse optical tomogra-
phy had heuristically chosen the filter factors [108–110], which can lead to bias in the
reconstructed images. Moreover, the back projected images were typically used to re-
duce the number of unknowns or deducting the resolution characteristics based on the
data-collection strategy [109, 110]. In here, filter factor choice was made using a simplex
method based optimization scheme, resulting in an automated choice. It is also impor-
tant to note that the traditional Fourier/Radon methods used in CT have filter factors
as vector. The automated selection scheme in the reconstruction procedure is not com-
putationally complex, resulting in computation times that are atleast half (excluding the
step of finding α) of the highly optimized traditional methods. The results that were ob-
tained in both numerical and experimental cases were promising and encouraging. The
studies here also indicate that niche for the proposed method could be in rapid dynamic
diffuse optical imaging studies, where typically the image reconstruction procedure is
handled off-line due to computational complexity associated with the traditional meth-
ods. In these dynamic studies, the main aim to capture the relative difference between
the successive frames, the proposed method is performing on par with the traditional
Full-Newton reconstruction methods (Figs. 5.6 and 5.8).
CHAPTER 5. RESULTS AND DISCUSSION 54

The algorithm was evaluated using both irregular and 3D experimental phantom
cases and was shown (Figs. 5.5 and 5.7) to be robust in reconstructing the optical pa-
rameters. The performance in recovery of contrast as well as quality of reconstructed
image is on par with the standard reconstruction method. Note that Eqs. 2.22 and 2.23
are equivalent [75], results in same solution within the round-off limits, and have same
convergence properties.

The algorithm was evaluated using a cylinderical mesh for robustness in reconstruct-
ing the optical properties with varying the noise level. The performance of proposed
method in terms of the contrast recovery was found to be poorer than standard Gauss-
Newton based reconstruction as shown in Fig 5.3. It can also be seen from Fig. 5.3 that
results obtained from the proposed method looks more smoothened compared to stan-
dard reconstruction. But the line profile in Fig. 5.4 indicates that even with increase in
the noise level a reasonably accurate contrast recovery was maintained by the proposed
method.

It is also evident from Table-5.1 that the number of iterations taken for the back pro-
jection type algorithm were atleast three times smaller than the standard reconstruction
method, resulting in fast convergence. As the back-projection algorithm is known to
be equivalent of analytical technique, these direct solutions converge faster compared to
non-linear algorithms. From Table-5.1, it could be deduced that the speed up factor by
using the proposed method could be anywhere between 2-13 depending on the update
equation and mesh size. One of the limitations of the proposed method is the accuracy of
the solutions compared to standard Gauss-Newton approaches, as it is an approximation
to the standard Gauss-Newton approach. The recovered µa using the proposed method
always has more blurring effect compared to standard reconstruction results (Figs. 5.1
and 5.2) as with any back-projection type algorithms [76].
CHAPTER 5. RESULTS AND DISCUSSION 55

It is important to note that generic back-projection algorithms that use radon-

tranforms will not work in diffuse optical imaging, as light does not travel in straight lines,
requiring the computation of sensitivity (Jacobian) matrix, along which the boundary
data needs to be back-projected [108]. As the boundary data is limited and incomplete,
a filter (or scaling factor) is necessary to achieve quantitatively acceptable results. In
here, α acts as that filter, which was chosen based on a optimization scheme resulting
from linearization of the minimization function.

The main advantage of the proposed method is that it eliminates any heuristic (em-
pirical) choice of reconstruction parameters, such as regularization parameter (λ). The
results obtained using such a reconstruction algorithm (the proposed one) are compared
with the standard method, where the λ was chosen based on outcome, i.e. the results that
were obtained using the standard method used the optimal λ that provided close esti-
mates of target µa distribution. If no such information about the target optical properties
were given, such a choice is not possible and may end up in biased reconstructed optical
images. Note that even λ could be obtained using a optimization scheme, which requires
repeated evaluation of (Eq. 2.22), making the order of computation to be O(q*NN3 )
with q as the number of function evaluations (Eq. 2.22), resulting in highly prohibitive
method to be used in any real time setting, especially in three-dimensions. Also for the
case of traditional method, the optimal regularization parameter is not computed using
any methods that are available, even then the comparison of the computational time of
traditional method with the proposed method that includes automated computation of
filter factor, has shown that the proposed method can be computationally efficient com-
pared to the traditional reconstructed methods, especially in three-dimensional imaging.
CHAPTER 5. RESULTS AND DISCUSSION 56

5.4 Numerical and Experimental Evaluation for the

proposed LSQR algorithm

5.4.1 Numerical Experiments

Initially, to asses the effectiveness the LSQR-type optimal estimation of regularization

parameter, a circular mesh with background optical properties as µa = 0.01 mm−1 , µ0s =
1 mm−1 and uniform refractive index of 1.33 is considered. The diameter of the circular
mesh is 86 mm. It had two regions mimicking fatty (with background optical properties),
and tumor regions. The tumor region is a circular region of radius 7.5 mm centered at
(15,0) having optical properties as µa = 0.02 mm−1 , µ0s = 1 mm−1 . To generate numer-
ical experimental data, a fine mesh having 10,249 nodes (corresponding to 20,160 linear
triangular elements) is used. This data is added with 1% and 3% normally distributed
gaussian noise to test the robustness of the method with increasing noise levels. The
reconstructions were carried on a coarser FEM mesh consisting of 1,933 nodes (corre-
sponding to 3,726 linear triangular elements). The data-collection system had 16 fibers
arranged on the boundary of the circular domain, where when one fiber acts as a source,
rest act as detectors. This set up results 16x15 = 240 number of measurements (NM).
Each source was modeled as Gaussian source having full width at half maximum of 3
mm to mimic the experimental case [104] and is placed at one mean transport length
inside the boundary.

The proposed LSQR-type method is tested for multiple anamolies case. The same
meshes were used as explained in the earlier case for generating the data and performing
the reconstruction. The anamolies having a radius of 7.5 mm were placed at (10,0) and
(-10,0) (hence the anamolies are separated by a distance of 5mm). The mesh had the
optical properties for the two regions similar as explained for circular mesh. The images
were reconstructed using the synthetic data with 1% Gaussian distributed noise to see
the effect of the noise level on the reconstructed image quality. The data collection
CHAPTER 5. RESULTS AND DISCUSSION 57

Figure 5.9: Gelatin phantom used in the experimental studies [111] to verify the proposed
LSQR-type algorithm, the two-dimensional cross-sectional geometry is as shown in the
top left of Fig. 5.12.

strategy was similar as explained for circular mesh.

5.4.2 Gelatin Phantom Experiment

The proposed LSQR based scheme was also tested with experimental data which was
generated by a gelatin phantom [59] of radius 43 mm, height 25 mm which was fabri-
cated using a mixture of India ink for producing absorption effect and Titanium oxide
(TiO2 ) for scattering. Layers of gelatin was fabricated by hardening heated gelatin solu-
tion (having a concentration of 20% of gelatin (G2625, Sigma Inc) and 80% of deionized
water) along with already mentioned mixture, which is shown in Fig. 5.9. The outer
layer mimicking the adipose region has a thickness of 10mm, and having µa = 0.0065
mm−1 and µ0s = 0.65 mm−1 . The middle layer was kept at 76 mm diameter, which
mimics the fibro-glandular layer, having the optical properties at µa = 0.01 mm−1 and
µ0s = 1.0 mm−1 . A cylinderical hole extending in Z direction was filled with intra-lipid
mixed with India ink acting as a tumor having optical properties of µa = 0.02 mm−1 and
CHAPTER 5. RESULTS AND DISCUSSION 58

µ0s = 1.2 mm−1 with a radius of 8 mm and heigth of 24 mm. Laser diode with 785 nm
wavelength was used as source for validation of individual layers optical properties from
the data obtained from large cylindrical samples of each layer. The data is collected
using a single layer of fibers (at z=0mm) leading to 240 lnA data points. The collected
data was calibrated using the coarser mesh with 1933 FEM nodes (corresponding to 3726
linear triangular elements).

Figure 5.10: The target µa distribution is shown on the top-left corner. The recon-
struction method along with the data noise level are given on top of each reconstructed
µa distribution. The one-dimensional cross-sectional plot of target and reconstructed
µa distributions along the line shown in the target image are given in the bottom-right
corner.

5.5 Results for the proposed LSQR-type algorithm

The reconstruction results using MRM and LSQR type methods for the case of numerical
experiment with single target and varying noise levels are shown in Fig. 5.10 along with
the one-dimensional profile along the line of the target image. The method along with
percentage of noise level in the data (in parenthesis) is given on top of corresponding
CHAPTER 5. RESULTS AND DISCUSSION 59

reconstructed µa distribution in Fig. 5.10. The total computation time for the MRM and
the proposed method to converge has been reported in first and second column of Table-
5.3 respectively. The results show that the proposed method performance is similar to
MRM method, leading to conclusion that the proposed methods is in spirit is similar
to the MRM method. Figure 5.11 shows the reconstructed results obtained using the
multiple target data. The method used for obtaining the reconstructed µa distribution
is given correspondingly on top. This results also indicate that LSQR type of optimal
regularization estimation may be better than traditional MRM based estimation of λ.

Target MRM

0.01 0.015 0.02 0.01 0.015 0.02

LSQR
Target
MRM
0.02 LSQR
µa

0.015

0.01

−50 0 50
0.010.0150.02 X

Figure 5.11: Similar effort as Fig. 5.10 except the target µa distribution has two anoma-
lies and the noise level in the data is kept at 1%

The experimental results that were obtained using gelatin phantom data are given
in Fig. 5.12 along with one-dimensional cross-sectional profiles on the right hand side of
CHAPTER 5. RESULTS AND DISCUSSION 60

Method MRM Method [45] LSQR Method

Target (Fig. 5.10) λ = 0.0083, 3.1e-04 λ = 0.0056, 2.6e-04
Target (Fig. 5.11) λ = 0.0139, 0.0010, 5.6e-05 λ = 0.0048, 5.7e-04, 2.99e-04,
2.93e-04, 2.02e-04, 1.91e-04, 1.91e-04
Experimental λ = 0.0126, 0.0029, 4.7e-04 λ = 0.0188, 0.0042, 0.0014,
Phantom (Fig. 5.12) 4.4e-04

Table 5.2: Comparison of regularization parameters obtained for the results presented in
this work. The λ values are decreasing with increasing number of iterations for all cases.

the same figure. The estimated λ’s and total computational time were given in Table-
5.2 and 5.3 correspondingly for all the results proposed in this work. The result in
Fig. 5.12 also indicates that the contrast recovery of the proposed method was found
to be higher than the standard method (MRM based) resulting in quantitatively better
reconstruction distributions using the proposed method. In all these methods the value
of optimal regularization parameter was estimated at each iteration, it is observed that
the optimally chosen regularization parameter reduces with each Gauss-Newton iteration
(Refer to Table-5.2). The computation time reported in Table-5.3 is the total time taken
for performing reconstruction (including estimation of optimal λ and solving the inverse
problem using this optimal λ).

Method MRM Method [45] LSQR Method

Target (Fig. 5.10) 224.90(2) 10.66(2)
Target (Fig. 5.11) 133.42(3) 31.88(7)
Experimental Phantom (Fig. 5.12) 179.98(3) 14.76(4)

Table 5.3: Comparison of total computational time (in seconds) including the overhead
time, for the three cases considered in this work. The last column indicates the proposed
method timing results. The second column indicates the image reconstruction using
regularized MRM method. The total number of iterations taken to converge are given in
the parenthesis. The corresponding regularization parameter values are given in Table-
5.2.
CHAPTER 5. RESULTS AND DISCUSSION 61

Target MRM

0.01 0.015 0.02 5 10 15

−3
LSQR x 10
0.025
Target
0.02 MRM
LSQR
0.015
µa

0.01

0.005

0
−40 −20 0 20 40
0.01 0.02 X

Figure 5.12: Performance comparison of MRM and the proposed LSQR method for the
case of gelatin phantom data

5.6 Discussions of results obtained using proposed

LSQR-type algorithm
The main advantage of the LSQR based optimal estimation of regularization parame-
ter is that it eliminates any heuristic (empirical) choice of reconstruction parameters,
such as regularization parameter (λ). The proposed algorithm is compared with existing
method of estimation of optimal regularization parameter (MRM), it could be seen that
the contrast recovery of the proposed method is better than the MRM based algorithm.
This justifies that estimation of optimal reconstruction parameter is better in case of
proposed algorithm when compared to traditional one. But both these schemes estimate
the optimal regularization parameter and there is no blowing of regularization param-
eter as in case of GCV [45]. As there are no reconstruction parameters involved and
relying on LSQR type method results in computationally efficient scheme. The proposed
algorithm is also efficient in terms of doing the reconstruction, compared to traditional
CHAPTER 5. RESULTS AND DISCUSSION 62

matrix inversion. For the results presented in here, the total time taken by traditional
reconstruction (Eq. 2.22) is 62.29 seconds for two anomalies case, 28.13 seconds for sin-
gle target case and 14.55 seconds for experimental case. The corresponding computation
time taken by LSQR method are reported in Table-5.3. The number of iterations taken
for convergence in traditional reconstruction for two anomalies target case is 35, single
anomaly target case is 16 and for experimental case is 7. For LSQR-type method the
corresponding number of iterations were 7, 2 and 4, leading to a computationally efficient
reconstruction procedure.

The proposed method requires solving of the forward model for estimating the optimal
number of Lanczos iterations required, but solving of forward model is performed only
for fixed number of times (k = 1, 2, . . . , 50), which is not the case with the MRM based
scheme (as it is solved repeatedly in a simplex based method [45]). In here, taking the
advantage of computational efficiency of LSQR method, we have estimated the optimal
regularization parameter using a simplex method. This automated choice was compared
with the traditional MRM method and found to be resulting in similar reconstructed
image quality. These type of methods will remove the unwanted bias induced by the
heuristic choice of regularization parameter, making the information provided by diffuse
optical tomographic image reconstruction more objective in nature.
Chapter 6

Conclusion

Diffuse optical tomographic inverse problem is non-linear and ill-posed requiring regu-
larization or filtering of high frequency components to converge to an acceptable solu-
tion. The usage of quadratic regularization with Gauss-Newton minimization scheme is
the most popular among the image reconstruction algorithms, where the regularization
parameter dictates the reconstructed image quality. This regularization parameter is
typically chosen empirically and requires either prior experience/information for obtain-
ing optimal reconstruction results. Often, such choice leads to bias in the reconstruction
results. In here, a simple back-projection type reconstruction algorithm is presented in
combination with an optimization scheme for an automated selection of filter factor to
speed up the computation and remove the biases induced by the empirical choice of the
reconstruction parameter. It is shown through numerical and phantom data that, even
though the back-projection algorithms performance for absolute quantitative diffuse op-
tical imaging is not on par with the traditional methods, they are effective in capturing
the relative difference in the optical properties.

In traditional Gauss-Newton minimization scheme, the estimation of regularization

parameter was done using a LSQR based method, the same was compared with the
MRM-based optimal regularization parameter estimation. It was found that LSQR based
scheme gives similar results as that of the MRM-based method with having an added

63
CHAPTER 6. CONCLUSION 64

advantage of being computationally inexpensive. This kind of estimation of optimal

hyper-parameters will remove the bias induced by the empirical choice of regularization
parameter and can be used to estimate optimal regularization parameter in real-time
settings. The developed algorithms are provided as open-source [89] (Matlab-based) for
interested users in Appendix-A.

Developed methodologies in this thesis work for finding hyper parameters in the dif-
fuse optical tomographic image reconstruction scheme, fall in the general inverse problem
schemes, hence these methods can be tested with various inverse problem based applica-
tions like image deblurring/denoising as well as any other ill-posed estimation problems.
The results that were presented here are pertained to diffuse optical tomography, the
proposed methods should be applicable to other tomographic inverse problems having
similar characteristics, such as electrical impedance tomography [112], electrical capaci-
tance tomography [113], fluorescence optical tomography [114] and bioluminescence to-
mography [115]. Also parallelizing these algorithms with usage of modern day graphic
processing units will further help in performing the image reconstruction in real-time
and remove the bottleneck of additional computation burden in the image reconstruc-
tion procedure [116].
Appendix A

List of Packages/Programs

This appendix gives the list of important MATLAB-programs that were created as a
part of this thesis work used for carrying out the computational work.

A.1 Packages
The MATLAB-based software packages that were used in this work are:
1. NIRFAST
Link: http://code.google.com/p/nirfast/
Version: 5
Accessed on: September 27, 2012.
The description of the NIRFAST is provided in Sec. 3.1 of this thesis.
2. Regularization Tools
Link: http://www2.imm.dtu.dk/ pch/Regutools/
Version:4.1
Accessed on: September 27, 2012.
This is a matlab package for analysis and solution of discrete ill-posed problems as
explained in Ref [103]

65
APPENDIX A. LIST OF PACKAGES/PROGRAMS 66

A.2 Developed Programs

A brief description of the developed programs for performing the work presented in
this thesis are given in the table below and are posted at:
https://sites.google.com/site/sercmig/home/prakashprograms/.

Program Description/Usage/Related Programs

recon dynamic simulation.m Dynamic imaging case with stationary target,

this runs both standard reconstruction (Eq.
2.22) and the proposed backprojection type al-
gorithms (Eq. 3.15). This program will also
plots the variation of contrast and frames and
rasterize the result as shown in Fig. 5.6. Related
programs: reconstruct stnd cw backproj.m, re-
construct stnd cw fwd.m
reconstruct linear data Stnd.m Dynamic case using for running an experimental
phantom case with standard image reconstruc-
tion using Eq. 2.22.
reconstruct linear data back.m Dynamic case using for running an experimental
phantom case with the proposed backprojection
type image reconstruction using Eq. 3.15. Also
this program plots the image in a rasterized for-
mat.
reconstruct stnd cw backproj.m This code performs the image reconstruction
using the backprojection algorithm with auto-
matic choice of filter factor (α). Usage: recon-
struct stnd cw backproj(fwd mesh, data fn, it-
eration, output fn, filter n). Related Programs:
opt mu objfun.m
APPENDIX A. LIST OF PACKAGES/PROGRAMS 67

Program Description/Usage/Related Programs

opt mu objfun.m This code is used to estimate the optimal filter factor
α by minimizing the objective function Ω. Usage:
omega = opt mu objfun(jacobian, data diff, alpha)
reconstruct stnd cw fwd.m Backprojection Type automatic estimation of opti-
mal filter factor is compared with this code (consid-
ered as standard image reconstruction). Usage: re-
construct stnd cw fwd(fwd mesh, data fn, iteration,
lambda, output fn, filter n)
reconstruct cw lsqr.m This code performs the image reconstruction us-
ing the LSQR based algorithm with automatic
choice of regularization parameter (λ) Usage:
reconstruct cw lsqr(fwd fn, recon basis, frequency,
data fn, iteration, lambda, output fn, filter n), Re-
lated program: opt lambda cw.m, lsqr b hybrid.m
opt lambda cw.m This code is used to estimate the optimal regulariza-
tion parameter λ by minimizing the objective func-
tion Ω. Usage: omega = opt lambda cw(Bidiagonal,
data diff, lambda, V ortho, Jacobian, I matrix,
U ortho)
lsqr b hybrid.m This code is used to perform the Lanczos Bidiag-
onalization of the jacobian matrix (J). This code
was modified by getting the original LSQR code from
the Regularization Tools package. Usage: [U ortho,
V ortho, Bidiagonal, F] = lsqr b hybrid(A, b, iter,
reorth, s)
APPENDIX A. LIST OF PACKAGES/PROGRAMS 68

Program Description/Usage/Related Programs

reconstruct stnd cw OMRM.m The code performs the image reconstruction

using the MRM based algorithm for auto-
matic selection of optimal regularization pa-
rameter (λ) [45]. This same code is used
as standard for comparison with the pro-
posed LSQR based algorithm. Usage: recon-
struct stnd cw OMRM(fwd mesh, data fn, it-
eration, output fn, filter n) Related Programs:
ObjectiveFunction.m, MRM.m
ObjectiveFunction.m This function is used for estimation optimal
regularization parameters by minimizing the
objective function (Ω = ky − G(µa )k2 ) in a
least-square sense [45]. Usage: Omega = Ob-
jectiveFunction(mesh, anom, foo, J, data diff,
lambda). Related Programs: MRM.m
MRM.m This code implements the Minimal Residual
Method which is used to perform the f oo =
(JT J + λI)−1 JT δ [45]. Usage: foo = MRM(J,
foo, data diff, alpha)
Bibliography

[1] American cancer society, statistics for 2012 : Cancer facts & figures 2012.
http://www.cancer.org/docroot/STT/stt 0.asp accessed Sep. 25 (2012).

[2] M. Cutler, “Transillumination of the breast,” Annals of Surg. 93, 223–234 (1931).

[3] R. J. J. Bartrum and H. C. Crow, “Transillumination lightscanning to diagnose

breast cancer: A feasibility study,” Am. J. Roentgenol. 142, 409–414 (1984).

[4] E. Carlsen, “Transillumination light scanning,” Diagn. Imaging 4, 28–34 (1982).

[5] J. J. Gisvold, L. R. Brown, R. G. Swee, D. J. Raygor, N. Dickerson, and M. K.

Ranfranz, “Comparison of mammography and transillumination light scanning in
the detection of breast lesions,” Am. J. Roentgenol. 147, 191–194 (1986).

[6] M. J. Homer, “Breast imaging: Pitfalls, controversies, and some practical thoughts,”
Radiol. Clin. of North Am. 23, 459–472 (1985).

[7] V. Marshall, D. C. Williams, and K. D. Smith, “Diaphanography as a means of

detecting breast cancer,” Radiology 150, 339–343 (1984).

[8] B. Monsees, J. M. Destouet, and W. G. Totty, “Light scanning versus mammography

in breast cancer detection,” Radiology 163, 463–465 (1987).

[9] A. Pera and A. K. Freimanis, “The choice of radiologic procedures in the diagnosis
of breast disease,” Obstet Gynecol Clin. North Am. 14, 635–650 (1987).

[10] A. E. Profio, G. A. Navarro and O. W. Sartorius, “Scientific basis of breast di-

aphanography,” Med. Phys. 16, 60–65 (1989).

69
BIBLIOGRAPHY 70

[11] E. A. Sickles, “Breast cancer detection with transillumination and mammography,”

Am. J. Roentgenol. 142, 841–844 (1984).

[12] H.Wallberg, A. Alveryd, U. Bergvall, K. Nasiell, P. Sundelin, and S. Troell, “Di-

aphanography in breast carcinoma. Correlation with clinical examination, mam-
mography, cytology and histology,” Acta Radiol. Diagn. 26, 33–44 (1985).

[13] D. J. Watmough, “Transillumination of breast tissues: factors governing optimal

imaging of lesions,” Radiology 147, 89–92 (1983).

[14] A. Alveryd, I. Andersson, K. Aspegren, G. Balldin, N. Bjurstam, G. Edstrom, G.

Fagerberg, U. Glas, O. Jarlman, S. A. Larsson, and et al, “Lightscanning versus
mammography for the detection of breast cancer in screening and clinical practice.
a Swedish multicenter study,” Cancer 65, 1671–1677 (1990).

[15] B. Chance and R. R. Alfano, “Optical tomography, photon migration, and spec-
troscopy of tissue and model media: Theory, human studies, and instrumentation,”
Proc. of SPIE 2389, (1995).

[16] A. Yodh and B. Chance, “Spectroscopy and imaging with diffusing light,” Phys.
Today 48, 34–40 (1995).

[17] A. G. Yodh and D. A. Boas, “Functional imaging with diffusing light,” Biomedical
Photonics. ed. Vo-Dinh, T. Boca Raton: CRC Press, 21-1–21-45 (2003).

[18] B. Chance, S. Nioka, J. Kent, K. McCully, M. Fountain, R. Greenfeld, and G.

Holtom, “Time-resolved spectroscopy of hemoglobin and myoglobin in resting and
ischemic muscle,” Anal. Biochem. 174, 698–707 (1988).

[19] A. Duncan, T. L. Whitlock, M. Cope, and D. T. Delpy, “Multiwavelength wide-

band, intensity modulated optical spectrometer for near infrared spectroscopy and
imaging,” Proc. of SPIE 1888, (1993).

[20] M. Ferrari, Q. Wei, L. Carraresi, R. De Blasi, and G. Zaccanti, “Time-resolved spec-

troscopy of the human forearm,”J. Photochem. Photobiol. B. 16, 141–153 (1992).
BIBLIOGRAPHY 71

[21] M. Ferrari, D. A. Wilson, D. F. Hanley, J. F. Hartmann, M. C. Rogers, and R. J.

Traystman, “Noninvasive determination of hemoglobin saturation in dogs by deriva-
tive near infrared spectroscopy,” Am. J. Physiol. 256, H1493–H1499 Part 2 (1989).

[22] A. Hielscher, F. K. Tittel, and S. L. Jacques, “Non-invasive monitoring of blood

oxygenation by phase-resolved transmission spectroscopy,” Proc. of SPIE, 275–288
(1993).

[23] J. P. Culver, T. Durduran, D. Furuya, C. Cheung, J. H. Greenberg, and A. G. Yodh,

“Diffuse optical tomography of cerebral blood flow, oxygenation, and metabolism
in rat during focal ischemia,” J. Cereb. Blood Flow Metab. 23, 911–924 (2003).

[24] T. Durduran, G. Yu, M. G. Burnett, J. A. Detre, J. H. Greenberg, J.Wang, C. Zhou,

and A. G. Yodh, “Diffuse optical measurement of blood flow, blood oxygenation,
and metabolism in a human brain during sensorimotor cortex activation,” Opt. Lett.
29, 1766–1768 (2004).

[25] Regine Choe - PhD Thesis, “Diffuse optical tomography and spectroscopy of breast
cancer and fetal brain” (2005).

[26] C. Menon, G. M. Polin, I. Prabakaran, A. Hsi, C. Cheung, J. P. Culver, J. F.

Pingpank, C. S. Sehgal, A. G. Yodh, D. G. Buerk, and D. L. Fraker, “An integrated
approach to measuring tumor oxygen status using human melanoma xenografts as
a mode,” Cancer Res. 63, 7232–7240 (2003).

[27] Website: http://omlc.ogi.edu/spectra/ (accessed on: August 20, 2007).

[28] S. L. Jacques, N. Ramanujam, G. Vishnoi, R. Choe, and B. Chance, “Modeling

photon transport in transabdominal fetal oximetry,” J. Biomed. Opt. 5, 277–282
(2000).

[29] A. Zourabian, A. Siegel, B. Chance, N. Ramanujam, M. Rode, and D. A. Boas,

“Transabdominal monitoring of fetal arterial blood oxygenation using pulse oxime-
try,” J. Biomed. Opt. 5, 391–405 (2000).
BIBLIOGRAPHY 72

[30] G. Vishnoi, A. H. Hielscher, N. Ramanujam, and B. Chance, “Photon migration

through fetal head in utero using continuous wave, near-infrared spectroscopy: de-
velopment and evaluation of experimental and numerical models,” J. Biomed. Opt.
5, 163–172 (2000).

[31] A. Prentice and T. Lind, “Fetal heart rate monitoring during labour - too frequent
intervention, too little benefit?,” Lancet 2, 1375–1377 (1987).

[32] F. A. Manning, C. R. Harman, I. Morrison, S. M. Menticoglou, I. R. Lange and J.

M. Johnson, “Fetal assessment based on fetal biophysical profile scoring,” Am. J.
Obstet. Gynecol 162, 703–709 (1990).

[33] J. P. Lavery, “Nonstress fetal heart rate testing,” Clin. Obstet. Gynecol. 25, 689–705
(1982).

[34] R. F. Brem, J. Baum, M. Lechner, S. Kaplan, S. Souders, L. G. Naul, and

J. Hoffmeister, “Improvement in sensitivity of screening mammography with
computer-aided detection: a multiinstitutional trial,” Am. J. Roentgenol. 181, 687–
693 (2003).

[35] K. Armstrong, E. Moye, S. Williams, J. A. Berlin, and E. E. Reynolds, “Screening

Mammography in Women 40 to 49 Years of Age: A Systematic Review for the
American College of Physicians,” Ann. Intern. Med. 146, 516–526 (2007).

[36] T. Hata, H. Takahashi, K. Watanabe, M. Takahashi, K. Taguchi, T. Itoh, and S.

Todo, “Magnetic resonance imaging for preoperative evaluation of breast cancer: a
comparative study with mammography and ultrasonography,” J. Am. Coll. Surg.
198, 190–197 (2004).

[37] B. Bone, Z. Pentek, L. Perbeck, and B. Veress, “Diagnostic accuracy of mammogra-

phy and contrastenhanced MR imaging in 238 histologically verified breast lesions,”
Acta. Radiol. 38, 489–496 (1997).
BIBLIOGRAPHY 73

[38] S. J. Nass, I. C. Henderson, and J. C. Lashof ( eds), “Mammography and Beyond:

Developing technologies for the early detection of breast cancer,” National Academy
Press: Washington DC (2001).

[39] M. F. Ernst and J. A. Roukema, “Diagnosis of non-palpable breast cancer: a re-

view,” Breast 11, 13–22 (2002).

[40] N. Hylton, “Magnetic Resonance Imaging of the Breast: Opportunities to Improve

Breast Cancer Management,” J. Clin. Oncol. 23, 1678–1684 (2005).

[41] E. S. Hwang, K. Kinkel, L. J. Esserman, Y. Lu, N. Weidner, and N. M. Hylton,

“Magnetic Resonance Imaging in Patients Diagnosed With Ductal Carcinoma-In-
Situ: Value in the Diagnosis of Residual Disease, Occult Invasion, and Multicentric-
ity,” Ann. Surg. Oncol. 10, 381–388 (2003).

[42] A. Hlawatsch, A. Teifke, M. Schmidt, and M. Tehelen, “Preoperative assessment of

breast cancer: Sonography versus MR imaging,” Am. J. Roentgenol. 179, 1493–1501
(2002).

[43] S. R. Arridge, “Optical tomography in medical imaging,” Inv. Problems 15, R41–
R93 (1999).

[44] J. P. Culver, R. Choe, M. J. Holboke, L. Zubkov, T. Durduran, A. Slemp, V. Ntzi-

achristos, B. Chance, and A. G. Yodh, “Three-dimensional diffuse optical tomogra-
phy in the parallel plane transmission geometry: evaluation of a hybrid frequency
domain/continuous wave clinical system for breast imaging,” Med. Phys. 30, 235–
247 (2003).

[45] R. P. K. Jagannath and P. K. Yalavarthy, “Minimum Residual Method Provides

Optimal Regularization Parameter for Diffuse Optical Tomography,” J. Biomed.
Opt. 17, 106015 (2012).

[46] S. R. Arridge and M. Schweiger, “Photon-measurement density functions. Part 2:

Finite-element-method calculations,” Appl. Opt. 34, 8026-8037 (1995).
BIBLIOGRAPHY 74

[47] H. Jiang, K. D. Paulsen, U. Osterberg, B. W. Pogue, and M. S. Patterson, “Optical

image reconstruction using frequency domain data: simulations and experiments,”
J. Opt. Soc. Am. A 13, 253–266 (1996).

[48] S. R. Arridge, M. Cope, and D. T. Delpy, “The theoretical basis for the determina-
tion of optical pathlengths in tissue: temporal and frequency analysis,” Phys. Med.
Biol. 37, 1531–1559 (1992).

[49] F. Martelli, A. Sassaroli, Y. Yamada, and G. Zaccanti, “Analytical approximate

solutions of the time-domain diffusion equation in layered slabs,” J. Opt. Soc. Am.
A 19, 71–80 (2002).

[50] F. A. Grunbaum, “Diffuse tomography: the isotropic case,” Inverse Problems 8,

409–419 (1992).

[51] R. F. Bonner, R. Nossal, S. Havlin, and G. H. Weiss, “Model for photon migration
in turbid biological media,” J. Opt. Soc. Am. A 4, 423–432 (1987).

[52] C. K. Hayakawa, J. Spanier, F. Bevilacqua, A. K. Dunn, J. S. You, B. J. Tromberg

and V. Venugopalan, “Perturbation Monte Carlo methods to solve inverse photon
migration problems in heterogeneous tissues,” Opt. Lett. 26, 1335–1337 (2001).

[53] Y. P. Kumar and R. M. Vasu, “Reconstruction of optical properties of low-scattering

tissue using derivative estimated through perturbation Monte-Carlo method,” J.
Biomed. Opt. 9, 1002–1012 (2004).

[54] K. Ren, G. S. Abdoulaev, G. Bal, and A. H. Hielscher, “Algorithm for solving the
equation of radiative transfer in the frequency domain,” Opt. Lett. 29, 578–580
(2004).

[55] H. Dehghani, B. W. Pogue, S. P. Poplack, and K. D. Paulsen, “Multiwavelength

threedimensional near-infrared tomography of the breast: initial simulation, phan-
tom, and clinical results,” Appl. Opt. 42, 135–145 (2003).
BIBLIOGRAPHY 75

[56] A. D. Zacharopoulos, S. R Arridge, O. Dorn, V. Kolehmainen, and J. Sikora, “Three-

dimensional reconstruction of shape and piecewise constant region values for opti-
cal tomography using spherical harmonic parametrization and a boundary element
method,” Inv. Problems 22, 1509–1532 (2006).

[57] H.Dehghani, S. Srinivasan, B. W. Pogue and A. Gibson, “Numerical modelling and

image reconstruction in diffuse optical tomography,” Phil. Trans. R. Soc. A 367,
3073–3093 (2009)

[58] M. Schweiger, S. R. Arridge, M. Hiroaka, and D. T. Delpy, “The Finite Element

Model for the Propagation of Light in Scattering Media: Boundary and Source
Conditions,” Med. Phys. 22, 1779–1792 (1995).

[59] P. K. Yalavarthy, B. W. Pogue, H. Dehghani, and K. D. Paulsen, “Weight-Matrix

Structured Regularization Provides Optimal Generalized Least-Squares Estimate in
Diffuse Optical Tomography,” Med. Phys. 34, 2085–2098 (2007).

[60] Phaneendra K. Yalavarthy, PhD Thesis - “A Generalized least-squares minimization

method for near infrared diffuse optical tomography” (2007)

[61] S. R. Arridge and M. Schwieger, “Gradient-based optimisation scheme for optical

tomography,” Opt. Exp. 2, 212-226 (1998).

[62] A. H. Hielscher, A. D. Klose, and K. M. Hanson, “Gradient-based iterative image

reconstruction scheme for time- resolved optical tomography,” IEEE Trans. Med.
Imag. 18, 262-271 (1999).

[63] K. D. Paulsen and H. Jiang, “Enhanced frequency domain optical image reconstruc-
tion in tissues through variation minimization,” App. Opt. 35, 3447–3458 (1996).

[64] J. C. Ye, C. A. Bouman, K. J. Webb, and R. P. Milane. “Nonlinear multigrid

algorithms for Bayesian optical diffusion tomography,” IEEE Trans. Imag. Process.
10, 909–922 (2001)
BIBLIOGRAPHY 76

[65] M. Guven, B. Yazici, X. Intes, and B. Chance, “Diffuse optical tomography with a
priori anatomical information,” Phys. Med. Biol. 50, 2837–2858 (2005).

[66] K. Levenberg, “A method for the solution of certain nonlinear problems in least
squares,” Q. Appl. Math. 2, 164-168 (1944).

[67] D. W. Marquardt, “An algorithm for least squares estimation of nonlinear parame-
ters,” J. Soc. Ind. Appl. Math. 11, 431-441 (1963).

[68] A. N. Tikhonov and V. A. Arsenin, “Solution of Ill-posed Problems,” Winston and

Sons, Washington 1977.

[69] A. Tarantola, “Inverse Problem Theory and Methods for Model Parameter Estima-
tion,” SIAM, 2004.

[70] B. W. Pogue, T. McBride, J. Prewitt, U. L. Osterberg, and K. D. Paulsen, “Spatially

variant regularization improves diffuse optical tomography,” Appl. Opt. 38, 2950–
2961 (1999).

[71] H. Niu, P. Guo, L. Ji, Q. Zhao, and T. Jiang, “Improving image quality of diffuse
optical tomography with a projection-error-based adaptive regularization method,”
Opt. Express 16, 12423–12434 (2008).

[72] N. Cao, A. Nehorai, and M. Jacob, “Image reconstruction for diffuse optical tomog-
raphy using sparsity regularization and expectation-maximization algorithm,” Opt.
Express 15, 13695–13708 (2007).

[73] S. H. Katamreddy and P. K. Yalavarthy, “Model-Resolution Based Regularization

Improves Near Infrared Diffuse Optical Tomography,” J. Opt. Soc. Am. A 29, 649–
656 (2012).

[74] X. Intes, C. Maloux, M. Guven, B. Yazici, and B. Chance, “Diffuse optical to-
mography with physiological and spatial a priori constraints,” Phys. Med. Biol. 49,
N155–N163 (2004).
BIBLIOGRAPHY 77

[75] P. K. Yalavarthy, D. R. Lynch, B. W. Pogue, H. Dehghani, and K. D. Paulsen,

“Implementation of a computationally efficient least-squares algorithm for highly
under-determined three-dimensional diffuse optical tomography problems,” Med.
Phys. 35, 1682–1697 (2008).

[76] A. Tarantola, “Linearized inversion of seismic reflection data,” Geophysical

Prospecting 32, 998–1015 (1984).

[77] S. A. Walker, S. Fantini, and E. Gratton, “Image reconstruction by backprojection

from frequency-domain optical measurements in highly scattering media,” Appl.
Opt. 36, 170–178 (1997).

[78] P. C. Jackson, P. H. Stevens, J. H. Smith, D. Kear, H. Key, and P. N. Wells,

“The development of a system for transillumination computed tomography,” Brit.
J. Radiol. 60, 375–380 (1987).

[79] A. C. Kak and M. Slaney, “Principles of Computerized Tomographic Imaging,”

IEEE, 49–75 (1988).

[80] P. Munshi, R. K. S. Rathore, K. S. Ram, and M. S. Kalra, “Error estimates for

tomographic inversion,” Inv. Problems 7, 399–408 (1991).

[81] P. Munshi, “Error analysis of tomographic filters I: Theory”, NDT&E International

25, 191–194 (1992).

[82] P. Munshi, R.K.S. Rathore, K.S. Ram, and M.S. Kalra, “Error analysis of Tomo-
graphic filters II: Results”, NDT&E International 26, 235–240 (1993).

[83] C. H. Schmitz, H. L. Graber, Y. Pei, M. Farber, M. Stewart, R. D. Levina, M. B.

Levin, Y. Xu, and R. L. Barbour, “Dynamic studies of small animals with a four-
color diffuse optical tomography imager,” Rev. Sci. Instrum. 76, 094302–094317
(2005).
BIBLIOGRAPHY 78

[84] D. Piao, S. Jiang, S. Srinivasan, H. Dehghani, and B. W. Pogue, “Video-rate near-

infrared optical tomography using spectrally encoded parallel light delivery,” Opt.
Lett. 30, 2593–2595 (2005).

[85] S. Gupta, P. K. Yalavarthy, D. Roy, D. Piao, and R. M. Vasu, “Singular Value

Decomposition based computationally efficient algorithm for rapid dynamic near-
infrared diffuse optical tomography,” Med. Phys. 36, 5559–5567 (2009).

[86] Z. Li, V. Krishnaswamy, S. Jiang, S. C. Davis, S. Srinivasan, K. D. Paulsen, and

B. W. Pogue, “Rapid magnetic resonance-guided near-infrared mapping to image
pulsatile hemoglobin in the breast,” Opt. Lett. 35, 3964–3966 (2010).

[87] T. Correia, A. Gibson, M. Schweiger, and J. Hebden, “Selection of regularization

parameter for optical topography,” J. Biomed. Opt. 14, 034–044 (2009)

[88] J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright, “Convergence Prop-

erties of the Nelder-Mead Simplex Method in Low Dimensions,” SIAM Journal of
Optimization 9, 112–147 (1998).

[89] Weblink: https://sites.google.com/site/sercmig/home/prakashprograms

[90] B. W. Pogue, S. C. Davis, X. Song, B. A. Brooksby, H. Dehghani, and K. D. Paulsen,

“Image analysis methods for diffuse optical tomography,” J. Biomed. Opt. 11, 33001
(2006).

[91] X. Song, B. W. Pogue, H. Dehghani, S. Jiang, K. D. Paulsen, and T. D. Toste-

son,“Receiver operating characteristic and location analysis of simulated near-
infrared tomography images,” J. Biomed. Opt. 1, 054013 (2007).

[92] H. Dehghani, M. E. Eames, P. K. Yalavarthy, S. C. Davis, S. Srinivasan, C. M.

Carpenter, B. W. Pogue, and K. D. Paulsen, “Near Infrared Optical Tomography
using NIRFAST: Algorithms for Numerical Model and Image Reconstruction Algo-
rithms,” Commun. Numer. Meth. Engng. 25, 711–732 (2009).
BIBLIOGRAPHY 79

[93] G. Golub and U. von Matt, “A Generalized cross-validation for large-scale prob-
lems,” J. Comput. Graph. Statist., 6, 1–34 (1997).

[94] C. C. Paige and M. A. Saunders, “LSQR: an algorithm for sparse linear equations
and sparse least squares ” ACM Trans. Math. Softw. 8, 43-71 (1982).

[95] C. C. Paige and M. A. Saunders, “LSQR sparse linear equations and least-squares
problems,” ACM Trans. Math. Softw. 8, 195-209 (1982).

[96] L. Hu, H. Wang, B. Zhao and W. Yang, “A hybrid reconstruction algorithm for
electrical impedance tomography”, Meas. Sci. Technol., 18, 813–818 (2007).

[97] G. M. Turner, G. Zacharakis, A. Soubret, J. Ripoll, V. Ntziachristos, “Complete-

angle projection diffuse optical tomography by use of early photons”, Opt. Lett. 30,
409–412 (2005).

[98] M. S. Zhdanov, Geophysical Inverse Theory and Regularization Problems 1st ed.
(Elsevier Science, 2002).

[99] C. Shaw and P. K. Yalavarthy, “Effective contrast recovery in rapid dynamic near-
infrared diffuse optical tomography using `1 -norm-based linear image reconstruction
method,” J. Biomed. Opt. 17, 086009 (2012).

[100] N. Nguyen, P. Milanfar, and G. Golub, “A computationally efficient superresolu-

tion image reconstruction algorithm,” IEEE Trans. Image Processing, 10, 573–583
(2001).

[101] P. C. Hansen, J. G. Nagy, and D. P. OLeary, “Deblurring Images: Matrices, Spec-

tra, and Filtering,” (Fundamentals of Algorithms, SIAM, Philadelphia, 2006)

[102] M. E. Kilmer, D. P. O’Leary, “Choosing regularization parameters in iterative

methods for ill-posed problems,” SIAM J. Matrix Anal. Appl, 22. 1204–1221 (2001).

[103] P. C. Hansen, “Regularization Tools Version 4.0 for Matlab 7.3,” Numerical Algo-
rithms 46, 189-194 (2007).
BIBLIOGRAPHY 80

[104] T. O. Mcbride, B. W. Pogue, S. Jiang, U. L. Osterberg, and K. D. Paulsen,

“A parallel-detection frequency-domain near-infrared tomography system for
hemoglobin imaging of the breast in vivo,” Rev. Sci. Instr. 72, 1817–1824 (2001).

[105] B. W. Pogue, K. D. Paulsen, H. Kaufman, and C. Abele, “Calibration of near

infrared frequency-domain tissue spectroscopy for absolute absorption coefficient
quantitation in neonatal head-simulating phantoms,” J. Biomed. Opt. 5, 182–193
(2000).

[106] B. W. Pogue and M. Patterson, “Review of Tissue Simulating Phantoms for Optical
Spectroscopy, Imaging and Dosimetry”, J. Biomed. Opt, 11, 041102, (2006).

[107] D. Piao, H. Dehghani, S. Jiang, S. Srinivasan, and B. W. Pogue, “Instrumenta-

tion for video-rate near-infrared diffuse optical tomography,” Rev. Sci. Instrum. 76,
124301–124313 (2005).

[108] S. R. Arridge and M. Schweiger, “Image reconstruction in optical tomography,”

Phil. Trans. R. Soc. Lond. B. 352, 717–726 (1997).

[109] D. A. Boas, K. Chen, D. Grebert, and M. A. Franceschini, “Improving the diffuse

optical imaging spatial resolution of the cerebral hemodynamic response to brain
activation in humans,” Opt. Letters 29, 1506–1508 (2004).

[110] B. Kanmani and R. M. Vasu, “Diffuse optical tomography using intensity measure-
ments and the a priori acquired regions of interest: theory and simulations,” Phys.
Med. Biol. 50, 247–264 (2005).

[111] P. K. Yalavarthy, B. W. Pogue, H. Dehghani, C. M. Carpenter, S. Jiang and K.

D. Paulsen, “Structural information within regularization matrices improves near
infrared diffuse optical tomography,” Opt. Exp. 15, 8043–8058 (2007).

[112] L. Borcea, “Electrical Impedance Tomography,” Inv. Problems 18, R99–R136

(2002).
BIBLIOGRAPHY 81

[113] M. Soleimani, P. K. Yalavarthy, and H. Dehghani, “Helmholtz-type regulariza-

tion method for permittivity reconstruction using experimental phantom data of
electrical capacitance tomography,” IEEE Trans. Instrum. Meas. 59, 78–83 (2010).

[114] S. C. Davis, H. Dehghani, J. Wang, S. Jiang, B. W. Pogue, and K. D. Paulsen,

“Image-guided diffuse optical fluorescence tomography implemented with laplacian-
type regularization,” Opt. Express 15, 4066–4082, (2007).

[115] S. Ahn, A. J. Chaudhari, F. Darvas, C. A. Bouman, and R. M. Leahy, “Fast

iterative image reconstruction methods for fully 3D multispectral bioluminescence
tomography,” Phys. Med. Biol. 53, 3921–3942 (2008).

[116] Jaya Prakash, Venkittarayan Chandrasekharan, Vishwajith Upendra, and Pha-

neendra K. Yalavarthy, “Accelerating frequency-domain diffuse optical tomographic
image reconstruction using graphics processing units,” J. Biomed. Opt. 15, 066009
(2010).