University of Southern California Dissertations and Theses

Pattern detection in medical imaging: pathology specific imaging contrast, features, and statistical models

(USC Thesis Other)

Pattern detection in medical imaging: pathology specific imaging contrast, features, and statistical models

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Pattern Detection in Medical Imaging: Pathology Specic Imaging
Contrast, Features, and Statistical Models
by
Sinchai Tsao
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements of the Degree
DOCTOR OF PHILOSOPHY
(Biomedical Engineering)
December 2013
Copyright 2013 Sinchai Tsao
Dedicated to the memory of Manbir Singh, PhD
ii
Born Oct. 7, 1945 in Amritsar, India, Manbir Singh earned his Ph.D. in physics
from UCLA in 1971, before conducting his postdoctoral studies in biomedical physics at
UCLAs Laboratory of Nuclear Medicine and Radiation Biology.
He later spent one year as a Visiting Scholar of the American Heart Association at the
Mayo Clinic in Rochester, Minn., where he conducted the rst studies in single-photon
emission computed tomography (SPECT). He joined the Department of Radiology at
USC in 1977 and received an appointment in biomedical engineering in 1988.
Singh pioneered the use of SPECT to detect and quantify acute myocardial infarctions
in three dimensions and was one of the rst investigators to demonstrate the synergism
of CT scans and nuclear medicine SPECT imaging in detecting and visualizing both the
anatomy and function of the heart.
He was the nationally elected AdCom representative for Nuclear Medical Sciences
in the IEEE Nuclear and Plasma Society (NPSS) from 1986 to 1989, co-founder of the
IEEE Medical Imaging Conference in 1990, the technical chair for nuclear medical sciences
within IEEE NPSS from 1991 to 1993 and scientic program chair of the 1992 and 1993
IEEE Medical Imaging conferences.
His latest interests were in functional magnetic resonance imaging (MRI) and diusion
tensor imaging (DTI) tractography with applications of DTI to Alzheimers disease and
traumatic brain injuries. He was founder and director of the neuroimaging core at USC,
which develops new methodology in functional MRI and DTI.
Singh also was founding director of the graduate program in biomedical imaging in
the Department of Biomedical Engineering.
Singh passed away due to unknown causes while visiting family in India in the winter
of 2011. He is survived by his wife, Heidi and their son, Kabir.
iii
Acknowledgements
I would like to rst and foremost thank my loving wife, Kori, who has been with
me through this long process and supported me through it all. I denitely couldn't
have done it without you! Thank you for being there through all the late nights, lab
changes and grumpiness of failed experiments. You've sacriced so much for this volume
of work. It belongs as much to you as much it is to me. To my parents, Winnie and
Fred, who raised me to think independently and not to be swayed by other's opinions.
My darling sister, Sonia, who continues to inspire me daily with her courage and never
ceasing enthusiasm for life. My in-laws, Sharon and Rodney, who generously shared their
academic journeys and who was always there to listen to my research problems. I would
like to especially recognize the members of my qualifying and defense committee members.
Your invaluable and thoughtful feedback and evaluation has made me the researcher I am
today. For my qualifying examination: Krishna Nayak (Chair), Vasilis Marmarelis, Walter
Wolf, Richard Leahy and Arkadiusz Gertych. For my defense: Natasha Lepore (Chair),
Norberto Grzywacz, Vasilis Marmarelis, Walter Wolf and Arkadiusz Gertych. Behind the
scene, we have Mischalgrace Diasanta, our under appreciated graduate student advisor
and my friend. I believe a piece of this thesis belongs to every one of you. Your support
through the years have been incredible. I truly could not have done this without you.
This thesis is dedicated to Manbir Singh, PhD. My mentor, advisor and friend who
passed away in 2011. Although many people have mentored and helped me through my
PhD training, Dr. Singh is denitely the most in
uential person in my academic life.
Dr. Singh was my PhD advisor from 2008-2011 but the bulk of the work of this PhD
is a continuation of his research program and legacy. I started my life at USC learning
Medical Imaging from him through his classes: BME 425 & 525. His lab was full at the
time and it was not until 2008 did I have the opportunity to join his group. I believe that
this fortuitous event was one of the best things that have ever happened to me. I would
like to thank him for his dedication to his many students including myself. His legacy
lives within this body of work as well as any other work I do in the future. Thank you
Dr. Singh! Mrs Heidi Singh and Kabir, Thank you for sharing your husband / father
with us, this would not have been possible without you or him.
I would also like to thank my many mentors and I've been so lucky and grateful to
have been able to learn from all of you (in no particular order):
The inspirational Natasha Lepore who adopted me when I was advisor-less and got
me through the tough last year of my PhD. I will forever be grateful for your mentorship
and how you helped me set the course of my career. Arek Gertych my constant mentor of
iv
throughout my PhD who taught me everything I know about up and downs of research
and how to do research with integrity. Meng Law who provide us guidance and support
during our toughest times. Jorge Documet who was my coding partner and teacher, you
truly inspired me with your ability to write elegant code. Dr. Helena Chui, who was
my perfect role model for a brilliant yet dedicated and humble medical researcher. Dr.
Walter Wolf, you encapsulate the researcher and person that I want to be. Thank you
for mentoring me and sharing your life experiences with me. Krishna Nayak who inspired
us with his passion for learning and drive for research. Norberto Gzywacz for all your
inspirational leadership of the BME department as well as your care and support through
our loss of Manbir Singh. I would not be here without Michael Khoo who as chair of
the BME department gave me the opportunity to prove myself as a researcher and never
stop believing in me. Tanja Adam the brilliant neurobiologist who directed me through
the pediatric obesity portion of this work. Katie Page for the opportunity to work on the
fructose vs glucose studies. Vidya Rajagopalan, I don't know what I would have done
had our paths not crossed. Not only did you support me through my last year of my PhD
but have been instrumental in getting me started with a new chapter of my academic life.
Brent Liu and Bernie Hwang who guided me through my rst years of my PhD. Alexander
Lin, for sharing his journey to faculty with me and his many words of encouragement and
guidance throughout the years. Ihab Hajjar, thank you for your support and being so
generous to me. I can't wait to see the results of your clinical trial. Tara Chklovski for
giving the opportunity to work with the kids at Iridescent and supporting my research.
Yiqiao Song and Sharron Lin, thank you so much for all your support over the years.
From the early days giving me advice over the phone to visiting you on the east coast. I'm
truly grateful for your encouragement and generosity. Karen Johnson, I really couldn't
have done this without you! The BME department at USC is so lucky to have you.
Thank you for your support through the ups and downs and especially when we loss
Dr. Singh. All the Professors who have guided me: Harry Hu, David D'Argenio, Ellis
Meng, Jesse Yen, Kirk Shung, Gerald Loeb, Ted Burger, Francis Richmond, Bob Scholtz,
Jay Kuo, Urbashi Mitra, Bosco Tjan, T.K. Hsiai, Bartlett Mel, Massoud Ghyam-Khah,
Manouchehr Moradi, Kathleen Allen, Heinz Lemke, Jianguo Zhang, Maria Law, Stephan
Buml, Anita Zimmerman, Sharon Swartz, Christine Kearney, Anubhav Tripathi, Michael
Lysaght, John Stein, Robert Pelcovits, Chung-I Tan, James Valles, Sean Ling, Thomas
Bancho, Jonathan Waage, Toby Cumberbatch, ...
My many many (lab) mates current and past! But especially ... Darryl Hwang, it's
been a pleasure working with you in the lab, hanging out and learning from you. I have
v
always and forever will be been impressed by your love for your work and will miss those
long days working hard and then playing hard. Niharika Gajawelli, it has been a pleasure
working with you. I believe you will make an amazing researcher and I stand by it. Bryce
Wilkins, thank you for coming down this long and dicult road with me. I know I will
continue to learn from you for many years to come. Sam Ma, I'm so excited for you.
Thank you so much for working with me. Nam (Namgyun Lee), can't wait until you
get back to the US after your military service and we can work together again. Yi Lao,
I'm so lucky to have so many inspiring lab mates you are denitely one of the top ones.
Fernando Yepes, wish we had more time to work together in Los Angeles. It has been
a great honor to have been able to work with you. You've been a great inspiration and
will continue to be for many years to come. Kevin Ma, my housemate, lab mate and
my friend. This volume of work in many ways document our learning journey together
at USC. You've contributed to it in so many ways. Anika Joseph, couldn't have done
it without your energy and companionship in this crazy journey to a doctorate. Anh
Le, I have learned so much from you. A lot this work reminds me of our rst years as
a PhD student taking classes and learning to be strong researchers. I thank you for all
the things you have taught me over the years. Witaya Sungkarat. P'Wit, I believe more
than half the things I know I have learned from you. I continue the legacy that you've
left behind at USC. You'll always be part of who I am as a researcher. David Woollard,
the committee loved all the suggestions you made for my qualifying examination. As
always I am ever impressed by your work and your passion. As I put my defense together,
I can see how I've applied what you've taught me! Thank you! Mark Haney, I wish
we have more chances to work together. I will never forget your drive, determination
and incredible spirit. Lauren Brown, it's been a pleasure working with you. I would
really like the chance to work together in the future. I'm so excited for you starting your
interdisciplinary work in Gerontology. Others including but limited to Jose Iriarte-Diaz,
Kristin Bishop, Kevin M. Middleton, ...
To the rest of my family and friends: Erick Soedjasa, thank you for your friendship
and support. My Tsao cousins Allegra and Spencer, you have both inspired me and more
ways than you know! My Lin cousins Charles, Stephanie, Michelle, Felicia, Andre, Jeremy
and Judith, thank you for loving support. In particular, Andre thanks for sharing your
USC college life with me. My uncle and aunts, Mary Ann, Bennett, Calvin, David, Zack,
Cheryd, Alfred, Liz, Michael, Winnie, John, Wendy and last but not least Rebecca. This
work is full of your love and support. To my grandparents, this would not have been
vi
possible if you had not braved the turmoils of our home country to give us all a better
life.
And so many others that have contributed in their own unique way to this volume of
work: [Again, in no particular order]
Bing G.
Lucy Z
Jasper L
Sharon J
Paul B
Alan S
Talia BY
Hugh P
Marvin N
Mary BN
Sam Valencerina
Vincent C
Brian Lim
Josh B
Travis P
Mandy L
Pirran T
Kenny K
Jin Ho O
Devyani N
Alice C
Shruthi N
Shivanth B
Ming S
Vanessa G
Alexis W
Wenli W
Angelica D
Brian Li
Caroline W
Dong S
Heidi G
Geo T
Harry S
Guru S
Jeerey C
Jay F
Leslie K
Melissa F
Maya C
Phillip H
Sadaf S
Rui Z
Sam Lin
Sang C
Sam Cunningham
Ude L
Peggy H
Je T
Kathryn D
Dan T
Patrick H
Jon C
Hsu-Lei Lee
Syed A
Wesley Z
Mahender M
Samir S
Chris Sandino
Wayne C
Davi L
Ahsan J
Yi Guo
Eamon D
Terrence J
Vanessa L
Hung D
Brian W
Yinghua Z
Yoon K
Ida A
Matt B
Ronalee L
Jay M
Manish K
Gabriela MO
Flavia O
Andrew W
Jae Youn H
Man N
Nadav I
Fei Yu
Chris N
Helene N
David F
Dennis M
Lauren H
Aneeta J
Jamuna M
Xenia Z
Yue Kang
Joseph M
Chi Hyung (JJ) Seo
Winston T
Djordje P
Jarree C
Limei C
Jonathan L
Susan C
Chaithanya R
Rosa C
Adam W
Leon C
Christian G
Vivek P
Rachelle B
Leanne C
Arthi S
Kelsey S
Orian L
Ami S
Bryce C
Debra F
Thad P
Mary H
Tim Chklovski
Julia Castro
Luz Rivas
Jenna Blanton
Mariana R
Paul Yarin
Kara Christianson
Lindsey JS
Erika A
Jennifer Hsu
Srinivas Yerramalli
All the Kids at Iridescent!!
vii
Table of Contents
Dedication ii
Acknowledgements iv
List of Figures xiv
List of Tables xix
Acronyms xxi
1 Introduction 1
1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 Part A: Imaging Informatics . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Part B & C: Functional and Structural Neuroimaging . . . . . . . 4
2.3 Part D: Statistical Analysis and Machine Learning . . . . . . . . . 5
3 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
A Imaging Informatics 9
2 Managing Healthcare Information Using Short Message Service (SMS)
in Wireless Broadband Networks 10
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Standard Work
ow . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 SMS System Architecture vis-a-vis PDA System Architecture . . . 18
4.2 System Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 System Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 User Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.5 Interface Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
viii
4.6 System Component and Operation Costs . . . . . . . . . . . . . . 21
5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 RadSearch: A RIS/PACS Integrated Query Tool 24
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 RadSearch Toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Natural Language Processing (NLP) -based Approaches . . . . . . 27
3.2 Lexicons and Synonymity . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 RIS/PACS Integration . . . . . . . . . . . . . . . . . . . . . . . . . 29
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Radiology Text Report Parsing . . . . . . . . . . . . . . . . . . . . 29
4.3 Forming Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.4 Integration with PACS and as a Study Management Tool (SMT) . 30
4.5 System Operation Considerations . . . . . . . . . . . . . . . . . . . 31
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1 RadSearch versus other RIS Data Mining Approaches . . . . . . . 31
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
B Functional Neuroimaging 33
4 Application of fMRI to Obesity Research: Dierences in Reward Path-
way Activation measured with fMRI BOLD during Visual Presentation
of High and Low Calorie Foods 34
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3 MRI Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4 SPM Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
ix
5 Novel MRI Methodology to detect Human Whole-Brain Connectivity
Changes after Ingestion of Fructose or Glucose 46
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Protocol Development . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Data Processing and Analysis of fMRI Baseline Trends . . . . . . . 51
3.4 Connectivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 51
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
C Structural Neuroimaging 64
6 Quantication of Fornix Tracts in MCI and AD 65
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7 Comparison of Limbic Regions FA Using Tractography-dened ROIs in
AD and MCI 71
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3 Results and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
8 CSF Contamination Correction in DTI Tractography of the Fornix in
Elderly Subjects 76
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
9 ICA-based Multi-Fiber DWI Tractography in Neurosurgical Planning 82
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2 Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3 Independent Component Analysis . . . . . . . . . . . . . . . . . . 86
3.4 Tractography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
x
5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
10 The Power of Hybrid / Fusion Imaging Metrics in Future PACS Sys-
tems: A Case Study into the White Matter Hyperintensity Prenumbra
using FLAIR and Diusion MR 93
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
11 FreeSurfer Parcellation of Brains Containing Large Infarcts 99
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
D Statistical Analysis and Machine Learning 113
12 Is Greulich and Pyle Atlas still a Good Reference for Bone Age Assess-
ment? 114
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.2 Data Visualization - Web GUI (Graphical User Interface) Develop-
ment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.3 Prototype of a Clinically Applicable GUI . . . . . . . . . . . . . . 123
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1 Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.2 Statistical analysis between races . . . . . . . . . . . . . . . . . . . 125
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
13 Automated Bone Age Assessment of Older Children using the Radius 129
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2.1 Current Issues in Bone Age Assessment Methods . . . . . . . . . . 131
2.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
xi
3.1 Localizing the Radius . . . . . . . . . . . . . . . . . . . . . . . . . 132
3.2 Features Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 134
3.3 Fuzzy Classication and Computation of Bone Age . . . . . . . . . 135
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
14 Evaluating the Predictive Power of Multivariate Tensor-based Mor-
phometry in Alzheimers Disease Progression via Convex Fused Sparse
Group Lasso 143
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
3 convex Fused Sparse Group Lasso (cFSGL) . . . . . . . . . . . . . . . . . 146
4 Multivariate Tensor-based Morphometry (mTBM) features . . . . . . . . 149
5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
15 Mapping of ApoE4 Related White Matter Damage using Diusion MRI152
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
16 Insulin sensitivity and brain reward activation in overweight Hispanic
girls: a pilot study 158
1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.2 Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.3 Anthropometric measures . . . . . . . . . . . . . . . . . . . . . . . 163
3.4 Body Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
3.5 Measures of Insulin Sensitivity . . . . . . . . . . . . . . . . . . . . 164
3.6 Analysis of blood parameters . . . . . . . . . . . . . . . . . . . . . 164
3.7 Food picture exposure paradigm . . . . . . . . . . . . . . . . . . . 164
3.8 Imaging and statistical analysis . . . . . . . . . . . . . . . . . . . . 165
4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.1 Brain responses to visual stimulation and SI . . . . . . . . . . . . . 167
4.2 Brain responses and fasting insulin . . . . . . . . . . . . . . . . . . 168
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
xii
17 Conclusion 175
1 Bone Age Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
2 Obesity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3 Aging and Dementias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4 Generating Meaningful Imaging Contrast . . . . . . . . . . . . . . . . . . 178
5 Pattern Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
References 185
xiii
List of Figures
1.1 Organizational chart for this thesis . . . . . . . . . . . . . . . . . . . . . 7
2.1 Component Interaction Diagram . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The System State diagram shows how the system keeps track of the users
selections. This state tracking system allows the user to use the simplistic
SMS text-based interface to select the study of interest from the patient
of interest to be sent to a designated DICOM node. . . . . . . . . . . . . 18
2.3 The SMS System Architecture is a powerful extension of the existing PDA
System Architecture, allowing any authorized user with a cellular phone
to direct DICOM images within the Hospital network. . . . . . . . . . . 19
3.1 Prototype Database Schema . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Overall System Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.1 fMRI visual stimulus presentation paradigm as presented through the
head coil mirror to a projection screen at the foot of the subject. All
images were based on previous studies or through the International Af-
fective Picture System. The rst three box car functions below the images
refer to the three stimulus reference functions for NF/Control, HC and
LC respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Group statistical parametric map (SPM) comparison of activated areas
during presentation of Low Calorie, High Calorie and Control Images
(P < 0:01, uncorrected) with an extent threshold of 10 Voxels. The ROIs
are rendered on to a template single subject's brain surface. . . . . . . 45
5.1 Scan Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 MRI Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.3 Demeaned Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.4 Default Mode Network before ingestion . . . . . . . . . . . . . . . . . . . 56
5.5 Default Mode Network post ingestion . . . . . . . . . . . . . . . . . . . . 57
5.6 Hypothalamic Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.7 Connectivity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.1 Normalized Fornix Tracts . . . . . . . . . . . . . . . . . . . . . . . . . . 66
xiv
6.2 ICV Normalized Freesurfer-based Hippocampus Volumes in Normal, MCI
and Probable AD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.3 ICV Normalized Tract Counts in Normal, MCI and Probable AD . . . . 68
6.4 Hippocampus Volumes and Tract Count Eect Sizes . . . . . . . . . . . 69
7.1 Fornix FA t-Score Comparison. MCI> AD (top row) NC> MCI (bottom
row). Colors denote the t-score dierence between the mean FA of the
various populations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.2 Cingulum FA t-Score Comparison. MCI >AD (top row) NC > MCI
(bottom row) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
8.1 Corrected Primary Eigenvectors after CSF compensation correction in
the anterior portion of the Fornix (Red Original Eigenvector) (Green
Corrected Eigenvector) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
8.2 Corrected Tractography (RED) and uncorrected Tractography (BLUE)
at FA threshold of 0.14 of tracts of the Hippocampus (green ROI). Figure
shows (1) correction recovering the superior portion of the fornix (2)
reduction of erroneous tract. . . . . . . . . . . . . . . . . . . . . . . . . . 81
9.1 Patient 1 CSTs using multi-ber ICA-based tractography with coronal
slice showing color FA; without tumor volume (L) and with tumor volume
(R). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2 Patient 1 CSTs using single-ber DTI tractography with coronal slice
showing color FA; without tumor volume (L) and with tumor volume (R) 90
9.3 Patient 2 CSTs using multi-ber ICA-based tractography with coronal
slice showing color FA; without tumor volume (L) and with tumor volume
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
9.4 Patient 2 CSTs using single-ber DTI tractography with coronal slice
showing color FA; without tumor volume (L) and with tumor volume (R) 91
10.1 Data for subject with very high WMH load. (A) FA image for WMH
candidate region mask. (B) Histogram tting for dynamic thresholding.
(C) WMH image (Red). (D) MD and FA histograms in WMH regions
vs. all WM regions.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
10.2 MD and FA histograms for low WMH load and high WMH load subjects
compared with MD and FA in all WM for each subject. . . . . . . . . . 97
11.1 Diagram showing the process for lling in the brain infarct and the output
after each step. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
11.2 FreeSurfer parcellation result for validation study. (a) T1 slices showing
simulated infarct; (b) Ground truth FreeSurfer parcellation; (c) FreeSurfer
parcellation before lling simulated stroke; (d) FreeSurfer parcellation af-
ter lling. Circles highlight regions where signicant distortions have
been corrected after lling. . . . . . . . . . . . . . . . . . . . . . . . . . . 110
xv
11.3 Ratios of various volumes of the FreeSurfer parcellation of the simulated
infarct with respect to the ground truth. (a) Averaged volumes ratios
of the before and after lling of cortical and subcortical regions in one
simulated lesion using 5 dierent normal brain volumes. The red curve
is the lled, the blue the unlled and the pink the lled and smoothed.
The red and pink plots are comparable. The gures on the lower rows
are the Histograms and Gaussian ts of volume ratios of the (b) original
to the lled brain, (c) original to the lled and smoothed brain, and (d)
original to the unlled brain. . . . . . . . . . . . . . . . . . . . . . . . . 111
11.4 Comparison of ratios of various volumes in the FreeSurfer parcellation
after lling in with the (a) SPM based method and (b) ANTS based
method. The results are quite similar with both showing a mean close to
1 and similar standard deviation. . . . . . . . . . . . . . . . . . . . . . . 111
11.5 Comparison of FreeSurfer parcellation before and after lling in the stroke
using SPM based approach. Dotted yellow circle indicates regions where
infarct was present. Distal regions as well as proximal regions are observed
to be aected signicantly. . . . . . . . . . . . . . . . . . . . . . . . . . . 112
12.1 Bone age assessment procedure in clinical practice. . . . . . . . . . . . . 117
12.2 Sample Hand Image from the CHLA data collection process. All pertinent
bones that were used as ROIs in the CAD program are labeled. . . . . . 118
12.3 Simple GUI to visualize dierences between the dierent Study Groups. 119
12.4 Clinical GUI to compare a patients CAD and Radiologists Readings rel-
ative to the accuracy of the Radiologists Readings from the Digital Hand
Atlas Database. The three black horizontal lines bisected by the vertical
line marks the mean as well as the upper and lower standard deviation of
the pediatric radiologists G&P-based readings of normal cases with that
specic chronological age. The yellow marker represents the radiologist
reading of the current case. The red marker represents the CAD output
of the current case. The GUI allows the user to assess the CAD and
Pediatric Radiologists readings against the accuracy of the gold standard
- the accuracy of G&P readings of the normal cases in the database. Any
reading variation away from the chronological age that is greater than the
accuracy as set by the gold standard indicates a possible growth disorder. 120
12.5 The gure illustrates that gender eects bone age assessment with both
reading and CAD methods. (a) Caucasian male; (b) Caucasian female.
The x-axis is chronological age and the y-axis is either radiologists reading
or CAD result. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
xvi
12.6 The gure illustrates that ethnic or racial origin eects bone age assess-
ment with both the reading and CAD methods. (a) Asian; (b) Caucasian.
The x-axis is chronological age and the y-axis is either radiologists read-
ing or CAD result. The x-axis is chronological age and the y-axis is
radiologists reading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
12.7 The divided age sets in evaluation. (a) for female and (b) for male. . . . 122
12.8 Plots of readings between race pairs (a) Asian female and Caucasian
female; (b) Hispanic female and African American female. The x-axis is
chronological age and the y-axis is radiologists reading. . . . . . . . . . . 123
13.1 Outline of Image Processing algorithm to extract Growth Plate ROI from
the larger Wrist ROI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
13.2 Enlarge Image showing the extracted Growth Plate ROI. Note the hor-
izontal pattern that denotes the boundary between the epiphysis and
metaphysis. This boundary lessens and eventual disappears as the child
grows. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
13.3 Mamdani Fuzzy Inference System used to aggregate the feature values
and derive a Bone Age based on feature values. The system requires
training with a data set of normal children. . . . . . . . . . . . . . . . . 135
13.4 Feature values for the male cohort. . . . . . . . . . . . . . . . . . . . . . 136
13.5 Feature values for the female cohort. . . . . . . . . . . . . . . . . . . . . 136
13.6 CAD Age versus Pediatric Radiologist's Reading for the Male Cohort . . 137
13.7 CAD Age versus Chronological Age for the Male Cohort . . . . . . . . . 138
13.8 CAD Age versus Pediatric Radiologist's Reading for the Female Cohort 139
13.9 CAD Age versus Chronological Age for the Female Cohort . . . . . . . . 140
14.1 Bar Chart of the rMSE of predictions with and without mTBM features
by time point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
14.2 Results without mTBM Features - Scatter plots of Predictions versus
Actual ADAS Cognition Scores at M06 through M48 . . . . . . . . . . . 151
14.3 Results with mTBM Features - Scatter plots of Predictions versus Actual
ADAS Cognition Scores at M06 through M48 . . . . . . . . . . . . . . . 151
15.1 Demographic Distribution of the Subjects . . . . . . . . . . . . . . . . . 155
15.2 SPM of ParaHippocampal FA values with ApoE-4 < no ApoE-4 with
P
uncorrected
< 0:01 and extent threshold < 10 voxels (all other covariates
were set as nuisance variables and controlled) . . . . . . . . . . . . . . . 156
15.3 SPM of t-scores of the Posterior Cingulum (left) and Hippocampus (right)
using FA values for ApoE4 < no ApoE4 (Red), Age (Green) and
CDR (Blue) with P
uncorrected
< 0:05 and extent threshold < 20 voxels. . 157
xvii
16.1 Association of insulin sensitivity (SI), insula activation (parameter esti-
mates, PE) (Figure 1A (top): r2 = 0.69; p < 0.005) and left anterior
cingulate activation (Figure 1B (bottom): r
2
= 0.65; p < 0.005), con-
trasting cerebral activation in response to high calorie visual stimuli >
non-food visual stimuli. . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
16.2 Association of Insulin Sensitivity, putamen activation (parameter esti-
mates, PE) (Figure 2A:r
2
= 0.74; p< 0.005) and insula activation (Fig-
ure 2B:r
2
= 0.66; p< 0.005), contrasting cerebral activation in response
to high calorie visual stimuli > low calorie visual stimuli. . . . . . . . . . 170
16.3 Association of fasting insulin concentrations, caudate activation (parame-
ter estimates, PE) (Figure 3A:r
2
= 0.70; p< 0.005) and insula activation
(Figure 3B: r
2
= 0.69; p < 0.005), contrasting cerebral activation in re-
sponse to high calorie visual stimuli > non-food visual stimuli. . . . . . 171
xviii
List of Tables
2.1 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Sample User - Server Interaction: Send Patient Doe, Joe (Patient ID:
X8787224)s Chest CT from February 1998 to EFILM workstation in Exam
Room 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1 Results when comparing High Calorie > Control ordered by z score with
Talairach Atlas Coordinates and the Automatic Labeling of the ROIs was
done using AAL toolkit for SPM [1] . . . . . . . . . . . . . . . . . . . . . 39
4.2 Results when comparing Low Calorie > Control ordered by z score with
Talairach Atlas Coordinates and the Automatic Labeling of the ROIs was
done using AAL toolkit for SPM [1] . . . . . . . . . . . . . . . . . . . . . 40
4.3 Results when comparing High Calorie > Low Calorie ordered by z score
with Talairach Atlas Coordinates and the Automatic Labeling of the ROIs
was done using AAL toolkit for SPM [1] . . . . . . . . . . . . . . . . . . 44
5.1 t-scores corresponding to the pair-wise ROIs shown in Figure 5.7 that ex-
ceed the threshold (t > 2). . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.1 Eect Sizes and t-test of Normalized Hippocampus Volume and Normalized
Tract Count against Normal Population . . . . . . . . . . . . . . . . . . . 70
8.1 Corrected FA scores of 4 Normal age matched elderly subjects . . . . . . . 80
9.1 DWI Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
9.2 Tractography Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
10.1 Summary of Subject Demographics . . . . . . . . . . . . . . . . . . . . . . 95
12.1 Total of 1,103 cases from rst Cycle Data Collection . . . . . . . . . . . . 127
12.2 CAD evaluation for eight categories by race and gender. Values in the table
represent the mean dierence between chronological age and the average
reading in year. Signicant mean dierence is indicated by an asterisk. . . 128
12.3 Cross racial comparison of chronological age vs. reading for female. Each
block was divided into four subblocks, each of which encompasses one age
set. (unit: year, -: over read) . . . . . . . . . . . . . . . . . . . . . . . . . 128
xix
13.1 Mean Absolute Error of the system for the Male Cohort using Phalanges
and Radius based on the Chronological Age and Reading of the Pediatric
Radiologist as the gold standards. The absolute mean dierence for the
male cohort between the Pediatric Radiologists reading and the normal
childs chronological age is 0.650 years. . . . . . . . . . . . . . . . . . . . . 137
13.2 Mean Absolute Error of the system for the Female Cohort using Phalanges
and Radius based on the Chronological Age and Reading of the Pediatric
Radiologist as the gold standards. The absolute mean dierence for the
female cohort between the Pediatric Radiologists reading and the normal
childs chronological age is 0.700 years. . . . . . . . . . . . . . . . . . . . . 138
14.1 Results showing rMSE and R for M06 through M48 with and without
mTBM features. (* P < 10
3
) . . . . . . . . . . . . . . . . . . . . . . . . 148
16.1 Anthropometric measures describing mean standard error of the means
(SE) and range (n = 10). . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
16.2 Comparison of regional brain activation in response to high calorie food >
non food stimuli in 10 overweight Hispanic girls. . . . . . . . . . . . . . . 168
xx
Acronyms
AD Alzheimer's Disease. xxi, 5, 8, 9
BOLD Blood Oxygen Level Dependent. 10, 183
CSF Cerebrospinal Fluid. 5, 8
CST Corticospinal Tract. 9
CT Computed Tomography. 5
DTI Diusion Tensor Imaging. 8, 12
EBM evidence-based medicine. xxi, 1, 7
EMR Electronic Medical Record. 3
FLAIR
uid attenuated inversion recovery. 12
fMRI functional magnetic resonance imaging. xxi, 4, 12
HRF Hemodynamic Response Function. 10, 183
MANN massive articial neural network. 185
MCI Mild Cognitive Impairment. 5
MRI magnetic resonance imaging. 5, 8, 12
PACS Picture Archival and Communication System. xxi, 3
QEBM quantitative evidence-based medicine. 7
RIS Radiology Information System. 3
SMS Short Message Service. xxi, 3
SVM Support Vector Machine. 185
WMH white matter hyperintensity. 12
xxi
Abstract
T
he motivation for this work is a vision of widespread adoption of a priori quantita-
tive epidemiological information for clinical decision-making, and can be seen as a
quantitative large-scale extension of evidence-based medicine (EBM) [2]. Medical images
can be seen as a spatially encoded map of physiological measurements that can be used to
predict prognosis and to drive treatment plans. This paradigm can be very powerful and
is driven by the recent big data revolution in computer science as well as the increasing
availability of medical imaging modalities due to decreases in manufacturing costs. In
order to achieve this overarching goal, three practical requirements must be reached and
correspond to the parts of this thesis: Part A: Developing IT infrastructure and technol-
ogy that enables the dataset to be properly collected and organized for analysis. Part B
& C: Generation of functional (Part B) and structural (Part C) medical imaging contrast
that are optimized for analysis. Part D: Pattern recognition techniques (including both
image processing and machine learning techniques) to mine information from the large
imaging datasets generated. As part of the thesis, I discuss my contribution to IT infras-
tructure (Part A) by developing a Short Message Service (SMS)-based system to control
the clinically used Picture Archival and Communication System (PACS) (Ch.2) as well
as an imaging study tool that categorizes patient imaging data for use in retrospective
studies(Ch.3). I then go on to detail my work with functional neuroimaging of obesity
using functional magnetic resonance imaging (fMRI)(Ch.4) and (Ch.5). Chapters 6-9 de-
tails my eorts at studying abnormal aging versus normal aging using diusion MRI as
well as applications of diusion MRI to surgical planning. Chapters 10 discusses my work
integrating diusion MR with FLAIR MRI to investigate the properties of white matter
lesions and how it can be used in the clinical setting. Chapter 11 then moves on to talk
about my work modifying standard brain parcellation techniques to allow them to work
with aged brains with large infarcts. Chapters 6-11 altogether represent my eorts in
structural neuroimaging using MRI (Part C). The thesis then closes with capstone work
in development staging using hand x-rays using fuzzy logic (Ch. 12 & 13). To close the
work with Alzheimer's Disease (AD) and aging, we used machine learning techniques to
predict disease progression based on a baseline MRI scan as well as higher order analysis of
our diusion MRI dataset by integrating MRI information with other clinical information
such as neuropsychological tests, cardiovascular status. This is all in an eort to compu-
tationally explore the relationship between MRI measurements and clinical presentation
of disease as measured by neuropsychological scores. Similarly with the Obesity work, we
related fMRI activation dierences between high and low calorie foods with non-imaging
information such as insulin resistance (Ch. 16).
xxii
Chapter 1
Introduction
Real knowledge is knowing what you know and what you don't know. Confucius
1 Motivation
T
he motivation for this work is a vision of widespread adoption of a priori quantita-
tive epidemiological information for clinical decision-making, and can be seen as
a quantitative large-scale extension of evidence-based medicine [2]. Medical Images can
be thought of a structured multi-dimensional form of medical data. The structure comes
from the fact that each voxel or image matrix value is related spatially to neighboring
values and its location in the image matrix has anatomical meaning. Pattern detection /
statistical analysis using this information can exploit not only the measurements or matrix
1
values but also their patterns related to matrix indices (i.e. in the hippocampus, certain
cortex region, etc) to predict prognosis and to drive treatment plans. This paradigm can
be very powerful and is driven by the recent big data revolution in computer science: As
data storage prices have plummeted, databases sizes have exponentially increased and
high computational power have become increasingly ubiquitous. Meanwhile, there have
been signicant advances in improving the availability of medical imaging modalities as
clinical diagnostic tools. This is mainly due to decreases in manufacturing costs, but also
because physicians have found them to be increasingly useful as screening tools. However,
in order to take advantage of this unique increase in medical data, we also require a rev-
olution in analysis techniques that can handle high dimensionality, multiple comparison
issues as well as the unique challenges that come with medical information and medical
imaging datasets (imaging artifacts, sampling issues, etc).
In order to achieve this overarching goal, three practical requirements must be reached
and correspond to the parts of this thesis:
1. Part A: Development of IT infrastructure and technologies that enables imaging
data to be properly collected and organized for analysis.
2. Part B & C: Generation of functional (Part B) and structural (Part C) medical
imaging contrast that are optimized for analysis.
3. Part D: Pattern recognition techniques (including both image processing and ma-
chine learning techniques) to mine information from the large imaging datasets
generated.
The overarching vision of this thesis is to contribute to building advanced computa-
tional techniques that leverage a priori information to characterize disease progression
2
and to support clinical decision-making. This is important not only in the context of
providing benchmarks for developing treatments but provides a way to understand basic
biological processes.
2 Overview
2.1 Part A: Imaging Informatics
Before being able to leverage advanced computational techniques, we have to address
the basic problem of getting information (in this case medical images) properly classied
and organized. To this end, in chapter 2, we developed a SMS-based routing system
to remotely manage the PACS database. In terms of image classication, the informa-
tion required currently lives on a number of separate databases in the clinic, including
the PACS, Radiology Information System (RIS) and increasingly the Electronic Medi-
cal Record (EMR). Important image meta-information such as the radiologist's report
as well as patient demographic information lives on the RIS database and the medical
images are stored in the PACS. The two systems are often from dierent vendors with
limited integration. It is, therefore, dicult to classify images for retrospective studies
and data mining. To compound this problem further, epidemiological information on the
RIS is unstructured and exists in a pure text form as a radiologist's report. To this end, I
developed a prototype system in chapter 3 that classies the medical images in the PACS
using extracted structured information from the RIS.
3
2.2 Part B & C: Functional and Structural Neuroimaging
Next we address the quality of the information. In medical imaging, this means generating
high signal to noise ratio (SNR) / low noise images with meaningful contrast. There are
two overall types of magnetic resonance imaging (MRI) contrast: Contrast based on
structure and those based on function. In structural MRI, the system recreates image
matrices that are representations of the object / person. We can take advantage of physical
properties of the magnetic dipoles in the body and alter the pulse sequence parameters
to change how contrast is generated. Functional techniques can be seen as an extension
of structural imaging which require the subject to perform a certain task. The dierences
in images acquired during dierent tasks versus control yield the contrast desired.
Improving sensitivity of the both the functional and structural image contrast allows
us to detect ever smaller variations. In Chapter 4, we were able to take a large number of
data samples and measure small but signicant changes in the T2* signal between a control
and a stimulated state. To give you an idea of what changes can be detected in state-of-
the-art medical imaging, the actual change in T2* weighted image/signal intensity is in
the order of 1-3%. With multiple repetition of the same stimulus / presentation paradigm
over time and averaging of this signal, the 1-3% raw signal change can be converted in to
a much more manageable change of about 2-3 standard deviations. This is only possible
because in fMRI, we trade image quality for sensitivity and speed.
The signal changes measured in fMRI are intra-subject. This problem becomes even
more dicult with inter-subject population studies using structural scans such as the
diusion MR studies in Part C of the thesis. In these studies, it is more important to
have high quality data that has high SNR / low error since it is rare to have 1000s of
4
subject scans to decrease our standard deviation sizes. In these cases, the actual signal
deviations are closer to 10% of the raw signal. Therefore, artifact correction techniques
such as Cerebrospinal Fluid (CSF) correction (Chapter 8) as well as proper feature ex-
traction (Chapter 13) can really improve statistical power when doing large group studies.
Moreover, the image intensity dierences are "perceived" very dierently by the human
eye versus the computer algorithm. For example, group-wise small signal changes un-
detectable by the human eye can be statistically signicant when using large number of
subjects. We demonstrated this with our MR studies comparing Normal Control (NC),
Mild Cognitive Impairment (MCI) and Probable AD subjects (Chapters 6,9,15).
2.3 Part D: Statistical Analysis and Machine Learning
All of these eorts allows us to create images that have high SNR and low noise with
respect to the desired contrast. These contrasts can be functional (Part B) or structural
(Part C) in nature. Once we have good contrast, the problem is to then design methods
that integrate these images with other clinical information so as to better understand the
complex relationship between laboratory diagnostic tests (such as hematological, genetic,
etc), neuropsychological examinations and the images created from the subject. This
work is contained in Part D of the thesis.
In recent years, there have been tremendous increases in the speed and ubiquity of
diagnostic systems, leading to a dramatic surge in the amount of clinical diagnostic infor-
mation, both imaging and non-imaging, available for driving treatment. For example, in
chapters 15 & 16, the studies contained both MRI information as well as large collection
of clinical measures such as neuropsychological scores, cardiovascular status and insulin
response.
5
In the clinic, however, physicians are usually able to only consider a select number
of information to make treatment decisions. This is where computational techniques are
useful. By integrating all signicant portions of large multi-dimensional datasets, machine
learning and other pattern recognition techniques allows us to nd trends and patterns in
large datasets that might gives us an edge on treatment planning. The question is then
how to design meaningful computation techniques of dimension reduction.
One approach is to use pieces of existing information to put a priori constraints on
the analysis. A lot of reasonable assumptions can be made: For example, in Chapter 14,
we assumed that disease progress is gradual and relatively monotonic. Hence, we put a
constraint on expecting our output disease progression measure (in this case, cognitive
ability) to be changing monotonically when sampled at 6 month intervals. In many cases
simpler approaches that such as multivariate permutation testing (Ch. 15) and fuzzy logic
(Ch. 12) are sucient in accurately doing classication or regression analysis.
Our results are promising, by using fuzzy logic we were able to mimic the pattern
detection work done by pediatric radiologists in matching pediatric hand x-rays to an
atlas to predict the patient's bone age. At geriatric end of research (Ch. 15), we have
began to be able to isolate the dierent parts of the brain that correlates with normal
aging, acute insult and long term genetic factors (such as ApoE). In our machine learning
work (Ch. 14), we are beginning to be able to predict disease progression in the next
48 months by using baseline measures from both clinical testing as well as structural
images. Finally, in our childhood obesity work (Ch.16), we were able to link the amount
of dierence in brain activation between low and high calorie with insulin resistance,
suggesting that the amount of insulin resistance may play a role in regulating the reward
pathway response to food or vice versa.
6
3 Organization
Project Bone Age Aging & Dementia Obesity PACS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Modality X-ray FLAIR DTI T1 fMRI fcMRI PACS
A
Imaging Ch. 2
Informatics Ch. 3
B
Functional Ch. 4
Neuroimaging Ch. 5
C
Ch. 10 Ch. 6 Ch. 11
Structural Ch. 7
Neuroimaging Ch. 8
Ch. 9
D
Statistics & Ch. 12 Ch. 14
M.L. Ch. 13 Ch. 15 Ch. 16
Figure 1.1: Organizational chart for this thesis
This thesis organized into four parts:
1. Imaging Informatics
2. Functional Neuroimaging
3. Structural Neuroimaging
4. Statistical Analysis and Machine Learning in Medical Imaging
7
Each part of this thesis outlines my contribution to a particular sub-discipline within
Medical Imaging; Part A outlines my work in improving access to medical imaging infor-
mation by optimizing how digital medical imaging information is transmitted (chapter 2)
and organized (chapter 3), Part B details my contribution to functional Neuroimaging
in particular fMRI in the study of obesity where we applied statistical techniques not
only in the time domain (chapters 4 & 5) but also in the subject group domain (chap-
ter 16). Part C talks about my contribution to structural MRI, it takes the functional
techniques in Part B and applies it to measure inter subject variation (chapter 15. It
also outlines my contribution to Diusion MRI, in particular, Diusion Tensor Imaging
(DTI) where I've worked on improving image contrast (chapter 8) as well as tractography
(chapter 6,7,9). Lastly, I outline my eorts at looking at white matter hyperintensity
(WMH)s, a characteristic of white matter damage shown in
uid attenuated inversion
recovery (FLAIR) images. Part D demonstrates applications of various machine learning
techniques in capstone applications of automated skeletal age assessment (chapter 12)
and predicting disease progression based on a baseline MRI scan.
These medical imaging methods were applied to three main research areas:
1. Obesity and Reward Pathway Disorders
2. Dementias and Aging
3. Skeletal developmental disorders and Bone Age Assessment.
The fourth area PACS can be seen as creating enabling imaging IT infrastructure for the
three main research areas.
8
Part A
Imaging Informatics
9
Chapter 2
Managing Healthcare Information Using Short
Message Service (SMS) in Wireless Broadband
Networks
Jorge Documet
a
, Sinchai Tsao
a
, Luis Documet
b
, Brent J. Liu
a
,
Zheng Zhou
a
, Anika O. Joseph
a
a
Image Processing and Informatics Lab, Department of Radiology, ISI/USC,
Marina Del Rey, CA USA 90292;
b
Saint Johns Health Center, 1328 Twenty-Second Street,
Santa Monica, CA USA 90404-2091;
Medical Imaging 2007: PACS and Imaging Informatics, edited by Steven C. Horii, Kather-
ine P. Andriole, Proc. of SPIE Vol. 6516, 65160T, (2007) 1605-7422/07/$18 doi:
10.1117/12.710160
10
1 Abstract
Due to the ubiquity of cell phones, SMS (Short Message Service) has become an ideal
means to wirelessly manage a Healthcare environment and in particular PACS (Picture
Archival and Communications System) data. SMS is a
exible and mobile method for
real-time access and control of Healthcare information systems such as HIS (Hospital In-
formation System) or PACS. Unlike conventional wireless access methods, SMS mobility
is not limited by the presence of a WiFi network or any other localized signal. It pro-
vides a simple, reliable yet
exible method to communicate with an information system.
In addition, SMS services are widely available for low costs from cellular phone service
providers and allows for more mobility than other services such as wireless internet. This
paper aims to describe a use case of SMS as a means of remotely communicating with a
PACS server. Remote access to a PACS server and its Query-Retrieve services allows for
a more convenient,
exible and streamlined radiology work
ow. Wireless access methods
such as SMS will increase dedicated PACS workstation availability for more specialized
DICOM (Digital Imaging and Communications in Medicine) work
ow management. This
implementation will address potential security, performance and cost issues of applying
SMS as part of a healthcare information management system. This is in an eort to de-
sign a wireless communication system with optimal mobility and
exibility at minimum
material and time costs.
Keywords: DICOM, PACS, PDA, SMS, Study Management Tool, Query-Retrieve
11
2 Introduction
The paradigm of remotely accessing and controlling Healthcare information via cell phone-
based SMS has yet to be explored. To illustrate the utility of this paradigm, a system
has been designed to perform DICOM Query-Retrieve on a standalone model PACS.
Traditionally, the distribution of PACS images in a standalone model PACS would require
a user performing DICOM Query-Retrieve operations on a free workstation and waiting
for the retrieve operation to complete before being able to view the images. Wireless
devices such as PDAs (Personal Digital Assistant) and cell phone-based SMSs have the
potential to streamline this process by allowing the user to wirelessly command the PACS
to transmit the images to a specied workstation beforehand. [3, 4, 5, 6, 7, 8, 9]These tools
are particularly useful when viewing cases that are not normally pre-fetched automatically.
In order to elicit key features vis-a-vis security, performance and costs, the design of the
aforementioned system has been outlined.
3 Methods and Materials
3.1 System Components
The system can reside on either two separate machines or a single machine. The soft-
ware components are distributed on both the Web Server as well as the Gateway Server
(see Table 1). Both single machine and dual machine congurations were tested. In an
attempt to provide a clear distinction between the Gateway and Web server, the systems
will be treated as two separate entities connected by an https (hypertext transfer protocol
- secure) web link as shown in Figure 1. This design utilizes Kannel Open Source SMS
Gateway as the Gateway Server software. The Gateway Server is serially connected to
12
the SMS modem (MultiModem GPRS Wireless Modem from MultiTech Systems, Mounds
View, Minnesota, USA) and communicates via AT commands1. The SMS modem is asso-
ciated with a GSM (Groupe Spcial Mobile or Global System for Mobile Communications)
cell phone account via a SIM (Subscriber Identity Module) card. The phone number of
the cell phone account is used as the server access number. On the Web Server machine, a
Perl capable web server such as Apache is required. The Perl scripts allow the web server
to dynamically create web pages by calling a C++ program that performs the DICOM
communications with the PACS archive.
3.2 Standard Work
ow
3.2.1 Component Perspective
The standard work
ow illustrated in Figure 1 shows how the dierent components interact
with each other. The following steps describe the work
ow: In step 1, a standard query
will start with the cell phone sending a SMS to the SMS gateway. Steps 2 and 3 shows
the gateway accessing a dynamic Perl-based web page from the Web Server via a build-in
web client, thus passing-on the query parameters indicated by the user to the perform a
DICOM operation. In order to respond back with dynamic data, the Web Server calls a
C++ program to perform a DICOM query with the PACS archive in step 4 and 5. The
data is then passed back to the user through the web page and SMS gateway. Once the
user has found the study of interest, the user can then tell the SMS Gateway to issue a
command down the system (steps 2 to 4) to the C++ program, which can then issue a
DICOM send command with the desired parameters as indicated in step 6.
13
Table 2.1: System Components
3.2.2 User's Perspective
Consider table 2, in a standard work
ow the user (1) sends an SMS to the Gateway with
a potential patient name. The system then takes the query from the user and performs
a query based on the criteria sent, generating output (2) in the form of an SMS reply
14
Figure 2.1: Component Interaction Diagram
to the user. To browse the potential matches to the initial query, the user can send two
commands n for next page, which sends a request to the gateway to send an SMS with the
next set of possible matches, or b to send a query to the gateway to resend the previous
choices. In the example, the user does not see the desired patient in the rst 4 results
and sends n to get the next set of results (see step 3). In (4) , the server returns the next
group of results. When the user nds the desired patient Doe, Joe ID number X8787224,
the user then sends 8 to select the desired patient (step 5). The server returns the studies
associated with patient Doe, Joe together with the examination date. In this case, the
server returns two Thorax CTs with its respective date of examination in brackets. Since
the user is looking for the examination performed in 1998 and therefore it is appropriate
to send the associated index number 2 (step 7). The user is then presented with a list
of destinations and as in the previous menu, chooses the desired destination by sending
15
the number 1 for the EFILM workstation 1 in examination room 1 (step 9). The system
has recorded all of the users choices in its cache and sends the appropriate DICOM send
command to the PACS archive (see number 5 in gure 1 or section 2.2.1). To conrm that
the command has been sent successfully, the SMS gateway responds with a conrmation
SMS in step 10. The system is now ready to accept a new query and has cleared the users
cache of the previous data.
3.2.3 Background Processes
To be able to perform a full DICOM-query and direct a specied study to a destination
via DICOM send, the state of the user must be tracked. In other words, the server must
be able to track the input associated with user so as to give the user the appropriate
options to choose from. This process is necessary because no session is kept on the user.
The users choices and data is stored in the users cache which is associated with the users
cell phone number as detected by caller ID. The users state is indicated by the state
diagram (see gure 2), the cache is updated every time the user state changes. Each row
of table 2 is associated with a single state indicated by a number in gure 2. The arrows
shows how the user moves from one state to another. For example, to move from state
1 to state 2, the user has to send a patient name. From state 2 to 3, the user sends a
number that associated the desired patient. The user is also able to send another Name
or patient name while in the patient state, which will elicit another list of possible patient
matches as seen in output 2 in table 2.
16
Table 2.2: Sample User - Server Interaction: Send Patient Doe, Joe (Patient ID:
X8787224)s Chest CT from February 1998 to EFILM workstation in Exam Room 1.
3.2.4 Other Features
To allow the user to return to the new query menu at anytime, there is a special command
0 that resets the system back to its initial state should the user make any mistakes and
17
wishes to start over. Other feathers include prexing the initial query with ID: allowing
the user to query by patient ID (see gure 2).
Figure 2.2: The System State diagram shows how the system keeps track of the users
selections. This state tracking system allows the user to use the simplistic SMS text-
based interface to select the study of interest from the patient of interest to be sent to a
designated DICOM node.
4 Results and Discussion
4.1 SMS System Architecture vis-a-vis PDA System Architecture
The SMS system adds the advantage of being able to direct DICOM studies in the PACS
archive without being requiring internet or local area network access. The only limitation
is a cellular phone carrier signal. It works o the same web server architecture as the
PDA system. [9] Therefore, the Web Server for the SMS is identical to the PDA server
18
except that the generated output is now modied for the SMS, with added functionality
to read from a users state tracking mechanism that allows for navigation between the
dierent states (see Figure 2). The limitation of the SMS is that it lacks the previewing
capability on the web browser as on the PDA system.
Figure 2.3: The SMS System Architecture is a powerful extension of the existing PDA
System Architecture, allowing any authorized user with a cellular phone to direct DICOM
images within the Hospital network.
4.2 System Performance
The average turn around time for a SMS command and a reply from the gateway was
measured to be 15-20 seconds. However, users noted that turn-around time depended
on cellular network congestion. Therefore, the navigation through the three states is
19
on the order of 2-3 minutes provided that the query does not yield too many possible
matches. The system was tested with the PDA server and SMS gateway server on the
same machine.
4.3 System Security
If the SMS gateway needs to be in a separate machine its communication with the PDA
server is via https, a secure protocol. The SMS gateway only allows queries of 3 types of
information: Patient Name, Patient ID and Study Information. This ensures that minimal
patient demographic information can be extracted from the PACS archive. Moreover, the
SMS gateway does not allow transfer of images outside of the pre-designated list of access-
protected destinations. The system can also allow or disallow user-access based on the
users phone number.
4.4 User Preferences
Since the users states are tracked internally, the system allows the administrator to set
user preferences such as authorized destinations. Other possible user-based preferences
could include user customized lters during searches, or a preference to search only by
Patient ID.
4.5 Interface Usability
Referring to Table 2 above, it shows the complete work
ow to perform a retrieve on a
study. The user only has to enter a number at each step following the initial query by
patient name. The only latency in the system is waiting for the server reply, which is about
15 seconds depending on cellular phone network congestion. For a large PACS archive,
20
the screen size and limited number of characters of the SMS protocol becomes a severe
limitation when querying with common last or rst names only. One way to overcome
this problem is to use query by patient ID. Otherwise, the system is user friendly and
requires simple inputs of numbers to make necessary choices.
4.6 System Component and Operation Costs
The system runs autonomously after setup and therefore should require minimal main-
tenance. Since the system can be run from a single machine, the hardware costs are
relatively low. Moreover, the scripts and web server does not put a heavy load on the
CPU or RAM, therefore the minimum hardware requirements are sucient. The cost
of the modem is also relatively low at about USD $300. Thus, the main burden lies in
the cellular phone subscription plans. Most normal cellular phone plans can be used and
usually gives 3000 SMS and 5 cents for each additional SMS received and transmitted a
month for about USD $60.00. However, large hospitals may consider a xed rate plan
with a cellular provider due to their high volume.
5 Conclusions and Future Work
5.1 Future Work
5.1.1 User Authentication
An additional authentication step is required to ensure proper access protection to the
SMS system. This can be easily implemented as extra user authentication state (see gure
2). The user will send an SMS with a password that will be authenticated against the
origin (user) phone number.
21
5.1.2 Hospital Information System (HIS) / Laboratory Information System
Applications (LIS)
Potential other use cases of such a system could include retrieving patient data from HIS
and LIS, provided ways of circumventing HIPAA requirements are found.
5.1.3 Multimedia Capabilities
With the advent of 3G (Third Generation) cellular technologies such as video streaming
and higher bandwidth data connections with cellular networks, it is possible that DICOM
images can also be relayed in the same manner as the SMS messages through the use of
MMS or Multimedia Messaging Service. One of the destinations in our system could be
to the users cellular phone. This would enable the user to see images on the cellular phone
display. Possibilities of features such as Window / Level and Zoom / Pan are possible
but may be limited by the processing power of the cellular phone. Display resolution is
another concern that needs to be addressed.
5.2 Conclusions
The SMS protocol provides an easy and simple way to deploy an application that can
manage the DICOM work
ow without being constrained by networking requirements. It
complements the PDA technology by providing the user with a simpler interface with less
network and bandwidth requirements. This paper shows that although SMS is a text-only
communication tool, it can allow a user to perform tasks on a PACS network wirelessly.
What remains is to see how receptive physicians and technologists are to such a method
of streamlining their work
ow. However, certain tasks and work
ow scenarios will lend
22
itself more to SMS-based wireless control. For example, tasks such as CD burning of
studies through a CD burning DICOM node would be well suited for this system.
SMS-based management of PACS work
ow is only one potential application of this sys-
tem. There are potentially a large variety of use cases for SMS in healthcare information
management. Certainly, text-based laboratory information could easily be transmitted
over SMS. Application of MMS to transmit images may open up greater possibilities in
the realm of image-based information.
23
Chapter 3
RadSearch: A RIS/PACS Integrated Query Tool
Sinchai Tsao
a
, Jorge Documet
a
, Paymann Moin
a
, Kevin Wang
a
, Brent J. Liu
a
a
Image Processing and Informatics Lab, Department of Radiology, ISI/USC,
Marina Del Rey, CA USA 90292;
Medical Imaging 2008: PACS and Imaging Informatics, edited by Katherine P. Andriole,
Khan M. Siddiqui, Proc. of SPIE Vol. 6919, 691909, (2008) 1605-7422/08/$18 doi:
10.1117/12.772983
24
1 Abstract
Radiology Information Systems (RIS) contain a wealth of information that can be used
for research, education, and practice management. However, the sheer amount of infor-
mation available makes querying specic data dicult and time consuming. Previous
work has shown that a clinical RIS database and its RIS text reports can be extracted,
duplicated and indexed for searches while complying with HIPAA and IRB requirements.
This projects intent is to provide a software tool, the RadSearch Toolkit, to allow intelli-
gent indexing and parsing of RIS reports for easy yet powerful searches. In addition, the
project aims to seamlessly query and retrieve associated images from the Picture Archiv-
ing and Communication System (PACS) in situations where an integrated RIS/PACS
is in place { even subselecting individual series, such as in an MRI study. RadSearchs
application of simple text parsing techniques to index text-based radiology reports will
allow the search engine to quickly return relevant results. -This powerful combination
will be useful in both private practice and academic settings; administrators can easily
obtain complex practice management information such as referral patterns; researchers
can conduct retrospective studies with specic, multiple criteria; teaching institutions can
quickly and eectively create thorough teaching les.
Keywords: NLP, PACS, RIS, Data Mining, Radiology, Radiology Reports
2 Introduction
2.1 Previous Work
Previous work has shown that a RIS database can be searched and radiology report text
can be parsed and indexed. Desjardins and Hamilton showed that radiology reports can be
25
extracted and indexed with minimal disturbance of the clinical work
ow while complying
with HIPAA and IRB requirements. Current trends in improving standards in Radiology
Reporting as well as current practices of report voice recognition dictation using standard
templates demonstrate predictability in the location of information within the format of
the radiology report. Classication tools such as Index Medicus (IM), Medical Subject
Headings (MeSH), International Classication of Disease (ICD), Systematized Nomencla-
ture of Medicine (SNOMED), Unied Medical Language System (UMLS) and RadLex
are built to classify large volumes of medical literature for easy searching. RadSearchs
end goal is to apply the same concept to Radiology Reports within the RIS.
2.2 RadSearch Toolkit
The toolkit builds upon current functionalities in previous work by (1) parsing the radiol-
ogy report intelligently into multiple clinically relevant subsections (2) linking the search
results to PACS studies. This will enable administrative users of the toolkit to track refer-
ral patterns, perform receiver operating characteristic (ROC) analysis, and track billing
codes for studies. The toolkit also allows researchers to easily perform large, complex
retrospective PACS-based studies by querying for one or more study indications, nd-
ings, impressions, diagnostic codes or any other indexed criteria. Finally, the RadSearch
toolkit allows educators to easily compile interesting cases and quickly assemble teaching
les. In short, RadSearch will enhance radiological informatics support for three main
groups: Administrators, Researchers, and Educators.
26
Figure 3.1: Prototype Database Schema
3 Methods
3.1 Natural Language Processing (NLP) -based Approaches
There have been numerous approaches to the data mining Radiology Text information in
the RIS database. Taira et al. , Friedman et al. as well as other groups have approached
the problem by classication for the specic type of Radiology Reports such as Chest X-
rays or Mammograms using Natural Language Processing Techniques. Leximer, a NLP
engine at Massachusetts General Hospital, went further to describe referral trends by
classifying pathological ndings. Nuance Inc. the company behind Leximer, also oers
applies similar techniques to improve billing by NLP-based classication of billing codes.
However, there have yet to be a tool that allow for generalized searches of the RIS database
for teaching le generation or retrospective studies that also have PACS integrated image
retrieval.
27
Figure 3.2: Overall System Diagram
3.2 Lexicons and Synonymity
Current work have also concentrated in including synonymity and uniformity to searches
of Medical Literature (Journals and other publications) using thesauruses such as Unied
Medical Language System or Radiological Society of North America (RSNA)s RadLex
(Radiology Lexicon) for Radiology. These lexicons can just as easily be employed in
Radiology Report searches and can be achieved in two ways. (1) Using the lexicons as a
pure thesaurus, therefore searching of one specic term will search for all related terms.
(2) Pre-Classication of the text with specic unied key terms for specic sub-headings
and convert all queries to use those key terms. The later approach has advantages with
increased search speed and accuracy since the text data can be indexed using those key
terms.
28
3.3 RIS/PACS Integration
Integrated PACS/RIS solutions have shown that it is possible to link RIS radiology reports
back to PACS image series using the Series Instance UID portion of the RIS radiology
report table. However, none of the current tools allow RIS data mining techniques with
PACS integration.
4 Results
4.1 System Architecture
The system is developed utilizing a SUSE linux box running a postgreSQL database as
well as an Apache web server with PHP enabled. It employs current PHP scripting
technology to perform database manipulations as well as to display results via a web
interface. Advantages to using this architecture is that server-side scripting languages
such as PHP do not require a compilation step and have an easy to use interface with
html for displaying results on the web. PACS connections will be run using the open-
source DCMTK toolkit executables called by PHP scripts.
4.2 Radiology Text Report Parsing
The current system has links from the RIS database to a text le that contains the
Radiology reports. Our goal is to breakdown the RIS text into broad elds such as:
Study Title, Location, Indications, Findings, Technique and Impressions. We can then
index the database based on these subsections. This allows the users to make more
precise searches based on the aforementioned elds. An initial parsing algorithm employs
a simple method of detecting each of the subtitles and parsing the text into each of
29
the subsections. Subsections not detected is left null. The detected values are then
populated into the database and indexed for expedited searching. The PHP scripts will
handle both the retrieval of the text le containing the Radiology Text report, parsing, as
well as population back into the database using the proposed schema in gure 1. Future
advanced Natural Language Processing algorithms can be applied to enhance and improve
the parsing step further and, thus improve the overall performance of the toolkit.
4.3 Forming Queries
Queries to the system is formed by specifying key terms for each of the subelds indicated
in gure 1. This can include patient demographic data as well as information contained
within the Radiology Report. The results will then be returned in the form of a list of
Radiology Reports that matched that query along with links to related images.
4.4 Integration with PACS and as a Study Management Tool (SMT)
RadSearch builds upon our current real time patient dashboard [10], which tracks the
status of current patients wait time and status within the radiology work
ow. It warns
clinicians and technicians if the patient has been waiting a long time or if the study hasnt
been read after a certain duration past study completion. The patient dashboard does
this via its real ime connection with the clinical RIS database and understanding of how
and when the RIS database is updated throughout the work
ow. RadSearch, therefore,
builds upon this understanding to construct a query tool for the same set of database
information. Data mining for the patient wait times, time from study completion to
reading by a radiologist and other matrices is also possible since it resides on the same
database system. PACS integration comes in when the user decides to view images related
30
to the desired Radiology Report. PACS images are query-retrieved and displayed in a
presentation state (JPEG) as specied in the DICOM convention for WADO (Web Access
for DICOM Objects). The system can also function as a study management tool that
allows the images to be moved via DICOM standards to other DICOM nodes for CD
burning, display on a PACS workstation, etc. This is done via the Series Instance UID
attached to the desired Radiology Report. The Series Instance UID allows the RadSearch
server to query-retrieve all series images from the PACS for the user or initiate a DICOM
C-MOVE to move images to another DICOM node.
4.5 System Operation Considerations
To minimize disruption of the clinical RIS system, the system database is duplicated
to a postgres SQL database on the RadSearch Server. The system will have weekly
duplication sessions during non-peak times such as Sunday nights. PACS connections
and query-retrieval will be based on an as-needed basis and should not eect the current
clinical operations of the PACS system. Early uses of the system are limited to teaching
le generation. IRB approval for use as a retrospective clinical research tool will be
implemented when the system has been tested for reliability. User authentication as well
as usage logging is also implemented to ensure HIPAA and IRB compliance.
5 Discussion
5.1 RadSearch versus other RIS Data Mining Approaches
RadSearchs approach diers from the other approaches in two folds (1) RadSearch includes
PACS integration (2) RadSearch allows for higher specicity in the Radiology Report by
31
parsing the report into subsections. PACS integration is essential as it allows the user to
preview the desired images on the same interface as the RIS search engine. Previously, the
user would have to obtain the patient name or ID from the search interface and search for
it within the PACS via a PACS workstation before being able to determine if those images
are what the user is looking for. With PACS integration, RadSearch will allow the user
to seamlessly preview the images together with the list of potential Radiology Reports.
Ability to search within specic elds is also important, as the appearance of a term in one
subsection of the report may have dierent meaning than if it appeared within another
subsection. Therefore, this ability to search within subsections will improve specicity,
accuracy and quality of the user query.
5.2 Future Work
We intend RadSearch to be a platform for future work. The next obvious step is to use
a Lexicon to build a classication engine so as to allow for synonimity in searching. We
plan to integrate the RadLex lexicon system, since it is build specically for Radiology.
Other possibilities include classication of ndings, techniques and indications in the
examinations to allow for statistical data mining.
32
Part B
Functional Neuroimaging
33
Chapter 4
Application of fMRI to Obesity Research:
Dierences in Reward Pathway Activation
measured with fMRI BOLD during Visual
Presentation of High and Low Calorie Foods
Sinchai Tsao
a
Tanja C. Adam
b
Michael I. Goran
c
and Manbir Singh
a
;d
a
Department of Biomedical Engineering, Viterbi School of Engineering, University of
Southern California, Los Angeles, California, USA
b
Department of Human Biology, Maastricht University, 6200 MD Maastricht,
Netherlands
c
Department of Preventive Medicine, Keck School of Medicine, University of Southern
California, Los Angeles, California, USA
d
Department of Radiology, Keck School of Medicine, University of Southern California,
Los Angeles, California, USA
Medical Imaging 2012: Biomedical Applications in Molecular, Structural, and Functional
Imaging, edited by Robert C. Molthen, John B. Weaver, Proc. of SPIE Vol. 8317, 83170B
c
2012 SPIE CCC code: 1605-7422/12/$18 doi: 10.1117/12.911854
34
1 Abstract
The factors behind the neural mechanisms that motivate food choice and obesity are not
well known [31, 32]. Furthermore, it is not known when these neural mechanisms develop
and how they are in
uenced by both genetic and environmental factors. This study uses
fMRI together with clinical data to shed light on the aforementioned questions by inves-
tigating how appetite-related activation in the brain changes with low versus high caloric
foods in pre-pubescent girls. Previous studies have shown that obese adults have less
striatal D2 receptors and thus reduced Dopamine (DA) signaling leading to the reward-
decit theory of obesity[33, 34]. However, overeating in itself reduces D2 receptor density,
D2 sensitivity and thus reward sensitivity. The results of this study will show how early
these neural mechanisms develop and what eect the drastic endocrinological changes
during puberty has on these mechanisms. Our preliminary results showed increased acti-
vations in the Putamen, Insula, Thalamus and Hippocampus when looking at activations
where High Calorie > Low Calorie. When comparing High Calorie > Control and Low
Calorie> Control, the High> Control test showed increased signicant activation in the
frontal lobe. The Low > Control also yielded signicant activation in the Left and Right
Fusiform Gyrus, which did not appear in the High > Control test. These results indicate
that the reward pathway activations previously shown in post-puberty and adults are
present in pre-pubescent teens. These results may suggest that some of the preferential
neural mechanisms of reward are already present pre-puberty.
Keywords: fMRI, MRI, Obesity, Reward Pathway, Food, Dopamine
35
2 Introduction
The factors behind the neural mechanisms that motivate food choice and obesity are
not well known [31, 32]. Furthermore, it is not known when these neural mechanisms
develop and how they are in
uenced by both genetic and environmental factors. This
study uses fMRI together with clinical data to shed light on the aforementioned questions
by investigating how appetite-related activation in the brain changes with low versus
high caloric foods in pre-pubescent girls. Previous studies have shown that obese adults
have less striatal D2 receptors and thus reduced Dopamine (DA) signaling leading to the
reward-decit theory of obesity[33, 34]. However, overeating in itself reduces D2 receptor
density, D2 sensitivity and thus reward sensitivity. Moreover, Killgore et al's studies have
shown that aective also eects the stratal activation in the brain [35, 33], suggesting that
a numerous temporary inputs and more permanent factors such as body weight [36, 37, 38]
could modulate the reward center. In this study we aim to use this fMRI techniques to
measure dierential activations due to the visual presentation of High Calorie (HC) and
Low Calorie (LC) foods to prepubescent children. The motivation was to investigate
these responses in the youngest population possible before factors like decreased Insulin
Sensitivity sets in. The hope is that dierences found will be due to responses that
is aecting their obese status and not secondary to the adiposity. Furthermore, the
presence of these dierential responses in the pre-pubescent populations and investigation
of potential treatment or weight management techniques will allow us to nd out whether
the endocrinological changes during puberty solidies the pathology in terms of neural
pathways and responses.
36
3 Methods
3.1 Participants
13 obese girls between the ages of 8 and 11 were selected to participate in this fMRI and
clinical study. 3 of the studies were rejected because of either low fMRI data quality due
to movement or violation of clinical protocol in measuring parameters such as Fasting
Insulin. The remaining 10 girls had a mean age of 9.90 years and a standard deviation
of 1.20 years. Their mean weight was 59.30 lbs with a standard deviation of 13.43 lbs.
Their mean height was 144.06 cm with a standard deviation of 7.12 cm.
fMRI
30 s
30 s
30 s
30 s
30 s
Control Control Control High Low
Figure 4.1: fMRI visual stimulus presentation paradigm as presented through the head
coil mirror to a projection screen at the foot of the subject. All images were based on
previous studies or through the International Aective Picture System. The rst three
box car functions below the images refer to the three stimulus reference functions for
NF/Control, HC and LC respectively.
37
3.2 Procedures
This study was approved by the University of Southern California IRB committee. After
informed consent by the subject and their parents, the subjects were required to partici-
pate in two separate visits. The rst visit was for anthropometric and body composition
measurements followed by an overnight stay and a Frequently Sampled Intravenous Glu-
cose Tolerance Test (FSIVGTT). The second visit was for the neuroimaging study outlined
in the next section. A pediatric health care provider conducted a medical history and
physical examination together with a determination of maturation using the citeria of
Tanner [39]. Body composition for all subjects was measured with dual-energy x-ray ab-
sorptiometry (DXA) using a Hologic Discovery A model (82702, Hologic, Bedford, MA).
3.3 MRI Protocol
An fMRI study ( TR= 3000 ms, TE = 25 ms, Slice Thickness = 5 mm, Number of Slices
= 33, Acquisition Matrix = 64 x 64, Voxel Size = 3.4375 x 3.4375) was performed on each
participant and consisted of presenting sets of 10 randomized images of High Calorie (HC)
Foods, Low Calorie (LC) Foods and Control Non-Food (NF) Images. The visual stimulus
was presented using the PsychToolbox for MATLAB [40] to ensure accurate timing. The
visual stimulus was an alternating set of NF, HC, NF, LC, NF images (Figure ??). Each
group of images were randomized to make sure that the results were not due to the eect
of any single image but the average of the set as a whole. Anatomical images using T1
FSE and FSPGR sequences were also acquired. All data was acquired on a GE 3T HDxT
scanner located at University of Southern Californias Health Consultation Center II.
38
3.4 SPM Post-processing
SPM8 was used to process the fMRI data. The data was rst motion corrected then
co-registered to it's corresponding T1-weighted FSPGR image. The T1-weighted FSPGR
images are then normalized to an MNI template. This transformation was then applied
to the motion-corrected fMRI images. The ROIs were extracted using the Anatomical
Labeling Toolkit for SPM [1] .
Table 4.1: Results when comparing High Calorie > Control ordered by z score with
Talairach Atlas Coordinates and the Automatic Labeling of the ROIs was done using
AAL toolkit for SPM [1]
Regions of Activation x y z z score
L. Cerebelum Crus I -26 -88 -26 3.33
L. Fusiform -42 -50 -18 3.27
R. Frontal Middle 32 50 32 3.07
L. Fusiform -38 -68 -16 3.02
L. Cerebelum Crus I -46 -76 -22 3.02
L. Occipital Middle -30 -84 6 2.99
R. Fusiform 28 -68 -8 2.59
4 Results
Our preliminary results showed increased activations in the Putamen, Insula, Thalamus
and Hippocampus when looking at activations where High Calorie > Low Calorie. When
comparing High Calorie > Control and Low Calorie > Control, the High > Control test
39
Table 4.2: Results when comparing Low Calorie > Control ordered by z score with Ta-
lairach Atlas Coordinates and the Automatic Labeling of the ROIs was done using AAL
toolkit for SPM [1]
Regions of Activation x y z z score
L. Frontal Superior -22 22 54 3.57
R. Frontal Superior 14 28 54 3.16
R. Cerebellum Crus I 24 -84 -24 2.92
R. Frontal Middle 30 14 54 2.80
L. Cerebelum Crus I -16 -86 -28 2.73
L. Occipital Inferior -50 -78 -12 2.55
showed increased signicant activation in the frontal lobe. The Low> Control also yielded
signicant activation in the Left and Right Fusiform Gyrus, which did not appear in the
High > Control test.
5 Discussion
Our preliminary results show that the dierential activations detected by numerous pre-
vious studies [41] are also present in children. This is important because the end goal
of the study is to determine how factors such as Insulin Sensitivity, Fasting Insulin lev-
els and Body Composition are driving forces behind the development of obesity. Obese
adults tend to have a low range of insulin sensitivity and therefore make investigating
the relationship between brain reward and insulin sensitivity dicult. However, with
obese children they still retain a large range of insulin sensitivity [42]. The hypothesis
40
is that insulin as a neurotransmitter plays a crucial role in regulating the reward centers
in the brain and therefore a loss in insulin sensitivity will cause subjects to be unable to
manage visual food stimulus appropriately. In other words, the loss of insulin sensitivity
means that the patient will be unable to temper food stimulus appropriately leading to
overeating and obesity. Previous work has shown to be consistent with this hypothesis
[42, 43, 44, 45]. Most of the insulin receptors in the brain lie within the limbic region,
our preliminary results show that a dierential in activation occurs when the subjects are
stimulated with high versus low calorie foods lie within the limbic regions (see Table 4.3).
This coupled with preliminary clinical data indicating that our subjects have a wide range
of insulin sensitivity suggests that it may be possible to correlated the degree of dieren-
tial activation with each subject's insulin sensitivity. Furthermore, our study also collects
body adiposity measurements, which have been known to be a factor in insulin resistance.
Our current hypothesis suggests that insulin sensitivity may be an earlier precursor to
body adiposity due to it's neurotransmitter role in the limbic system. This hypothesis
can be tested by carefully controlling for body adiposity in our study thus showing that
perhaps insulin sensitivity and not body adiposity is the culprit in mis-regulation of visual
food stimulus in the obese.
6 Conclusion
Our study shows how we are able to extend traditional fMRI studies to investigate patho-
logical neurological pathways in the obese. This example simply shows that if clinical
parameters such as insulin sensitivity and adiposity can be incorporated into fMRI image
analysis, a large range of clinical hypothesis can be tested. This study outlines the initial
41
steps of this type of analysis. These initial steps are important in establishing that the
responses that we are correlating are indeed in areas that are related to the pathway in
question. In our case, multiple parts of the limbic region and other brain reward areas
were found to be signicant in our High > Low Calorie analysis. These regions did not
appear to be signicant in our High> Control and Low> Control comparisons indicating
that these regions are indeed involved only in dierentiating between low and high visual
food stimulus [41].
This paradigm of combining traditional clinical studies of pathologies such as obesity
with fMRI techniques have traditional used only for primary neuroscience research but
can easily be applied to a variety of disorders, obvious targets include Alzheimer Disease
and Mildly Cognitively Impaired. Other applications include relating the change in fMRI
activity with therapeutic outcomes and treatments.
ACKNOWLEDGMENTS
The authors would like to acknowledge the help of Kori Wong Kazama Tsao for her
clinical psychology expertise in handling the anxiety of the young female subjects in the
MRI. Darryl H. Hwang for his help in the operation of the MRI as well as handling of
subjects. Kathleen A. Page for introducing the authors and her contribution to the use
of fMRI in Obesity Research. Houchun Harry Hu for developing the initial version of the
fMRI visual stimulus. Samuel Valencerina for his help with the operation of the MRI at
USC. Last but not least, I'd like to dedicate this work to the late Professor Manbir Singh
PhD, my mentor, friend and inspiration who's sudden passing this past winter has been
42
a tragic loss for us all. This work could not have been possible if not for every single one
of you.
This study was supported by the National Cancer Institute (U54 CA 116848), a Pzer
Fellowship in Health Disparities awarded to Tanja C. Adam and by National Center of
Research Resources (NCRR) grant 1S10RR019942 that provided the 3T MRI used in this
study.
43
Table 4.3: Results when comparing High Calorie > Low Calorie ordered by z score with
Talairach Atlas Coordinates and the Automatic Labeling of the ROIs was done using
AAL toolkit for SPM [1]
Regions of Activation x y z z score
R. Temporal Superior 60 -52 24 3.24
R. Frontal Middle 46 18 34 3.11
L. Frontal Medial Orbital -6 50 -4 3.09
R. Frontal Superior 28 32 48 3.05
L. Frontal Middle -24 6 48 3.01
L. Precentral / Postcentral -30 32 62 2.89
L. Frontal Middle -24 32 42 2.88
R. Hippocampus / R. Thalamus 18 -36 6 2.85
L. Frontal Middle / Superior -20 20 52 2.82
R. Temporal Superior 58 -18 -2 2.77
R. Putamen / R. Insula 32 -20 0 2.74
Caudate 0 14 -2 2.73
L. Frontal Inferior Opercularis / Middle -48 16 34 2.63
R. Frontal Middle 30 18 54 2.48
44
Differences in Appetite-Related Brain Activity with High and Low Calorie Food in Pre-Pubescent Girls
S. Tsao
1
, T. Adam
1
, D. H. Hwang
1
, M. I. Goran
1
, and M. Singh
1

1
University of Southern California, Los Angeles, CA

Introduction: The factors behind the neural mechanisms that motivate food choice and obesity are not well known.
Furthermore, it is not known when these neural mechanisms develop and how they are influence by both genetic and
environmental factors. This study uses fMRI together with clinical data to shed light on the aforementioned questions by
investigating how appetite-related activation in the brain changes with low versus high caloric foods in pre-pubescent girls.
Previous studies have shown that obese humans have less striatal D2 receptors and thus reduced Dopamine (DA) signaling
leading to the reward-deficity theory of obesity. However,
overeating reduces D2 receptor density, D2 sensitivity and thus
reward sensitivity. The results of this study will show how early
these neural mechanisms develop and what effect the drastic
endocrinological changes during puberty has on these mechanisms.
Materials and Methods: 13 obese girls between the ages of 8 and
11 were selected to participate in the study. An fMRI study was
performed on each participant and consisted of presenting sets of 10
randomized images of High Calorie (HC) Foods, Low Calorie (LC)
Foods and Control Non-Food Images. Each participant’s percentage
body fat was also measured along with body height & weight, hip
and waist circumference, Intravenous Glucose Tolerance Test
results (IVGTT), Body Mass Index (BMI), Tanner Stage as well as
Resting Heart Rate.

Results and Discussion: Our preliminary results showed increased
activations in the Putamen, Insula, Thalamus and Hippocampus
when looking at activations where High Calorie > Low Calorie.
When comparing High Calorie > Control and Low Calorie >
Control, the High > Control test showed increased significant
activation in the frontal lobe. The Low > Control also yielded
significant activation in the Left and Right Fusiform Gyrus, which
did not appear in the High > Control test.
Conclusions: Our preliminary results show that the
reward pathway activations previously shown in post-puberty and adults are
present in pre-pubescent teens. These results may suggest that the some of the
preferential neural mechanisms of reward are already present pre-puberty. We
hope that by correlating our lab results such as IGVTT with the degree of neural
activation in our SPMs we can see which regions of the pre-pubescent brain
have been affected by obesity related physiological factors such as insulin
resistance.
Figure 1. Group statistical parametric map (SPM)
comparison of activated areas during presentation
of Low Calorie, High Calorie and Control Images
(P<0.01, uncorrected) with an extent threshold of 10
Voxels. The SPMs are presented in two ways
(LEFT) Rendered onto a single subject’s brain
(RIGHT) Maximum Intesity Projections (MIP) in
three views.
For Tables Note:
L. Left Hemisphere
R. Right Hemisphere.
Atlas coordinates x, y,z
in mm are from the
Talairach and
Tournoux (1988) Atlas.
Only regions with z
score significant
beyond P<0.01
(Uncorrected are
shown)
Table 1. High Calorie > Control
Table 2. Low Calorie > Control
Table 3. High Calorie > Low Calorie
Figure 4.2: Group statistical parametric map (SPM) comparison of activated areas during
presentation of Low Calorie, High Calorie and Control Images (P < 0:01, uncorrected)
with an extent threshold of 10 Voxels. The ROIs are rendered on to a template single
subject's brain surface.
45
Chapter 5
Novel MRI Methodology to detect Human
Whole-Brain Connectivity Changes after
Ingestion of Fructose or Glucose
Sinchai Tsao
a
Bryce Wilkins
a
Kathleen A. Page
b
and Manbir Singh
a;c
a
Department of Biomedical Engineering, Viterbi School of Engineering,
University of Southern California, Los Angeles, California, USA
b
Department of Internal Medicine, Division of Endocrinology, Keck School of Medicine,
University of Southern California, Los Angeles, California, USA
c
Department of Radiology, Keck School of Medicine,
University of Southern California, Los Angeles, California, USA
Medical Imaging 2012: Biomedical Applications in Molecular, Structural, and Functional
Imaging, edited by Robert C. Molthen, John B. Weaver, Proc. of SPIE Vol. 8317, 83170B
c
2012 SPIE CCC code: 1605-7422/12/$18 doi: 10.1117/12.911854
46
1 Abstract
A novel MRI protocol has been developed to investigate the dierential eects of glucose
or fructose consumption on whole-brain functional brain connectivity. A previous study
has reported a decrease in the fMRI blood oxygen level dependent (BOLD) signal of the
hypothalamus following glucose ingestion, but due to technical limitations, was restricted
to a single slice covering the hypothalamus, and thus unable to detect whole-brain con-
nectivity. In another previous study, a protocol was devised to acquire whole-brain fMRI
data following food intake, but only after restricting image acquisition to an MR sam-
pling or repetition time (TR) of 20s, making the protocol unsuitable to detect functional
connectivity above 0.025Hz. We have successfully implemented a continuous 36-min, 40
contiguous slices, whole-brain BOLD acquisition protocol on a 3T scanner with TR=4.5s
to ensure detection of up to 0.1Hz frequencies for whole-brain functional connectivity anal-
ysis. Human data were acquired rst with ingestion of water only, followed by a glucose or
fructose drink within the scanner, without interrupting the scanning. Whole-brain con-
nectivity was analyzed using standard correlation methodology in the 0.01-0.1 Hz range.
The correlation coecient dierences between fructose and glucose ingestion among tar-
geted regions were converted to t-scores using the water-only correlation coecients as
a null condition. Results show a dramatic increase in the hypothalamic connectivity to
the hippocampus, amygdala, insula, caudate and the nucleus accumben for fructose over
glucose. As these regions are known to be key components of the feeding and reward
brain circuits, these results suggest a preference for fructose ingestion.
Keywords: MRI, fMRI, Functional Connectivity, Glucose, Fructose, Obesity, Insulin,
Hypothalamus
47
2 Introduction
The main objective of these pilot studies was to develop a MRI protocol and conduct
initial evaluation studies of the dierential eects of glucose and fructose ingestion on
the feeding pathways in normal (healthy) and obese adults. By identifying brain regions
that show dierential activity across normal and obese subjects following glucose or fruc-
tose ingestion, we hope that the data will permit new understandings of the eects of
glucose/fructose ingestion and their implication in obesity. Recent published research
[46, 47, 48, 49] has shown that glucose ingestion in healthy humans leads to a decrease in
hypothalamic nuclei activity and consequently a decrease in the fMRI blood oxygen level
dependent (BOLD) signal. Furthermore, it has been shown that glucose ingestion fails
to inhibit hypothalamic nuclei activity in patients with type 2 diabetes[50]. These scans
typically require continuous acquisition of fMRI data for a period of about 30-45 minutes,
and due to MRI technical limitations, previous research in this area has been restricted
to acquisition of fMRI data from a single mid-sagittal slice, covering only the hypothala-
mus. A fundamental dierence in our proposed studies was to design a protocol to obtain
fMRI sampling over the entire brain, permitting the detection of glucose/fructose related
changes throughout the brain. A successful protocol would then enable us to conduct an
analysis of whole-brain functional connectivity following glucose/fructose ingestion. To
our knowledge, no such connectivity analysis has been reported previously by any group.
Thus, development of a MRI scan protocol for this project, which presents many key
technological challenges, was considered an essential rst step.
48
Figure 5.1: Three blood samples are taken during the experiment to track the subjects
blood sugar level. The sequence of the individual scans, and times at which blood-
sampling occurs, are indicated in the gure above.
3 Methods
3.1 Subjects
This protocol was tested with 4 Hispanic subjects between the ages of 15-25 who have no
history of diabetes.
3.2 Protocol Development
The challenge was to develop a whole-brain fMRI protocol with continuous data ac-
quisition for approximately 45 minutes that would include a baseline, a relatively long
post-glucose/fructose ingestion period, and a short time for ingesting the glucose/fructose
during the scan. Typically a TR of 2 seconds is used in fMRI connectivity studies and
for TR=2s, continuous acquisition of 4mm thick images covering the whole brain would
require a time-series of 40,500 images. As such long time-series have never been reported
before in fMRI and are not allowed by hardware constraints of many scanners, we did sev-
eral developmental scans to push the technical limits of the USC GE 3T MRI to increase
49
the number of images that could be acquired in a continuous time-series. It was found
that our 3T system could acquire at most 20,000 images continuously before exhausting
its current memory capabilities. However during some developmental scans, the gradients
overheated and the scans had to be terminated.
We argued that as glucose/fructose were diluted in water during ingestion, it would
be important to add a baseline study corresponding to ingestion of water only, so that
confounding eects of water could be separated from those attributed to sugar ingestion.
We conducted several developmental studies to assess the fMRI signals acquired with
water ingestion only and those after ingestion of a sugar drink and came to the conclusion
that because all images must be acquired continuously (stopping and restarting a scan is
not an option as it changes many parameters including the baseline making an accurate
analysis virtually impossible) it was essential to acquire data from a water only period
prior to sugar ingestion in the same continuous scan. Incorporating these requirements,
we came up with our nal 36 min fMRI protocol during which the subject rst drinks
pure water, followed by 10 min of data acquisition, then drinks water containing glucose
(or fructose), followed by about 24 min of data acquisition. Thus the fMRI scan was
divided into three regions: a 10 min baseline condition before ingestion of the sugar
drink; drink period of 2-4 min; and lastly 22-24 min of post-drink data (see Figure ??).
After several trials, the repetition time (TR) of the fMRI scan was chosen to be 4.5s. This
TR is short enough for adequate temporal ltering of the data for functional connectivity
analysis, and long enough to enable us to acquire data continuously for 36 min for a total
of approximately 19,000 fMRI images. As additional images and time are required for
anatomical and diusion tensor imaging (DTI) analysis, 36 minutes were considered as
close to the upper limit for the fMRI portion to nish the entire scan in about one hour.
50
Immediately following the fMRI acquisition, an anatomical T1-weighted scan is ac-
quired to enable registration of subject fMRI data to an anatomical reference template.
This is followed by another higher-resolution (SPGR) anatomical scan that is extensively
used to spatially normalize images and delineate small anatomical regions of interest such
as the hypothalamus.
Following the anatomical scans a diusion-weighted data set is obtained in accordance
with a standard DTI protocol that we have implemented on this GE scanner. This data
may be used to investigate structural connectivity dierences between normal (healthy)
and obese subjects.
3.3 Data Processing and Analysis of fMRI Baseline Trends
As subjects are bound to move their heads during a relatively long scan and certainly
during the drinking period, the rst step in processing is to correct the fMRI time-series
for head-movements. Each subjects' fMRI data is motion-corrected to the last acquired
head-volume, registered to the individuals' high-resolution SPGR anatomical scan, and
warped to normalized space through non-linear mapping techniques. The Colin Brain
template [51] in Talairach Tournoux coordinates was used to dene normalized space in
this study though other templates such as the SPM MNI template could also be used.
Regions of Interest (ROIs) are derived from the San Antonio Talairach Daemon [52, 53],
which is also in the Talairach-Tournoux normalized space.
3.4 Connectivity Analysis
Whole-brain connectivity was analyzed using a standard technique where the movement
and slice-timing corrected fMRI data are band-pass ltered (0.01 0.10 Hz), a seed region
51
is selected and then the time courses of all voxels throughout the brain volume are corre-
lated to the time-course of the seed region to obtain a measure of functional connectivity.
An alternative approach is to rst dene ROIs based on anatomy and then compute the
pair-wise connectivity among averaged time-courses of these ROIs. Both methods were
pursued. In addition, the resting state default connectivity network was also examined.
A methodology was developed to answer the following questions: a) Is there a signicant
dierence between the resting state functional connectivity of the default mode network,
which generally shows strong functional connectivity among the posterior cingulate and
certain parietal and frontal regions, after ingesting glucose vs. fructose? b) Is there a sig-
nicant dierence between the whole-brain functional connectivity of the hypothalamus
after ingesting glucose vs. fructose? c) Is there a signicant dierence in the pair-wise
functional connectivity among specied ROIs after ingesting glucose vs. fructose? To
reduce the eects of inter-subject variations, a pair-wise connectivity analysis was con-
ducted after averaging the time-courses in targeted ROIs of 3 subjects who completed the
glucose and fructose scans. The fourth subjects drinking times were signicantly dierent
to other subjects and as a result data from this subject was not included in the pair-wise
connectivity analysis. The selected ROIs included key brain regions previously suggested
to be involved in brain feeding and reward circuits. The pair-wise connectivity scores
were converted to t-scores by comparing to the water-only baseline connectivities of the
same pairs.
52
4 Results
An example of raw fMRI results from a single subject are shown in Figure 5.2. Due to
scanner drift, these signal cannot be interpreted directly. To account for scanner drift we
demeaned the signal as shown in Figure 5.3. Using a seed region and time course of the
signal before the ingestion of glucose / fructose in the posterior cingulum we were able
to derive correlation maps for the default mode / resting state network (see Figure 5.4
). After the ingestion of glucose / fructose we again used the same seed region and we
were able to qualitative observe changes in the network (see Figure 5.5). We repeated
the same procedure but used the hypothalamus as the seed regions in pre-glucose/pre-
fructose and post-glucose/post-fructose (see Figure 5.6). To expand the search beyond
manual selection of seed regions such as the posterior cingulum and the hypothalamus
we used ROIs derived from the San Antonio Talairach Daemon [52, 53] and correlated
each of the ROIs with each other as shown in Figure 5.7. The colors on the matrix grid
indicate the degree of correlation between the two ROIs. ROIs with t-scores > 2.00 are
shown in Table 5.1.
5 Discussion
Examples of the connectivity images obtained from one such subject are presented in
Figures 5.4,5.5 showing potential dierences in the default network between glucose and
fructose ingestion. A correlation threshold of 0.5 was applied. Similar connectivity dier-
ences between the hypothalamic ROI and voxels throughout the brain were also noted.
As the hypothalamus is composed of just a few voxels and is anatomically located in a
region of the brain where susceptibility artifacts are strong, the signal-to-noise ratio in
53
Figure 5.2: The hypothalamic fMRI time-course for a subject showing an initial water-
only drink period, the actual drinking period (shaded) followed by either a glucose (top) or
fructose (bottom) drink period. Glucose and fructose experiments were done on separate
days.
the time-course is not very high. Thus more variation is expected in the hypothalamic
connectivity analysis than the default network. The hypothalamic connectivity images
are shown in Figure 5.6. Overall it appears that connectivity of several regions is altered
post-fructose vs. post-glucose but there were signicant inter-subject variations (which is
expected).
Though the sample size is very small, the results show a dramatic increase in the
hypothalamic connectivity to very specic regions known to be key components of the
feeding and reward circuits after fructose ingestion when compared to glucose ingestion.
Specically, an increase in hypothalamic connectivity to the hippocampus, amygdala,
insula, caudate and the nucleus accumbens stands out. The largest increase (t-score of 4.5,
which means an increase of 4.5 standard deviations with respect to glucose connectivity)
54
Figure 5.3: (top) The hypothalamic fMRI time-course includes a pre-drink baseline period
preceding a water-only drink and then a second water-only drink. The data is de-trended
by the rst water-only period. (bottom) Same as top except the subject ingested a glucose
drink during the second drinking period. The change in baseline trend following glucose
ingestion is obvious.
55
Figure 5.4: The default-mode network in a pilot subject during the pre-glucose (left) or
the pre-fructose (right) period. These are expected to be the same but as the glucose
and fructose experiments were done on separate days, minor dierences can be seen. The
default network is clearly identied by our protocol.
56
Figure 5.5: The default-mode network in a pilot subject during the post-glucose (left)
or the post-fructose (right) period. Several dierences can be seen, which need further
analysis to interpret their signicance.
57
Figure 5.6: The hypothalamic connectivity in a pilot subject during the pre and post-
glucose periods (top and bottom left respectively) or the pre and post-fructose periods
(top and bottom right respectively).
58
Figure 5.7: Matrix of thresholded t-scores (t > 2) from pair-wise functional connectivity
between select ROIs averaged over 3 subjects; the color-bar gives the t-score to color
mapping. The most prominent connectivity exists between the left hypothalamus and
left/right hippocampus, left/right amygdala, left/right insula, right caudate head, and
left nucleus accumbens. The highest t-score (4.57) is between the left hypothalamus and
left nucleus accumbens.
59
Table 5.1: t-scores corresponding to the pair-wise ROIs shown in Figure 5.7 that exceed
the threshold (t > 2).
ROI 1 ROI 2 t-Score ROI 1 ROI 2 t-Score
L:Hippocampus L:Hypothalamus 2.97 R:Insula L:Caudate Head 2.30
L:Hippocampus R:Caudate Head 2.02 R:Insula R:Nucleus Accumbens 2.31
R:Hippocampus R:Insula 2.08 L:Hypothalamus L:Hippocampus 2.97
R:Hippocampus L:Hypothalamus 3.01 L:Hypothalamus R:Hippocampus 3.01
R:Hippocampus R:Hypothalamus 2.64 L:Hypothalamus L:Amygdala 3.41
R:Hippocampus R:Caudate 2.23 L:Hypothalamus R:Amygdala 2.68
R:Hippocampus R:Caudate Body 2.02 L:Hypothalamus L:Insula 2.22
R:Hippocampus R:Caudate Head 2.99 L:Hypothalamus R:Insula 2.76
L:Amygdala L:Hypothalamus 3.41 L:Hypothalamus R:Caudate Head 2.46
L:Amygdala R:Hypothalamus 3.31 L:Hypothalamus L:Nucleus Accumbens 4.57
L:Amygdala L:Caudate Head 3.47 R:Hypothalamus R:Hippocampus 2.64
L:Amygdala R:Caudate Head 2.03 R:Hypothalamus L:Amygdala 3.31
L:Amygdala R:Nucleus Accumbens 3.31 R:Hypothalamus R:Amygdala 2.45
L:Amygdala L:Anterior Cingulate 2.36 R:Caudate R:Hippocampus 2.23
L:Amygdala R:Anterior Cingulate 2.27 R:Caudate Body R:Hippocampus 2.02
R:Amygdala L:Hypothalamus 2.68 L:Caudate Head L:Amygdala 3.47
R:Amygdala R:Hypothalamus 2.45 L:Caudate Head R:Amygdala 2.41
R:Amygdala L:Caudate Head 2.41 L:Caudate Head L:Insula 2.05
L:Insula L:Hypothalamus 2.22 L:Caudate Head R:Insula 2.30
L:Insula L:Caudate Head 2.05 R:Caudate Head L:Hippocampus 2.02
L:Insula R:Nucleus Accumbens 2.41 R:Caudate Head R:Hippocampus 2.99
R:Insula R:Hippocampus 2.08 R:Caudate Head L:Amygdala 2.03
R:Insula L:Hypothalamus 2.76 R:Caudate Head L:Hypothalamus 2.46
60
is seen for the hypothalamic-nucleus accumbens connection, which is known to be a key
component of the feeding circuit. The selected ROIs and their pair-wise connectivity
results are shown in Figure 5.7. The t-scores for all ROIs above the threshold of t> 2 are
listed in Table 5.1. No signicant increases in pair-wise connectivity among the selected
ROIs were detected for glucose compared to fructose ingestion.
6 Conclusion
After the publication of Liu et al. 2000's Nature article [49], there has been a lot of
interest in being able to investigate how brain activity is in
uenced by food. The tool of
choice has been fMRI. Liu et al 2000 showed how the use of the raw fMRI signal without
deconvolution of the Hemodynamic Response Function (HRF) can give us repeatable
detection of temporal changes in brain activations even though responses from HRF-
based analysis can be seen [54, 33]. Recent developments of functional connectivity and
better knowledge about the default mode network has given us a new framework from
which to investigate the basis of these changes. This paper outlines our initial eorts at
attempting to investigate food related functional connectivity changes using the premise
that the default mode network somehow modulates these responses.
We know that the default mode network is crucial in our ability to maintain homeosta-
sis [55]. Feeding behavior is one complex component of supplying nutrition to maintain
homeostasis and therefore the default mode network should respond accordingly to food
stimulus. The question is whether when and where do these stimuluses get translated into
neurological pathways that solidify our behavior. In particular, how do these dierences
in default mode network responses re
ect pathological states where the human body is
61
unable to regulate feeding behavior properly, leading to the development of obesity. To
make matters more complex, the causality is unknown. In other words, does the patho-
logical feeding behavior cause the dierences in responses or do the responses indicate
an underlying condition that causes the pathological behavior. This coupled with the
dierential activations between fructose and glucose raises a large number of questions
about the neurological mechanisms that govern our feeding behavior. To our knowledge,
no such results have been reported before and if these results can be veried in a larger
study, they would represent a key discovery in how brain connectivity among key regions
is altered by fructose compared to glucose ingestion.
With regards to fMRI processing and techniques, this study serves to show that there
is still much work that needs to be done to properly understand the results of fMRI
studies. The underlying biological mechanisms that lead to the partial correlations in
functional connectivity are not well know [55]. Even if we were to be able to nd changes
that correlate well with parameters such as the ingestion of glucose versus fructose, we
are not able to draw denitive conclusions unless the underlying mechanisms are better
understood. Current glucose ingestion studies together with studies that use pharmaceu-
ticals alter our homeostatic state [56] will help to elucidate some of these mechanisms.
This together with other anthropometric techniques as well as PET studies will further
clarify the phenomenon observed in our fMRI studies [54, 33]. Techniques that better
utilize the temporal information within the fMRI signal will continue to push the enve-
lope in terms of the information elicited from imaging studies. Functional Connectivity
and other techniques will continue to push the evolution of techniques that do not use
the HRF. This will hopefully help us establish a better understanding of how each part
62
of the brain responds vis- a-vis another part of the brain and not just be able to isolate
individual activities as predicted by the BOLD HRF.
Acknowledgements
The authors would like to acknowledge Samuel Valencerina for his help with the MRI at
USC. Lab-mates Darryl Hwang for helping with the development of the MRI protocol.
In particular, his help with the glucose ingestion delivery system. I'd like to dedicate this
work to the late Professor Manbir Singh PhD, my mentor, friend and inspiration who's
sudden passing this past winter has been a tragic loss for us all. This study was partially
supported by National Center of Research Resources (NCRR) grant 1S10RR019942 that
provided the 3T MRI used in this study.
63
Part C
Structural Neuroimaging
64
Chapter 6
Quantication of Fornix Tracts in MCI and AD
D. H. Hwang
1
, S. Tsao
1
, and M. Singh
1
1
Depts of Biomedical Engineering and Radiology,
University of Southern California, Los Angeles, CA, United States
Proc. Intl. Soc. Mag. Reson. Med. 17 (2009)
65
1 Introduction
Numerous studies have shown a pattern of progressive hippocampal shrinkage as a result
of mild cognitively impairment (MCI) and Alzheimer Disease (AD). One of the major
pathways aected by the hippocampus is the fornix, and it has been recently shown
that fractional anisotropy (FA) is reduced in normal controls verses AD [DeCarli et al.,
Alzheimers Imaging Consortium IC-P1,July 2008]. The purpose of this study was to
(a) objectively dene hippocampal ROIs as most previous studies rely on subjectively
delineated ROIs to perform DTI analysis, (b) correct for varying subject head sizes using
intracranial volume( ICV) of subjects in a novel way to equalize the number of seed points
and their anatomical distribution so that total number of tracts becomes independent of
the subject's head size, and (c) examine the dierences in the equalized streamline fornix
tracts.
Normal MCI AD
Figure 6.1: Normalized Fornix Tracts
2 Methods
Data were acquired on a 3T GE scanner from three groups (10 normal control, 8 MCI, and
7 AD probable) as classied by neuropsychological testing. Multi-slice (2:04 2:04 4
66
Figure 6.2: ICV Normalized Freesurfer-based Hippocampus Volumes in Normal, MCI and
Probable AD
mm
3
voxel size, 4 mm thick slices no gap) DTI data were acquired using 25 encod-
ing gradient directions at bvalues of 0 and 1000 sec/mm
2
with an acquisition matrix of
128 128. Anatomical data were acquired using a 256 256 acquisition matrix em-
ploying a SPGR sequence (1 1 1 mm
3
voxel size, 1 mm thick slices no gap) with
TE=3.128ms, TR=7.832ms and TI=450ms. The hippocampi were identied using high
resolution SPGR images via the Freesurfer software package (Harvard/MGH). The auto-
matically segmented hippocampal ROIs were then co-registered back to the DTI subject
space using SPM. Seed points where normalized by adjusting spacing according to subject
intracranial volume (ICV). ICVs were calculated based on a weighted combination of grey
and white matter, and CSF segmented volumes generated by SPM. By using a control
average ICV of 1.5x106 mm3, all seed spacing was adjusted to compensate for ICV in
individuals. Streamline tractography was performed using tensor interpolation with a
0.2 mm step-size utilizing in-house DTI software. Whole-brain tractography data were
67
Figure 6.3: ICV Normalized Tract Counts in Normal, MCI and Probable AD
processed by several stages of ltering to obtain the fornix tracts. The steps are based
on two known facts (1) fornix tracts connect the hippocampus and hypothalamus and (2)
these tracts do not cross right/left hemispheres. First, only the tracts which intersect the
hippocampal ROIs are considered. Next, all tracts originating from seed points outside of
the fornix region are eliminated. Tracts originating from the fornix regions that progress
out of the fornix region are kept for tract counting. However, the aforementioned tracts
are cut back to where they deviate from the fornix for display purposes (g. 1).
3 Results and Discussion
Results of hippocampal volume ( g. 2) conrm the progressive degenerative eects of
MCI and AD and validate the usability of Freesurfer dened hippocampal ROIs without
any subjective input. ICVNormalized tract counts for the fornix (g.3) show a similar
progressive reduction. Although there is a measurable decrease in connectivity based on
68
Figure 6.4: Hippocampus Volumes and Tract Count Eect Sizes
normalized tract count (Prob AD p=0.039,e size = 0.9), the eect seems to be less than
what is measured using normalized hippocampal volumes (Prob AD p=0.0007, e size=
1.8) (table 1). Although the eect of Alzheimer Disease on hippocampal volume is known,
its eects on the aerent and eerent connections between the hippocampus and other
structures via the crucial fornix tract is relatively unknown. Relatively large voxel sizes
were used in this study ( 2x2x4mm3) creating partial volume confounds. We expect that
as higher resolution DTI data are acquired, the eects of partial volume will be reduced
leading to better tract quantitation and improvements in eect sizes corresponding to
fornix and other aected tracts.
69
Table 6.1: Eect Sizes and t-test of Normalized Hippocampus Volume and Normalized
Tract Count against Normal Population
70
Chapter 7
Comparison of Limbic Regions FA Using
Tractography-dened ROIs in AD and MCI
D. H. Hwang
1
, S. Tsao
1
, and M. Singh
1
1
Depts of Biomedical Engineering and Radiology,
University of Southern California, Los Angeles, CA, United States
Proc. Intl. Soc. Mag. Reson. Med. 18 (2010)
71
1 Introduction
The fornix and cingulum (limbic regions) are suspected of being aected by mild cognitive
impairment (MCI) and Alzheimer Disease (AD). Recent studies have reported observable
reductions in FA of the cingulum between normal and AD populations. Most studies
resort to manual tracing of the anatomy in patient acquisition space. Objective isolation
of the fornix and cingulum has remained elusive. The purpose of this study was to (a)
bring DTI including tractography to a common template space, (b) dene the fornix and
cingulum objectively using tractography, and (c) examine the FA changes among normal
control, MCI and probable AD populations.
Figure 7.1: Fornix FA t-Score Comparison. MCI > AD (top row) NC > MCI (bottom
row). Colors denote the t-score dierence between the mean FA of the various populations.
72
Figure 7.2: Cingulum FA t-Score Comparison. MCI >AD (top row) NC> MCI (bottom
row)
2 Methods
Data were acquired on a 3T GE scanner from three populations (26 normal control, 16
MCI, and 13 AD probable) as classied by neuropsychological testing. Multi-slice (2.04
x 2.04 x 4 mm3 voxel size, 4 mm thick slices no gap) data were acquired using 25 encod-
ing gradient directions at b-values of 0 and 1000 sec/mm2 with an acquisition matrix of
128x128. Anatomical data were acquired using a 256 x 256 acquisition matrix employing
a SPGR sequence (1 x 1 x 1 mm3 voxel size, 1 mm thick slices no gap) with TE=3.128ms,
TR=7.832ms and TI=450ms. SPM8 was used to co-register the DTI b-value = 0 data
73
to the SPGR images. A custom template was created using the 26 normal control scans.
All anatomical scans were normalized to the template using SPM8's DARTEL suite. The
custom template was then spatially normalized to MNI space. The three step transforma-
tion allows for the FA map to be warped into MNI space. Tractography was conducted in
subject space using normalized seeds inversely mapped from the MNI space and individ-
ual tracts were mapped from subject space to MNI space thereby avoiding the problems
associated with spatial warping of diusion tensors. A single set of ROIs were dened in
MNI space and applied to the normal control group whole-brain tractography to isolate
the fornix and cingulum. Four processing masks were dened (left fornix, right fornix, left
cingulum, and right cingulum) as voxels where tracts overlapped in at least 50% of the
control group. A voxel-based comparison of the t-score statistics among the FA map dif-
ference of the three dierent populations was conducted using the four masks to highlight
changes observed in the fornix and cingulum tracts.
3 Results and Conclusions
SPM8's DARTEL suite provided a much improved normalization of the aged brain pop-
ulations. Visual inspection of the ltered fornix and cingulum tracts conrmed greatly
improved alignment when compared to previous results. The t-score comparisons (Fig.
1 and 2) visually illustrate areas where FA in one population is higher than the other.
There appears to be a progressive degeneration of the fornix and the cingulum. The right
and left fornix branches both exhibit pattern of damage which begins at the hippocampal
end of the fornix and propagates toward the hypothalamus with right greater than left
reduction in FA. The cingulum shows less damage, but does appear to be show progressive
74
damage extending to the posterior end, more in right than left. The complete posterior
cingulum portion could not be reliably identied in the common space for all subjects,
suggesting a need for further improvement of normalization in AD. These preliminary re-
sults shows promise in tracking axonal damage with the progression of Alzheimer Disease.
The right greater than left reduction of FA in the fornix and cingulum is consistent with
several reports of white matter atrophy and hypometabolism in AD (e.g., Villain, N. et.
al., J. NeuroSci. 28(24):6174-6181).
75
Chapter 8
CSF Contamination Correction in DTI
Tractography of the Fornix in Elderly Subjects
S. Tsao
1
, D. H. Hwang
1
and M. Singh
1
1
Depts of Biomedical Engineering and Radiology,
University of Southern California, Los Angeles, CA, United States
Proc. Intl. Soc. Mag. Reson. Med. 18 (2010)
76
1 Introduction
The microstructural integrity of the limbic regions is frequently compromised in neurode-
generative diseases such as Alzheimer Disease (AD). A key limbic region is the fornix
located proximal to the ventricles. Given the relatively large voxel size used in most
clinical DTI acquisitions, the probability of CSF contamination in the fornix is high, of-
ten leading to interruption of tracts due to either a reduction in FA or misdirection due
to erroneous eigenvector estimation, particularly in AD where ventricles are enlarged.
FLAIR DTI has been used by many investigators to suppress CSF (e.g. Kwong et al.
MRM 21, 157-163, 1991) but at the expense of SNR and data acquisition time, and to
our knowledge, FLAIR DTI is rarely used in clinical studies. Aiming toward eventual
quantication of DTI metrics such as FA and tract density in the fornix and other limbic
pathways in AD, the objective of this work was to develop a post-processing strategy to
correct partial volume eects such that it could be used to analyze existing clinical DTI
data.
2 Methods
Existing DTI data acquired on a 3T GE Signa HDx MRI using a multi-slice, twice
refocused sequence with 128x128 single-shot EPI readout, 25 gradient directions with
b=1000s/mm2, one b=0 image, TR/TE=8000ms/86.1ms, FOV 26cm, two averages, 4mm
thick contiguous 28 slices covering the entire head in 2.03 x 2.03 x 4 mm3 voxels (scan time:
approximately 7 minutes) were used. Relying on a straightforward bi-compartmental
model (e.g., Koo et al. NeuroImage 44:136-144, 2009)
77
Figure 8.1: Corrected Primary Eigenvectors after CSF compensation correction in the
anterior portion of the Fornix (Red Original Eigenvector) (Green Corrected Eigenvector)
S =f
WM
S
WM
+f
CSF
S
CSF
(8.1)
where S
CSF
, S
WM
are the signals from a pure CSF and white matter (WM) voxel and
f
CSF
,f
WM
are corresponding fractions of CSF and WM. We can isolate the white matter
signal within any particular voxel by extending this model to the DTI signal for the 25
gradient directions given by
S
i
S
0
=
f
WM
S
i
WM
+f
CSF
S
i
CSF
f
WM
S
0
WM
+f
CSF
S
0
CSF
(8.2)
78
This allows us to extract the WM only DTI signal given by
S
i
WM
S
0WM
=
Sf
CSF
S
i
CSF
S
0
f
CSF
S
0CSF
(8.3)
As we are only concerned with the mixing between CSF and white matter in this study,
the gray matter (GM) fraction has been ignored but can be readily incorporated in a
three-compartmental model. Derivation of f
CSF
was based on using MD values of the
voxel. K-means clustering was used to separate all voxels into three clusters: CSF, WM
and GM. A linear interpolation was done between the CSF and the WM MD values as
an estimate of CSF content. To get S
CSF
and S
0CSF
, we used only the voxels from the
ventricles and with CSF content f
CSF
greater than 99%. Correction was then performed
only on voxels with f
CSF
less than 70 percent to avoid problems with overcompensation
and possible errors in our estimate of f
CSF
.
3 Results
An increase in FA was found in tract-based ROIs of the fornix after correction, due to
an increase in anisotropy after the isotropic CSF component has been removed (see table
1). Results also indicate that the primary eigenvector and therefore the tractographic
vector orientation can also been corrected (see g. 1), leading to less erroneous tracts and
allowing for the tractography of fornix at a lower FA threshold (see g. 2).
79
Table 8.1: Corrected FA scores of 4 Normal age matched elderly subjects
Subject Right Fornix Left Fornix
Corrected FA Original FA Corrected FA Original FA
1 0.4150 0.0372 0.3398 0.0219 0.4428 0.0312 0.3633 0.0210
2 0.2149 0.0182 0.1980 0.0096 0.2801 0.0355 0.2465 0.0170
3 0.2179 0.0344 0.1884 0.0133 0.1997 0.0094 0.1968 0.0058
4 0.2476 0.0370 0.2010 0.0147 0.3175 0.0405 0.2627 0.0227
4 Conclusion and Discussion
Our results suggest that WM voxels that have been contaminated by CSF can be corrected
via a simple compartmental model that boosts FA values (table 1) and corrects for primary
eigenvector orientation estimation (g. 1). This is particularly useful in aged subjects
that have high degrees of cerebral atrophy that increases CSF contamination in a large
number of voxels.
80
Figure 8.2: Corrected Tractography (RED) and uncorrected Tractography (BLUE) at FA
threshold of 0.14 of tracts of the Hippocampus (green ROI). Figure shows (1) correction
recovering the superior portion of the fornix (2) reduction of erroneous tract.
81
Chapter 9
ICA-based Multi-Fiber DWI Tractography in
Neurosurgical Planning
Sinchai Tsao*
1;2
Niharika Gajawelli*
1;2
Peter A. Michels
1;2
Darryl Hwang
1;2
Yi Lao
1;2
Fernando Yepes
1
Vidya Rajagopalan
1
Meng Law**
2
Natasha Lepore**
1;2
1
University of Southern California, Los Angeles, California, USA
2
Children's Hospital Los Angeles, Los Angeles, California, USA
*,** These authors share equal contribution
Proceedings of MICCAI 2013 DTI Challenge Workshop (In Press Springer LNCS)
82
1 Abstract
Diusion Tensor Imaging is a powerful imaging modality that allows us to investigate the
underlying white matter ber structure of the brain. However, it has many limitations,
including its inability to resolving ber crossings. Many multi-ber techniques attempt
to solve this problem, but they often require high power computation as well as High
Angular Resolution Diusion Imaging (HARDI) data [57]. Since diusion imaging has
the potential to contribute to the presurgical planning by clearly delineating white matter
anatomy and integrity, it is important to determine methods that solve the crossing-ber
problem in the clinical setting. Here, we explore the use of a recently developed tractog-
raphy technique that utilizes Independent Component Analysis (ICA)-based multi-ber
orientation estimation. This ICA-based technique leverages neighborhood voxel informa-
tion for ber estimation and has been shown to function especially well with clinical data
that have small number of gradient directions and b-values, using limited computation
power and time [58, 59]. We were able to reconstruct estimations of the white matter
corticospinal tracts from both of the MICCAI 2013 Challenge neurosurgical cases. Pa-
tient 1 demonstrates penetration of the corticospinal tract by the mass, suggesting that
complete retraction of the glioblastoma would result in severe distruption of the right
CST. Any remaining right CST related function could be lost. Patient 2 shows some
displacement of the right CST. When comparing to the contralateral left CST, the right
CST seems to be relatively intact. The mass also seems to have displaced a small portion
of the right CST. In conclusion, the ICA-based method performed relatively well in both
patients, clearly delineating the CSTs relative to their respectives masses. Our method
also showed marked improvements when compared with the single-ber DTI model.
Keywords: MRI, ICA, DTI, Tractography, DWI, Neurosurgery, Multi-Fiber
83
2 Introduction
Radiological examinations play a crucial role in surgical planning and guidance. Magnetic
Resonance Imaging (MRI), in particular, is of considerable use due to the lack of ionizing
radiation, and can serve not only for treatment planning but also for evaluating outcomes.
In surgical oncology, preoperative radiological examinations not only provide information
about the extent of masses as well as indication of the degree of necrosis and edema,
but also give a measure of tissue integrity not observable intraoperatively. In particular,
preservation of white matter connectivity can severely aect postoperative outcomes in
terms of both behavior and function. In this challenge, we attempt to mediate postopera-
tive motor function loss by accurate delineation of the corticospinal tracts using diusion
weighted imaging and tractography.
Diusion weighted imaging (DWI) is a unique modality in medical imaging which
enables us to model the diusion pattern of water in the brain. The most commonly
used DWI method is diusion tensor imaging (DTI), in which rank 2 tensors or 3 x 3
matrices are computed at each voxel using the diusion weighted images. The primary
eigenvector indicates the primary direction of unrestricted diusion in a particular voxel
and is thought to be aligned with neuronal ber paths in the brain.
This model of diusion, however, does not take into account ber crossings which have
been hypothesized to occur in the majority white matter voxels in the brain[60]. Currently,
there are many other models using high b-values, and large numbers of gradient directions
such as HARDI[57] or diusion spectrum imaging[61] data, but these datasets are hard to
acquire in a clinical settings due to limitations in acquisition time. Our method attempts
to circumvent these issues by estimating multiple bers in each voxel using clinical data
84
Table 9.1: DWI Information
Patient 1 Patient 2
Number of Gradient Directions 20 20
Image Size (mmmm) 1.14 x 1.14 1.14 x 1.14
Slice Thickness(mm) 5.2 5.2
b-value (mm=s) 1000 1000
Number of Slices 30 29
Number of Repetitions 4 4
with limited b-values and gradient directions, and utilizing much lower computational
resources when compared to methods such as spherical deconvolution[62].
3 Methods
3.1 Data
The 2013 MICCAI DTI Challenge provided participants with DWI datasets from two
neurosurgical cases. Patient 1 has a recurrent/residual glioblastoma W.H.O Grade IV.
Patient 2 is diagnosed as 'Inltrated Glioma, no grade'. Diusion weighted images were
provided in the NRRD format with 20 gradient directions, each with a b-value of 1000
s=mm
2
. Patient 1's image had 30 slices and Patient 2 had 29 slices. The acquisition was
repeated 4 times with voxel dimensions: 1.14 mm x 1.14 mm x 5.2 mm. Co-registered
anatomical images as well as a labeled tumor volume were also provided. Two additional
datasets of Patient 3 with diusive low grade WHO II glioma were released for further
85
evaluation. The Patient 3 data was aquired at 2 timepoints, 6 weeks and 2 weeks prior to
surgery and acquired with 30 gradient directions and 20 gradient directions respectively.
3.2 Pipeline
All data including the DWI data was rst converted to a NIFTI le format and checked
for obliqueness and consistency in orientation. Next, we computed a mean DWI volume,
using the 4 base volumes, as well as brainmask with AFNI[63]. We then use Indepen-
dent Component Analysis to recover the directions of individual bers from a mixture
of bers within a voxel. Display and render of tracts as well as ltering was done using
TrackVis[64].
3.3 Independent Component Analysis
Consider the diusion eld of a ber as the source, and the voxel as a sensor; in this
formulation each voxel is considered to be "sensing" signals from multiple sources. The
implementation of our multi-ber "unmixing" paradigm is based on the paper by Singh et
al 2010[58]. The attenuated diusion signal from each of the gradient directions, X
i
can
be written written as X
i
= S
i
=S
0
, where i is each gradient direction, S
i
is the diusion
weighted image and S
0
is the non-diusion weighted scan.
In using ICA, a few assumptions were made: (1) Fibers cross through multiple voxels
maintaining their orientation and (2) Fiber signal sources are independent. The latter
assumption is made considering the rather large size of voxels (on the order of mm)
compared to the order of diusion of water molecules in axons (on the order of 10um).
Unless the diusion time is extremely long, molecules diusing around axons of one ber
will not have a signicant interaction with those diusing around axons of another ber.
86
Therefore similar to the conventional ICA construction, assuming we havek bers crossing
in a small neighborhood ofn voxels, we can formulate the bers ask independent sources,
s
1
;:::;s
k
from x
1
;:::;x
n
measurements:
x
i
=
k
X
j=1
a
ij
sj8i = 1; 2;:::;n (9.1)
where a
ij
is the mixing fraction of j-th source in the i-th voxel.
The aim in ICA is to nd a linear transformation of the dependent sensor signalx
i
that
makes the outputs as independent as possible. In vector-matrix notation: x = Asor s = A
1
x.
Each linear combinationy
j
=
P
w
ij
xi would be an estimate of a single source if each row
of matrix A
1
is w
ij
. ICA assumes that the sources are non-Gaussian and Singh et
al 2010[58] showed that the S
i
=S
0
distributions of a single ber are approximately non-
Gaussian. Therefore ICA should correctly assume that the sum of non-Gaussian variables
would be increasingly more Gaussian than our original source. From this fact, we postu-
late that the W
ij
that maximizes the non-Gaussianity of yi give an accurate estimation
of the orientation of our sources. To solve for this linear transformation, we started with
a random set of w
ij
, and used a center voxel surrounded by n 1 voxels to maximize
the non-gaussianity of y
j
with the fastICA algorithm[65], which relies on negentropy to
maximize non-gaussianity. After computing the sources using ICA, DTI processing was
done to calculate the tensors and estimate the ber directions.
87
Table 9.2: Tractography Parameters
Type: Streamline
FA Stopping Threshold: 0.02
MD Stopping Threshold: 4.00
Track Step Length: 0.1 mm
Max Number of Steps: 4000
Maximum Change in Angle: 45

Number of Neighborhood Voxels used in ICA Estimation: 8
3.4 Tractography
Whole-brain streamline tractography was performed from the ICA ber estimations. At
each point, the consecutive direction was estimated as the weighted sum of correlated
directions from neighborhood voxels. The parameters used were as follows:
The ROIs used for tractography ltering includes a slice lter placed in the z-direction,
to ensure tracts meet a certain length, a slice lter placed in Y-plane at the middle of
the volume, and a large sphere lter posterior to the brainstem, both to remove spurious
tracts.
4 Results
Figures 9.1 and 9.3 show the results of our ICA-based multi-ber tractography for Patient
1 and 2, respectively. For comparison, we ran the exact same streamline algorithm but
instead used the primary eigenvector orientation from single-ber DTI as our estimate of
ber orientation in Figures 9.2 and 9.4. We rendered the tracts in front of a coronal color
88
Figure 9.1: Patient 1 CSTs using multi-ber ICA-based tractography with coronal slice
showing color FA; without tumor volume (L) and with tumor volume (R).
FA slice to give a sense of overall orientation in the brain. The tumor volume is rendered
as a semi-opaque volume on the right.
5 Discussion and Conclusion
This paper demonstrates the performance of ICA-based DWI tractography of the CST in
two neurosurgery cases. Patient 1 has a tumor that is positioned central to the CST tract
and is more challenging than Patient 2. In the former case, the dierences between single-
ber tractography versus multi-ber ICA-based tractography is more obvious. However,
upon closer inspection the dierences between single-ber tractography versus multi-ber
tractography become increasingly stark in the peritumoral regions as well as regions of
known crossings near the corpus collosum, cerebral cortex as well as lower down in the
brain stem.
89
Figure 9.2: Patient 1 CSTs using single-ber DTI tractography with coronal slice showing
color FA; without tumor volume (L) and with tumor volume (R)
Figure 9.3: Patient 2 CSTs using multi-ber ICA-based tractography with coronal slice
showing color FA; without tumor volume (L) and with tumor volume (R)
90
Figure 9.4: Patient 2 CSTs using single-ber DTI tractography with coronal slice showing
color FA; without tumor volume (L) and with tumor volume (R)
The authors believe that the multi-ber neighborhood technique may have mitigated
some of the eects of edema from the tumor. This is possible because the edema compo-
nent can be seen as a independent source of signal consistent across neighborhood voxels
close to the tumor. We also believe that the large 5.2 mm slice thickness may be causing
severe partial volume eects. The non-isometric voxel sizes may also cause directional
biases downstream from the tensor estimation and may aect tractography results. In
previous work with non-isometric voxels, the authors observed dierential performance
from the ICA methodology between isometric and non-isometric voxels.
More work needs to be done to determine whether the increased signal-to-noise ratio of
the larger slices outweigh the distortion eects it may have on multi-ber and streamline
tractography techniques that rely on neighborhood information. The geometry of the
tracts in question may also determine the optimal geometric conguration of voxel sizes
and location. In this case, the CST tractography may seem denser simply because of
91
the direction bias in the Ventral-Rostral direction caused by increased slice thickness
in the Z-direction and may be mistaken for improved results. This can be seen in the
comparison between our single-ber and multi-ber methods in patient 2, where the ICA-
based methods seems to have decreased ber density, but where there is also less noise at
the peritumoral regions as well as the end points of the tracts.
In conclusion, this paper (1) demonstrates the application of a computationally thrifty
multi-ber orientation estimation methodology suitable for neurosurgical planning appli-
cations, (2) revisits and examines its performance against traditional tractograpy using
DTI's primary eigenvector but in a clinical neurosurgical setting and (3) aims to be
the rst step towards ne tuning both imaging as well as tractographic parameters for
eventual vetting of the ICA methodology as a viable multi-ber tractographic tool for
neurosurgical planning.
Acknowledgements
The authors would like thank their colleagues Alec CW Wong, PhD and Bryce Wilkins for
all their eorts in developing and testing ICA-based tractography. This work would not
be possible without their tireless dedication to science. We would also like to dedicate this
publication in memory of their late advisor Professor Manbir Singh, PhD who championed
the ICA methodology and whose guidance we miss dearly daily.
92
Chapter 10
The Power of Hybrid / Fusion Imaging Metrics
in Future PACS Systems: A Case Study into the
White Matter Hyperintensity Prenumbra using
FLAIR and Diusion MR
Sinchai Tsao
a;b
Samantha J. Ma
a;b
Peter A. Michels
a;b
Niharika Gajawelli
a;b
Meng Law
a
Helena Chui
a
and Natasha Lepore
a;b
a
University of Southern California, Los Angeles, California, USA
b
Children's Hospital Los Angeles, Los Angeles, California, USA
Submitted SPIE Medical Imaging 2014
93
1 Abstract
Most white matter related neurological disease exhibit a large number of White Matter
Hyperintensities (WMHs) on FLAIR MRI images. However, these lesions are not well
understood. At the same time, Diusion MRI has been gaining popularity as a powerful
method of characterizing White Matter (WM) integrity. This work aims to study the
behavior of the diusion signal within the WMH voxels. The goal is to develop hybrid
MR metrics that leverage information from multiple MR acquisitions to solve clinical
problems. In our case, we are trying to address the WMH penumbra (as dened by
Maillard et al 2011[66]) where WMH delineates a foci that is more widespread than than
the actual damage area presumably due to acute in
ammation. Our results show that
diusion MR metrics may be able to better delineate tissue that is in
amed versus scar
tissue but may be less specic to lesions than FLAIR. Therefore, a hybrid metric that en-
codes information from both FLAIR and Diusion MR may yield new and novel imaging
information about the progression of white matter disease progression. We hope that this
work also demonstrates how future PACS systems could have image fusion capabilities
that would be able to leverage information from multiple imaging series to yield new and
novel imaging contrast.
Keywords: DTI, Diusion MRI, MRI, Alzheimer's Disease, Dementia, WMH, White
Matter, FLAIR, Neuroimaging
94
2 Purpose
White matter hyperintensity (WMH) burden as detect on FLAIR MR images has been
associated with cognitive impairment[67]. However, the degenerative process that leads
to such lesions is not well understood. In this study, we investigate the diusion proper-
ties of white matter lesions using diusion MRI at 3T. To this end, a segmentation and
coregistration routine was developed to segment WMHs and coregister them to diusion
measurements. The objective of this work was to characterize the diusion characteris-
tics of WMHs to allow for more accurate staging of white matter disease as well as to
dierentiate between early-stage lesion formation and end-stage white matter scars.
Table 10.1: Summary of Subject Demographics
Subject WMH Voxels Total WM Voxels Percent WMH Sex Age
1 6,230 1,168,404 0:53% M 73
2 1,411 1,113,728 0:13% F 70
3 9,767 1,146,941 0:85% F 86
4 13,546 1,334,280 1:02% M 75
5 1,282 1,096,588 0:12% F 78
3 Methods
Diusion tensor imaging (DTI) data was acquired on a 3T GE Signa HDx MRI using a
multi-slice, twice refocused sequence with 128x128 single-shot EPI readout, 25 gradient
directions with b=1000s/mm2, three b=0 images, TR/TE=8000ms/86.1ms with 2.03 x
2.03 x 2 mm3 voxels. 3D T1-weighted and FLAIR images from 5 cognitively normal aged
95
Figure
1
–
Data
for
subject
with
very
high
WMH

load.
(A)
FA
image
for
WMH
candidate
region

mask.
(B)
Histogram
fitting
for
dynamic

thresholding.
(C)
WMH
image
(Red).
(D)
MD
and

FA
histograms
in
WMH
regions
vs.
all
WM
regions.

Figure 10.1: Data for subject with very high WMH load. (A) FA image for WMH
candidate region mask. (B) Histogram tting for dynamic thresholding. (C) WMH image
(Red). (D) MD and FA histograms in WMH regions vs. all WM regions.)
subjects. Initial preprocessing involved several standard toolkits available in the public
domain. FreeSurfer was used for intensity non-uniformity correction and the removal of
non-brain tissues. SPM8 was used to coregister the DTI b=0 data to the T1-weighted and
FLAIR images. A WMH candidate region mask was generated using an in-house written
MATLAB code that was driven by DTI fractional anisotropy (FA) measurements. WMH
segmentation was processed in the WMH candidate region using dynamic thresholding
based on histogram analyses (see Fig. 10.1: A, B, C). Diusion measurements were taken
96
in the WMH regions and compared to those measured in all white matter regions of the
brain.
Figure
2
–
MD
and
FA
histograms
for
low
WMH

load
and
high
WMH
load
subjects
compared
with

MD
and
FA
in
all
WM
for
each
subject.

Figure 10.2: MD and FA histograms for low WMH load and high WMH load subjects
compared with MD and FA in all WM for each subject.
4 Results
For subjects with low WMH load, mean diusivity (MD) values within the WMH regions
were distributed dierently in the histogram compared to MD values in all white matter
97
(WM) regions. The histogram of WMH FA values for these subjects maintained a similar
distribution to all WM regions. Histograms for subjects with high WMH load exhibited
the opposite relationship (see Fig. 10.2). For the subject with a very high WMH load,
the MD and FA histograms for WMH regions were similar (see Fig. 1 D).
5 Conclusion
Our results suggest that early-stage lesion formation is may be characterized by MD
changes, while regions with long-term lesions display FA changes. End-stage WM scars
exhibit similar DTI measurement distributions, indicating stabilization of the tissue.
This is consistent with previous observations of what investigators have termed WMH
penumbra[66], where the WMH delineated a foci of more widespread and subtle white
matter damage that can be characterized with Diusion MRI.
Acknowledgements
This work has been supported by USC ADRC's National Institutes of Aging Program
Grant 5P01AG012435-18 as well as National Institute of Biomedical Imaging and Bio-
engineering grant 5R21EB013456-02.
98
Chapter 11
FreeSurfer Parcellation of Brains Containing
Large Infarcts
Niharika Gajawelli
1
Sinchai Tsao
1
Darryl H. Hwang
1
Bryce Wilkins
1
Stephen Kriger
2
Susanne Mueller
2
Diana Truran
2
Meng Law
3
Helena Chui
4
Michael Weiner
2
Manbir Singh
1;3
1
Department of Biomedical Engineering, University of Southern California,
Los Angeles, California, United States
2
Center for Imaging of Neurodegenerative Diseases, VA Medical Center,
San Francisco, California, United States
3
Department of Radiology, University of Southern California,
Los Angeles, California, United States
4
Department of Neurology, University of Southern California,
Los Angeles, California, United States
Resubmission in Progress
99
1 Abstract
Purpose: To compensate for imaging changes in areas of cerebral infarct to facilitate
successful brain segmentation in FreeSurfer.
Materials and Methods: A non-linear co-registration method to and from an atlas
space allowed for voxel intensities from the contralateral hemisphere to ll in areas of
cerebral infarct in order to facilitate successful segmentation. A validation study was
conducted using T1 volumes of 5 normal age-matched subjects, with known he FreeSurfer
analysis, were simulated with cerebral infarcts from 8 stroke patients.
Results: Before application of our method, FreeSurfer had a completion rate of 50%
with our data. Application of our method yielded 100% completion rate in FreeSurfer.
Conclusion: Our method allows for the inclusion of data with cerebral infarcts into
automated brain segmentation work
ows.
Keywords: FreeSurfer, stroke, infarct, parcellation
100
2 Introduction
Volumetric analyses of brain images play an important role in investigating cerebral dis-
eases that aect brain structure such as Alzheimers disease or Vascular dementia in elderly
subjects (1). Brain volumes decrease with age, and changes in particular regional volumes
may be associated with certain neurological disorders. For example, it is shown that the
hippocampal volumes are particularly aected in patients with Alzheimers disease (2).
Segmentation of brain volumes into various regions requires a good deal of anatomical
knowledge and has traditionally been done manually by experts trained in this area (3).
However, this process is tedious, requires the resources of skilled experts, and is prone to
human error, making it unfeasible for large scale studies, especially those that require the
processing of data of hundreds of subjects.
As an alternative to manual segmentation, there exist several software packages that
facilitate automatic brain segmentation. Many packages are capable of segmenting the
brain volume into white matter, gray matter and cerebrospinal
uid, but only a few
are capable of segmentation and parcellation into more specic regions. FreeSurfer (4),
frequently used in the neuro-imaging community, is a robust, validated package that
utilizes volumetric methods for the cortical and sub-cortical segmentation of structural
MRI. In the processing pipeline, FreeSurfer incorporates prior probabilities of tissue classes
(the probability that an anatomical feature exists at a certain global position), as well
as the relative spatial positioning pattern of various tissues for the segmentation and
parcellation processes (5), (6). Therefore, the classication of various brain regions is
highly dependent on the integrity of neighborhood voxels and their structure, which leads
to frequent failure of FreeSurfer processing in the presence of pathologic lesions or cerebral
101
infarcts often present in elderly subjects. For these volumes, not only structures within,
or in the vicinity of infarcts are aected, but distal regions may also be misclassied.
Classication of brains with infarcts is indeed important in order to gather information
pertinent to brain volumes with stroke. In addition, neglecting such data would not
only accrue unnecessary costs, but would also hinder analysis for regions not aected by
the stroke. Eorts have been made to segment stroke volumes using multimodal image
information (7); however, this information may not always be available. An important
pre-processing step for large-scale studies is registration to a common template. Brain
volumes with stroke have a signicant amount of missing tissue as well as sulcal and
morphological variability, thus, aecting the registration precision (8), (9). Registration
of brain volumes with lesions using SPM (Statistical Parametric Mapping) software can be
facilitated by lling-in the lesion with tissue from the contralateral hemisphere (10). The
accuracy of this method, however, depends on the symmetry of the brain, and therefore
is impractical in large-scale studies involving brain volumes of the elderly.
Taking into account the issues stated above, Solodkin et al. (11) proposed lling-in
the stroke region using tissue from the homologous region of the contralateral hemisphere
as follows. First, a mask was created by manually outlining infarct boundaries. This mask
was used to acquire voxels from the contralateral hemisphere to transplant and morph
into the infarct region. The alignment to the region surrounding the infarct was improved
by adding manually traced sulci in the hemisphere with the infarct. The brain was then
transformed into the Talairach space, where hemispheres were separated, interchanged,
and rigidly registered to the ipsilateral hemisphere before being inversely deformed back
into the native subject space. The scar lines created due to lack of continuity of tissue were
removed using suturing software and an expanded lesion mask. Validation in this study
102
was done by comparing voxel dierences between hemispheres, under the assumption that
the overall structure would be the same for both hemispheres in a normal brain.
In this paper, we have demonstrated a similar technique that simplies the lling
process by completely eliminating the labor-intensive manual sulci-drawing step. It is
shown below that generating a lled volume by automatically replacing the region from
the contralateral hemisphere produces results reasonably similar to the original volume.
To validate our technique, we manually simulated strokes in brain volumes of healthy
elderly normal subjects and compared ratios of the cortical and sub-cortical volumes to
their ground truth respectively, before and after lling the stroke. This approach requires
minimal manual intervention and therefore facilitates analysis of large-scale volumetric
studies requiring FreeSurfer parcellation of cortical/sub-cortical regions. In this method,
we used SPM as it is a widely accepted MATLAB based tool, which can be used on any
operating system.
As subjects undergoing an MRI scan are typically not positioned exactly symmetri-
cal with respect to the midline of the brain, identication of homologous regions in the
contralateral hemisphere requires an accurate realignment and spatial normalization of
the brain. Spatial normalization of T1-weighted images being subjected to FreeSurfer
analysis was accomplished by SPM ane and non-linear transformations (12) that map
the coordinates of the subjects brain to the MNI template. First, the infarct boundaries
were outlined manually to create a mask. The infarct was then lled-in with the mean
intensity of the gray and white matter of the contralateral hemisphere of the T1 brain
volume, as without such uniform lling, normalization often failed (10). The template
used for spatial normalization was the MNI152 template acquired from the Montreal
103
Neurological Institute website. The result of the normalization in MNI space was man-
ually compared to the template to check the correspondence of major features. After
satisfactory correspondence between the two brain volumes was conrmed, the left and
right hemispheres of the brain were interchanged (in MNI space) and normalized a second
time. This step is equivalent to co-registering the contralateral hemisphere to the infarct-
containing hemisphere and reduces the inter-hemispherical dierences from morphological
distortions likely to be caused by the infarct.
The volume was then mapped back to the subject space using the inverse deformation
eld. Finally, voxels corresponding to the infarct were inserted into the original subject
data from this inverse mapped image. The output after each step along with the procedure
is shown in Figure 1.
3 Validation
To validate our technique, cerebral infarcts from 8 stroke patients were simulated one
at a time in T1 volumes of 5 normal elderly subjects in whom the FreeSurfer analysis
was known. The normal elderly brains were determined by visually conrming no strokes
were present and by checking the CDR scores of the subjects. Insertions of infarcts
into the normal brains were done in normalized MNI space to account for the variances
in brain symmetry and shape in native space. To simulate a realistic case, the infarct
intensities were normalized to match that of the volume they were being placed into.
After infarct insertion was completed in the MNI space, the inverse deformation eld was
applied to transform the volume back into the subjects native space. The infarct region
was then transferred and placed into the original healthy brain. This nal volume was
104
used as the starting point to apply the steps of our lling technique described earlier in
the Methods section. To avoid inaccuracies caused by interpolation, we avoided directly
using the deformed volumes. FreeSurfer was then run on all the original normal elderly
brain volumes, all the un-lled volumes generated after various stroke insertions, and all
of the lled volumes after implementing our approach. The outputs from all studies were
evaluated quantitatively by comparing ratios of the cortical and sub-cortical volumes to
their ground truth respectively, before and after lling the stroke. The regions entirely
or partially within the infarct mask, as well as regions not relevant for this study such
as ventricles and vesicles were discarded from the nal cortical and sub-cortical volumes.
This was done by using the FreeSurfer command mri segstats with the {mask option to
nd regions that are partially or completely included in the mask. The mean and standard
deviation of the ratios for the various brain regions acquired from the 40 validation studies
were computed and the histograms were t to a Gaussian distribution for comparison.
4 Results
An example of FreeSurfer brain parcellation in a validation study from a subject with a
relatively large infarct is shown in Figure 2. Signicant changes were seen before and after
lling-in the infarcted region. Fig. 2(a) shows an example of a simulated stroke/infarct
region. This particular infarct was in the right pre-frontal region and therefore results
for the right hemisphere are displayed. The regions encompassed within the infarct are
the Caudal Middle Frontal Gyrus, the Precentral Gyrus, the Superior Frontal Gyrus,
the Insula, and the Parsopercularis. Fig.2 (b) displays the ground truth parcellation,
while (c) the result after inserting this infarct. The yellow circles indicate regions outside
105
the infarcted region that show high deviation from the ground truth. The similarity
of the parcellation between the ground truth and the result after lling, shown in (d) is
remarkable. Brain volumes inserted with the same infarct had in general many overlapping
areas, which were eliminated when calculating the aggregate ratios for analysis. Regions
that had a partial infarct were also eliminated and the average was computed only with
volumes that did not have any infarct in the regional volume. Fig. 3(a) shows the ratios
for one simulation study using one infarct in ve dierent normal brain volumes. Less
variation exists between the parcellated brains as exhibited by the smaller error bar for the
lled case and the case when smoothing is also applied after lling. Histograms generated
from averaged ratios of all combined simulation studies (8 brains x 5 lesions) for the lled,
lled and smoothed, and unlled cases are t to a Gaussian distribution and shown in Fig.
4 (b). The lled volumes have a mean slightly closer to 1.0 and about 30% smaller standard
deviation, re
ecting the closeness of the reconstruction to the ground truth. Although
the mean for all cases is close to 1, it is important to note that larger strokes limit
the reconstruction aecting distal regions as well as the proximal, causing the standard
deviation for the unlled brain volume ratios to increase. A comparison of our SPM-based
approach was made to using the symmetric dieomorphic mapping normalization option
(SyN) included in ANTs (Advanced Normalization Tools) software (13). ANTs provides
open-source functionality for deformable normalization with large deformations. SyN was
placed as the top performer amongst 14 other tools in an independent evaluation of ANTS
normalization tools (14) and therefore is used for comparison.
For this evaluation, the brain volume was rst normalized to a template space and
then separated into the left and right hemispheres and inverse normalized to the subject
space. The hemisphere without the lesion is then normalized to the hemisphere with the
106
lesion using the SyN option in the ANTS tool. The procedure outlined in the paper by
Solodkin et. al (11) was followed except for the manual delineation of sulci. FreeSurfer
was run again on these volumes and volume ratios from various regions of the original to
the lled using our SPM based method was compared to that using ANTs dieomorphic
mapping. For this comparison, fewer subjects were used (4 infarcts in 3 dierent normal
brains) but as shown in Figure 4, the histogram and the distributions are very similar,
with the standard distribution acquired by ANTS being slightly better. Our approach,
when applied to the stroke subject volumes, works well even in brains with larger lesions.
Although for this comparison, the ground truth is unknown, FreeSurfer parcellations and
segmentations are manually examined for quality and there improvement noticed after
lling in the infarct region. The results of infarct lling in a stroke subject are shown in
Figure 5. The dotted yellow circle indicates the stroke and we can clearly observe that
there is a signicant dierence in some of the distal as well as proximal volume regions
before and after the technique is applied.
5 Discussion
Out of 16 stroke brain volumes investigated in this study, FreeSurfer failed to complete
processing on half of them before our technique was applied. All 16 volumes were pro-
cessed successfully after lling. Therefore, our SPM-based approach is useful as a simple,
automated pre-processing step before running FreeSurfer analysis for brain volumes con-
taining infarcts. This method is very important in AD volumetric analysis as it would
help quantify the regional volumes more accurately. Our FreeSurfer comparisons were
also done with infarcts lled with uniform grayscale intensity, which was the mean of the
107
gray and white matter. However, the accuracy of the output varied widely depending
on the stroke location (i.e. brain volumes with infarcts located in regions with greater
intensity changes had inaccurate segmentation/parcellation). The location and size of the
stroke is a limitation to be considered in future work. These factors aect the amount
of sulci the stroke spans and may limit the accuracy of FreeSurfer results as our current
process does not correct for the mismatch of sulci.
6 Conclusion
In conclusion, the results above suggest that our lling approach is a promising technique
to conduct accurate volumetric analyses of brain structure even in the presence of cerebral
infarcts, uses the widely accepted SPM software, requires minimal manual interaction,
and is thus suited to automate processing of large-scale clinical studies. The methods
discussed in this paper enable us to improve the ability of template methods to work
when there exist pathologies that either remove or distort the brain in a major way. This
is not limited to stroke and can be utilized for brain tumors, surgical removal, traumatic
brain injury, as well as other neurological pathologies. Also, this method could be used
to improve the functionality of any template based method with not only FreeSurfer but
other software as well.
7 Acknowledgements
I would like to thank all members of the Biomedical Imaging Lab in the University
of Southern California and the VA Medical Center in San Francisco for their help and
support.
108
Figure 11.1: Diagram showing the process for lling in the brain infarct and the output
after each step.
109
Figure 11.2: FreeSurfer parcellation result for validation study. (a) T1 slices showing
simulated infarct; (b) Ground truth FreeSurfer parcellation; (c) FreeSurfer parcellation
before lling simulated stroke; (d) FreeSurfer parcellation after lling. Circles highlight
regions where signicant distortions have been corrected after lling.
110
Figure 11.3: Ratios of various volumes of the FreeSurfer parcellation of the simulated
infarct with respect to the ground truth. (a) Averaged volumes ratios of the before and
after lling of cortical and subcortical regions in one simulated lesion using 5 dierent
normal brain volumes. The red curve is the lled, the blue the unlled and the pink the
lled and smoothed. The red and pink plots are comparable. The gures on the lower
rows are the Histograms and Gaussian ts of volume ratios of the (b) original to the lled
brain, (c) original to the lled and smoothed brain, and (d) original to the unlled brain.
Figure 11.4: Comparison of ratios of various volumes in the FreeSurfer parcellation after
lling in with the (a) SPM based method and (b) ANTS based method. The results are
quite similar with both showing a mean close to 1 and similar standard deviation.
111
Figure 11.5: Comparison of FreeSurfer parcellation before and after lling in the stroke
using SPM based approach. Dotted yellow circle indicates regions where infarct was
present. Distal regions as well as proximal regions are observed to be aected signicantly.
112
Part D
Statistical Analysis and Machine
Learning
113
Chapter 12
Is Greulich and Pyle Atlas still a Good Reference
for Bone Age Assessment?
Aifeng Zhang
a
, Sinchai Tsao
a
, James W. Sayre
b
, Arkadiusz Gertych
a
, Brent J. Liu
a
,
H.K. Huang
a
a
Image Processing and Informatics Lab, Department of Radiology, ISI/USC,
Marina Del Rey, CA USA 90292;
b
Department of Biostatistics and Radiological Sciences, UCLA School of Public Health,
Los Angeles, CA USA 90095-1772
Medical Imaging 2007: PACS and Imaging Informatics, edited by Steven C. Horii, Kather-
ine P. Andriole, Proc. of SPIE Vol. 6516, 65160T, (2007) 1605-7422/07/$18 doi:
10.1117/12.710160
114
1 Abstract
The most commonly used method for bone age assessment in clinical practice is the book
atlas matching method developed by Greulich and Pyle in the 1950s. Due to changes in
both population diversity and nutrition in the United States, this atlas may no longer be a
good reference. An updated data set becomes crucial to improve the bone age assessment
process. Therefore, a digital hand atlas was built with 1,100 children hand images, along
with patient information and radiologists readings, of normal Caucasian (CAU), African
American (BLK), Hispanic (HIS), and Asian (ASI) males (M) and females (F) with ages
ranging from 0 18 years. This data was collected from Childrens Hospital Los Angeles. A
computer-aided-diagnosis (CAD) method has been developed based on features extracted
from phalangeal regions of interest (ROIs) and carpal bone ROIs from this digital hand
atlas. Using the data collected along with the Greulich and Pyle Atlas-based readings and
CAD results, this paper addresses this question: Do dierent ethnicities and gender have
dierent bone growth patterns? To help with data analysis, a novel web-based visualiza-
tion tool was developed to demonstrate bone growth diversity amongst diering gender
and ethnic groups using data collected from the Digital Atlas. The application eectively
demonstrates a discrepancy of bone growth pattern amongst dierent populations based
on race and gender. It also has the capability of helping a radiologist determine the
normality of skeletal development of a particular patient by visualizing his or her chrono-
logical age, radiologist reading, and CAD assessed bone age relative to the accuracy of
the P&G method.
Keywords: Bone Age Assessment, Digital Hand Atlas, Web-based application, Computer-
aided-diagnosis
115
2 Introduction
Pediatric bone age assessment is an important procedure that allows a physician to judge
the degree of a patients skeletal bone development relative to the patients chronological
age. Discrepancies between bone age and chronological can occur due to growth disorders
that can stem from endocrine disorders or malnutrition. Accurate diagnosis is required
to help determine proper treatment. The most commonly used method for bone age
assessment in clinical practice is the book atlas matching method developed by Greulich
and Pyle [80, 81]. This method is based on visual comparison of the patients hand x-
ray with images collected in the atlas. The closest match is subjectively selected by the
radiologist and yields the bone age of the patient (Figure 1). However the book atlas
has not been updated since its initial publication in the early 1950s. The data collected
in the original atlas was derived solely from upper middle class Caucasian populations
residing in the midwest of the United States. Due to changes in both population diversity
and nutrition in the United States, this atlas may no longer be a good reference. An
updated data set has become crucial to improve the bone age assessment process. For
this purpose, left hand X-ray lms of normal children along with pertinent data related to
patient growth factors were acquired at Childrens Hospital Los Angeles. [82, 83, 84] The
data was divided into eight categories by race and gender. A computer aided diagnosis
(CAD) method was developed based on features extracted from phalangeal regions of
interest (ROI) and carpal bone ROI. [85, 86, 87, 88, 89] For data analysis, a web-based
statistical visualization application was developed to help address the following issues: 1)
Does each race have distinct bone growth patterns? and 2) More importantly, is the G &
P atlas still a good reference for bone age assessment of todays children?
116
Figure 12.1: Bone age assessment procedure in clinical practice.
3 Methods and Materials
3.1 Data Collection
Building of the Digital Hand Atlas A digital hand atlas was developed with data from
eleven hundreds children. This consisted of hand images collected from Childrens Hospital
Los Angeles accompanied by patient information and radiologists readings. The data
consisted of normal Caucasian (CAU), African American (BLK), Hispanic (HIS), and
Asian (ASI) males (M) and females (F), with ages ranging from 0 18 years. [82, 83, 84]
Each hand image and the subjects demographic data, along with two radiologists readings
and CAD-based bone age were populated into a mysql database. The database was
integrated with a web server at the Image Processing and Informatics Laboratory, USC.
To simplify statistical testing as well as data visualization, the two radiologists readings
117
Figure 12.2: Sample Hand Image from the CHLA data collection process. All pertinent
bones that were used as ROIs in the CAD program are labeled.
118
Figure 12.3: Simple GUI to visualize dierences between the dierent Study Groups.
were averaged to get one reading-based age per subject. Figure 2 is a sample hand image
with phalangeal and carpal ROIs outlined.
3.2 Data Visualization - Web GUI (Graphical User Interface) Develop-
ment
A separate JAVA applet-based platform-independent software was developed to assess
relevant data in the Digital Hand Atlas database and to plot results. The GUI programs
119
Figure 12.4: Clinical GUI to compare a patients CAD and Radiologists Readings relative
to the accuracy of the Radiologists Readings from the Digital Hand Atlas Database. The
three black horizontal lines bisected by the vertical line marks the mean as well as the
upper and lower standard deviation of the pediatric radiologists G&P-based readings
of normal cases with that specic chronological age. The yellow marker represents the
radiologist reading of the current case. The red marker represents the CAD output of
the current case. The GUI allows the user to assess the CAD and Pediatric Radiologists
readings against the accuracy of the gold standard - the accuracy of G&P readings of
the normal cases in the database. Any reading variation away from the chronological age
that is greater than the accuracy as set by the gold standard indicates a possible growth
disorder.
were developed to compare Pediatric Radiologist Readings, CAD results with Chronolog-
ical Age amongst dierent racial and gender groups. The rst GUI program (see gure
120
Figure 12.5: The gure illustrates that gender eects bone age assessment with both
reading and CAD methods. (a) Caucasian male; (b) Caucasian female. The x-axis is
chronological age and the y-axis is either radiologists reading or CAD result.
2) was used as a tool to plot graphs with the users choice of Radiologists Readings, CAD-
based result or Chronological age as the X or Y axis. The JAVA applet runs on a web
server of choice and connects to a mysql database on the same machine for its input data.
[90, 91] A php script that connected to the mysql database was used as an intermediary
data source for the JAVA applet. The JAVA applet passes the necessary inputs to the
php script, which then retrieves the relevant information from the database without com-
promising data integrity. One feature of this design is that as the data is updated in the
database, the graphing output will change accordingly. In this study, four comparison
studies based on the specic population combinations were conducted using this GUI:
1) All eight categories are individually compared, 2) Male and Female are combined to
form four ethnic groups for comparison, 3) All male and all female are combined to form
121
Figure 12.6: The gure illustrates that ethnic or racial origin eects bone age assessment
with both the reading and CAD methods. (a) Asian; (b) Caucasian. The x-axis is
chronological age and the y-axis is either radiologists reading or CAD result. The x-axis
is chronological age and the y-axis is radiologists reading.
Figure 12.7: The divided age sets in evaluation. (a) for female and (b) for male.
two gender groups for comparison, and 4) All eight categories were combined to form one
group.
122
Figure 12.8: Plots of readings between race pairs (a) Asian female and Caucasian female;
(b) Hispanic female and African American female. The x-axis is chronological age and
the y-axis is radiologists reading.
3.3 Prototype of a Clinically Applicable GUI
To demonstrate the clinical viability of the CAD methodology, a second GUI was devel-
oped to allow the user to visualize the current patients data against information in the
database. The mean and standard deviation of the radiologists readings of the normal
cases in the database for that particular chronological age is shown along with the current
patients results. If the CAD assessed bone age or the Pediatric Radiologist Readings falls
outside of the range of error, the patient could have a case of suspected abnormal bone
development. (see gure 3)
123
4 Results
4.1 Plots
The plotting application allows users to visualize the normal data from the digital hand
atlas, by race, gender or combination among race and gender. Figures 4-5 shows the
output of the web applet when using CAD and Readings against Chronological Age.
However, the user can compare a particular patient using any three inputs: Chronological
age, Radiologists readings and CAD assessed bone age. In this section, general trends
shown from the graphical plots will be outlined. Statistical studies will be outlined in the
following section. Comparing radiologists readings and CAD results versus chronological
age, the graphs show that ethnicity and gender does have in
uence on the growth pattern
as shown by the variability of both CAD results as well as radiologists reading in gure
4. Variation in both the radiologists reading and the CAD results between the dierent
gender (gure 6) and dierent ethnic groups (gure 5) show that both ethnicity and gender
have an independent eect on bone age assessment. With gender, it is well documented
that females bone structure variations end earlier, making it more dicult to dierentiate
between the later age groups. However, the use of dierent ethnic groups / populations
is novel. Assessing the G&P-based Radiologists readings, it can be seen that it correlates
well with chronological age (see gure 7), except at later ages. This can be primarily
attributed to the lack of variation in older females (see gure 6). The CAD method
sensitivity decreases at age 15+ (see gure 7). This is attributed to the lack of variation
in the current ROIs. At the time of writing of this paper an addition ROI, the Radius,
with known variation from 15-19 years is being added to remedy this problem.
124
4.2 Statistical analysis between races
To study the discrepancy between races, paired-samples t-test was performed between
the chronological age and the average of two readings. Table 2 shows the mean dierence
between average reading and chronological age for eight categories by race and gender.
The numbers with an asterisk represents a dierence with a signicant p-value of <
.05. As we can see from Table 2, radiologists using the G&P atlas are able to assess
bone age very accurately for the African American sample population and the Caucasian
male sample population. However, signicant discrepancies were observed in Asian and
Hispanic groups. To study this in greater detail, the entire age range was divided into
four age groups. The Males and Females were divided in dierently according to Figure 6.
Comparisons of the dierence between radiologists average reading and chronological age
for two races was conducted by analysis of variance (ANOVA). The results for the four age
sets are presented in Table 3 for female and Table 4 for male. As the data in Table 3 shows,
signicant mean dierences of readings between races are observed at age group from 10-
14 years old. Radiologist over-read Asian cases by approximately 0.80 years compared
with their African American and Caucasian peers. In the same fashion, over-reading was
observed between the Hispanic and African American groups. Figure 6 shows the plots
of readings between the Asian female population and the Caucasian female population
as well as the Hispanic female and African American female populations. For the male
populations from the digital atlas, signicant dierences were the groups ranging from 11
to 16 years of age. In comparison with the African American population, Radiologists
tends to over-read Asian and Hispanic cases about a year. (Figure 7)
125
5 Conclusions
The aforementioned study represents a novel statistical analysis of bone age assessment
for diering gender and ethnic populations. Through the study we attempted to answer
the following questions: 1) Does each race have distinct bone growth patterns? and 2)
Is the G & P atlas still a good reference for bone age assessment of todays children?
Addressing the rst question, the hypothesis is that genetic dierences with in
uences
from diet and nutrition may cause variation in growth patterns, thereby in
uencing the
bone age assessment process. The statistical analysis (Section 3.2) discovered signicant
dierences for Asian and Hispanic ethnic populations, which to a certain extent nullies
the null hypothesis. This therefore illustrates that the accuracy of the Greulich and Pyle
atlas can indeed be improved by taking ethnic population and gender into account. This
potential can be captured by a CAD trained to adapt to reading images of dierent
genders and populations as previously outlined. The answer to the latter question is that
until this CAD methodology is complete and compared against the Greulich and Pyle
method, the extent of which patient demographic background can be exploited to give a
more accurate bone age assessment cannot be fully determined.
Acknowledgements
This work has been supported by NIH R01 EB 00298
126
Table 12.1: Total of 1,103 cases from rst Cycle Data Collection
Age group/Category ASIF ASIM BLKF BLKM CAUF CAUM HISF HISM
00 1 2 4 5 3 3 1 4
01 5 5 5 5 5 5 5 5
02 5 5 5 5 5 5 5 5
03 5 5 5 5 5 5 5 5
04 5 5 5 5 5 5 5 5
05 5 5 5 5 5 5 5 5
06 5 5 5 5 5 5 5 5
07 5 5 5 5 5 5 5 5
08 5 5 5 5 5 5 5 5
09 5 5 5 5 5 5 5 5
10 10 10 10 10 10 10 10 10
11 10 10 10 10 10 10 10 10
12 10 10 10 10 10 10 10 10
13 10 10 10 10 10 10 10 10
14 10 10 10 10 10 10 10 10
15 10 10 10 10 10 10 10 10
16 10 10 10 10 10 10 10 10
17 10 10 10 10 10 10 10 10
18 10 10 10 10 10 10 10 10
Total 136 137 139 140 138 138 136 139
127
Table 12.2: CAD evaluation for eight categories by race and gender. Values in the table
represent the mean dierence between chronological age and the average reading in year.
Signicant mean dierence is indicated by an asterisk.
* p value < 0.05
Table 12.3: Cross racial comparison of chronological age vs. reading for female. Each
block was divided into four subblocks, each of which encompasses one age set. (unit: year,
-: over read)
* p value < 0.05
128
Chapter 13
Automated Bone Age Assessment of Older
Children using the Radius
Sinchai Tsao
a
, Arkadiusz Gertych
b
, Aifeng Zhang
a
, Brent J. Liu
a
, H.K. Huang
a
a
Image Processing and Informatics Lab, Department of Radiology, ISI/USC,
Marina Del Rey, CA USA 90292;
b
Department of Surgery, Minimally Invasive Surgical Technologies Institute,
Cedars-Sinai Medical Center, 8700 Beverly Blvd, Davis 4021, Los Angeles, CA 90048,
USA
Medical Imaging 2008: PACS and Imaging Informatics, edited by Katherine P. Andriole,
Khan M. Siddiqui, Proc. of SPIE Vol. 6919, 69190E, (2008) 1605-7422/08/$18 doi:
10.1117/12.770018
129
1 Abstract
The Digital Hand Atlas in Assessment of Skeletal Development is a large-scale Computer
Aided Diagnosis (CAD) project for automating the process of grading Skeletal Develop-
ment of children from 0-18 years of age. It includes a complete collection of 1,400 normal
hand X-rays of children between the ages of 0-18 years of age. Bone Age Assessment is
used as an index of skeletal development for detection of growth pathologies that can be
related to endocrine, malnutrition and other disease types. Previous work at the Image
Processing and Informatics Lab (IPILab) allowed the bone age CAD algorithm to accu-
rately assess bone age of children from 1 to 16 (male) or 14 (female) years of age using
the Phalanges as well as the Carpal Bones. At the older ages (16(male) or 14(female) -19
years of age) the Phalanges as well as the Carpal Bones are fully developed and do not
provide well-dened features for accurate bone age assessment. Therefore integration of
the Radius Bone as a region of interest (ROI) is greatly needed and will signicantly im-
prove the ability to accurately assess the bone age of older children. Preliminary studies
show that an integrated Bone Age CAD that utilizes the Phalanges, Carpal Bones and
Radius forms a robust method for automatic bone age assessment throughout the entire
age range (1-19 years of age).
Keywords: Bone Age Assessment, CAD, Image Processing, Skeletal Imaging
130
2 Introduction
2.1 Current Issues in Bone Age Assessment Methods
Bone Age Assessment is used in Pediatrics to determine the stage of a subjects bone
maturity during its growth from 0 to 19 years of age. Delayed or accelerated bone growth
relative to a childs chronological age may be an indication of a pathological growth pat-
tern. Examples of diseases that may cause pathological growth patterns include growth
hormone deciency, hypothyroidism, Sotos syndrome and other genetic and non-genetic
pathologies. The Greulich and Pyle atlas published in 1950 form the basis of clinical bone
age assessment. Radiologists use the hand radiographs of children from the GP atlas
along with descriptions of hand bone growth stages to assess a patients bone age. Other
more recent but less commonly used methods include the Tanner and Whitehouse or TW2
Method developed in 1975. However both methods require the Radiologist to compare
and manually judge stages of growth relative to an atlas. Our CAD method attempts
to quantify features that correlate well with bone maturity and growth and completely
automate the process of bone age assessment.
2.2 Previous Work
Previous work by Pietka, Gertych, Pospiech-Kurkowska, et al [92] used the Phalanges
as the ROI for feature extraction. The dynamic changes in the Phalanges in children
between the ages of 7 to 16 (male) or 14 (female) years of age provided information for
bone maturity assessment within this age range. Later work by Zhang et al [89] using size
and shape-based features of the Carpal Bones allowed for assessment of younger children.
To allow for complete assessment of Bone Age in all children, another ROI was necessary.
131
The Radius was chosen as the ROI of choice for older children because of its late fusion
of the proximal Metaphysis and distal Epiphysis. This phenomenon is documented in
the Greulich and Pyle Atlas as being later in the Radius as compared to the Phalanges.
Thus allowing for assessment in the degree of fusion at the older age range using wavelet
transformation based features similar to those used in the Phalanges. Due to the diering
morphology between the Phalanges and the Radius, however, a new methodology is still
necessary to extract the smaller sub-image of the growth plate region that will be used
for feature extraction via wavelet transform. This segmentation technique is based on
knowledge about the generalized morphology of the wrist. Using this knowledge along
with dened landmarks, the Radius is identied and the bounding box for the sub-image
is segmented from the larger wrist ROI.
3 Methods
3.1 Localizing the Radius
This initial methodology was developed based on previous work in the phalangeal ROI in
the hand. The wrist region is segmented from the overall image via a set of landmarks
based on the location of the Phalanges and the hand midline. The wrist ROI is then
extracted as shown in gure 1. Through textural analysis, the pixels are grouped into
three subtextures (Bone, Soft Tissue and Background) as shown in the second image in
gure 1. The bone-based texture is then removed from the other two textures. A binary
image shows the location of all pixels that are part of that texture group. Using knowledge
of the morphology of the wrist, the Radius is selected from all other bone structure and
an overall outline of the bone is formed using image processing techniques such as erosion
132
Figure 13.1: Outline of Image Processing algorithm to extract Growth Plate ROI from
the larger Wrist ROI.
and dilation. In addition, knowledge of the location of the Epiphysis and Metaphysis of
the Radius allows missing smaller structures that may be identied as separate objects
to be merged to form the image of the Radius as seen in image number 4 in gure 1.
Using the width and centerline of the Radius, a box containing the growth plate is drawn
133
Figure 13.2: Enlarge Image showing the extracted Growth Plate ROI. Note the horizontal
pattern that denotes the boundary between the epiphysis and metaphysis. This boundary
lessens and eventual disappears as the child grows.
automatically by the CAD algorithm. The subimage extracted from the segmentation
process illustrated by gure 1 is shown in gure 2.
3.2 Features Extraction
After the growth plate ROI is segmented (gure 2), the sub-image is then passed onto
the features extraction algorithm, which calculates the energy, orientation and number
of edges in the sub-image and provides 12 quantitative measurements using the wavelet
transform developed for the phalanges [92]. Six of the features that show most correlated
variation with bone maturity are selected and passed on to a Mamdani fuzzy inference
system. For each feature, a set of membership functions are formed from training using
48 images from males of ages 14 to 18 years of age. Using statistical analysis of feature
values between age groups of 14, 15,16,17,18 and 19 years, groups that showed too much
variation were merged.
134
3.3 Fuzzy Classication and Computation of Bone Age
The feature set for the images are divided into the two male and female cohorts for
bone age computation, since the growth patterns and in male and female children are
dierent[93]. For each feature, a set of 3 to 6 Gaussian membership functions were formed
using the standard deviation and means from the data. Each of these membership func-
tions represented a specic age group and depending on the feature value will contribute
to each of the computed bone age output dierently. A set of simple aggregation rules
dictated how the membership functions were aggregated into the output membership
function. The output membership functions were also Gaussian-based and depended on
the standard deviation and mean of the chronological age of the 149 image training set.
For our preliminary system testing we used the same image set for training and testing.
Figure 13.3: Mamdani Fuzzy Inference System used to aggregate the feature values and
derive a Bone Age based on feature values. The system requires training with a data set
of normal children.
135
Figure 13.4: Feature values for the male cohort.
Figure 13.5: Feature values for the female cohort.
4 Results
With our initial test group of Caucasian Children between the ages of 14 and 19, there
were 149 images that were used for testing. The chronological age of these children does
136
Figure 13.6: CAD Age versus Pediatric Radiologist's Reading for the Male Cohort
Table 13.1: Mean Absolute Error of the system for the Male Cohort using Phalanges and
Radius based on the Chronological Age and Reading of the Pediatric Radiologist as the
gold standards. The absolute mean dierence for the male cohort between the Pediatric
Radiologists reading and the normal childs chronological age is 0.650 years.
Phalanges Radius
ChrAge 1.246 yrs 0.814 yrs
Reading 1.030 yrs 0.918 yrs
not dier more than 3 years from the projected bone age that the Pediatric Radiolo-
gist determined. The Caucasian Children also provide us with homogeneity in ethnic
background in our dataset in case there are dierences in growth patterns due to ethnic
background. From previous experience we also found that the Caucasian dataset has
137
Figure 13.7: CAD Age versus Chronological Age for the Male Cohort
Table 13.2: Mean Absolute Error of the system for the Female Cohort using Phalanges
and Radius based on the Chronological Age and Reading of the Pediatric Radiologist
as the gold standards. The absolute mean dierence for the female cohort between the
Pediatric Radiologists reading and the normal childs chronological age is 0.700 years.
Phalanges Radius
ChrAge 1.583 yrs 1.156 yrs
Reading 1.181 yrs 0.908 yrs
the most consistent in image quality. Our initial results show that we were able to cor-
rectly segment out the growth plate region approximately 50% of the time. Errors were
a result from (1) an inability to dierentiate between the carpal bones and the radius
138
Figure 13.8: CAD Age versus Pediatric Radiologist's Reading for the Female Cohort
(2) mistaking the Ulna for the radius (3) unable to locate the growth plate sub-region
of the carpal bone accurately. More work needs to be done to improve the accuracy of
segmentation or have fail safe mechanisms when the Radius or the growth plate region
cannot be segmented. In the nal integrated system, images where the Radius cannot be
used will rely on the phalangeal data for bone age assessment. The results of the feature
extraction are shown in gure 3 and 4, correlation coecients are calculated from either
a linear regression or exponential regression depending on the feature and were indicated
by either POLY for linear polynomial or EXP for exponential. We chose to use both mea-
sures because the fuzzy inference system was able to mimic both types of trends during
training. The trends indicate that the features allow for dierentiation between the lower
139
Figure 13.9: CAD Age versus Chronological Age for the Female Cohort
age range of 14 to 15 and the higher age range of 17 to 19 years of age but not any higher
in age range resolution. Figures 6 to 9 shows our preliminary results of the system after
the fuzzy inference engine has computed the bone age based on the radius features. The
computed CAD bone age for the new Radius results and the previous Phalangeal results
are compared to the chronological age as well as the Radiologists Readings. Tables 1 and
2 show the generalized error for the overall testing set. Relative to chronological age, the
new system showed a decrease of about 0.3 years for males and 0.4 years for females.
5 Discussion
The GP atlas acknowledges natural growth variation in males from 14 to 17 years of age
to have a standard deviation of 10.72 to 13.05 months. [93] Therefore, any variation of
140
our result to within the order of 1 year can be attributed to natural variation in the
population. In addition, our current system based on the Phalanges tapers o at 16 years
of age for females and 17 years of age for males. Therefore any ability to dierentiate
between the pre-16 or pre-17 years of age and the higher age group of 18 and 19 will
be contributed signicantly to the overall fuzzy bone age result. This is because nal
system integration will allow the Phalangeal features to be aggregated with the Radii
features and weighed using the fuzzy inference system to yield a nal CAD bone age
result. The R-squared correlation coecients for Radius are much lower than that of
the Phalanges (approx 0.8 and above) for their optimal operational age range from 5 to
14 years. The computed bone age results show the eects of the lack of correlation in
the features. Relative to the gold standard of the chronological age, it shows a stepwise
output, with many of the computed bone ages clustered at 15.5 years and 17.5 years. The
system however still show improvements over the Phalanges in this age range but most
likely re
ect the fact that Phalanges have already fully developed at this age range and
no longer have enough variability for bone age assessment.
6 Future Work
The current image segmentation techniques have a high failure rate (50 %), improvements
can be made looking at current error cases. Common problems include mistaking the Ulna
for the Radius and having an incomplete segmentation of the Radius. Adding improve-
ments to the current image segmentation algorithm will decrease the failure rates and
increase the accuracy of the segmentation. The features are highly sensitive to whether
141
the demarcation of the epiphysis and metaphysis is captured therefore better segmenta-
tion will yield better feature values. For CAD system to have accuracy over the entire
0-19 year age range, the radii information has to be integrated with the current system
that uses the carpal and phalangeal information for Bone Age Assessment. How this will
be done and to ensure maximum reliability and performance has yet to be determined.
The diculty lies in the fact that at lower age ranges the current Radii methodology will
not work as a method to determine bone age, similarly the carpal data will not work
on older children, therefore, a reliable method must be developed to cover the whole age
range by eciently aggregating information from each of the dierent regions.
Acknowledgements
The authors would like to thank National Institutes of Health for their support of this
research as well as the Society of Imaging Informatics in Medicine for their continuing
support for this project through their imaging informatics training grant.
142
Chapter 14
Evaluating the Predictive Power of Multivariate
Tensor-based Morphometry in Alzheimers
Disease Progression via Convex Fused Sparse
Group Lasso
Sinchai Tsao
a
Jiayu Zhou
b
Jie Shi
b
Jieping Ye
b
Yalin Wang
b
Natasha Lepore
a
a
University of Southern California and Children's Hospital Los Angeles, Los Angeles,
California, USA
b
Arizona State University, Phoenix, Arizona, USA
Submitted SPIE Medical Imaging 2014
143
1 Abstract
Previous work by Zhou et al. 2013[100] has shown that a multi-task learning framework
can be used to encode both sparsity as well as temporal smoothness in predicting cogni-
tive outcomes of Alzheimer's Disease Neuroimaging Initiative (ADNI) subjects based on
MRI baseline features as well as other subject information. Wang et al. 2013[101] has
developed a multivariate tensor-based morphometry (mTBM) features based on para-
metric surface analysis for sparsifying machine learning methods. They have shown these
features when applied to the cortical surface outperforms standard volumetric or other
features when used for disease classication. This study aims to combine this multi-task
framework with novel multivariate tensor-based morphometry applied on to hippocampus
surface to improve prediction performance of cognitive scores 6, 12, 24, 36 and 48 months
from baseline. The hippocampus has been well studied in relation to AD and suggests
that mTBM analysis should give us a highly accurate assessment of cognitive outcomes.
We also assess the predictive power of these novel features relative to existing imaging-
based features provided by FreeSurfer as well as non-imaging features such as age, sex,
baseline test scores and genetic information.
Keywords: Alzheimers Disease, Disease Progression, Multi-task learning, fused Lasso,
ADAS-Cog, Tensor-based Morphometry, Hippocampus, Feature Selection
144
2 Introduction
Recent work in psychological testing[102], genetic studies[103], magnetic resonance (MR)
imaging[104], positron emission tomography (PET) imaging[105], cerebral spinal
uid
(CSF) measurements[106], cardiovascular status[107] and others have yielded tremendous
amounts of diagnostic data for diagnosing and staging dementias, especially Alzheimers
disease (AD). Moreover, many of these studies now also include longitudinal information[108,
102]. This has lead to a problem often referred to as the curse of dimensionality, where
the size (number of dimensions) of the dataset make it dicult to do various numeri-
cal analysis on the data, thereby making it to draw all-inclusive conclusions from the
dataset. Statistical analysis together with clinical disease models have helped with de-
termine how the dierent sets of diagnostic information interacts with one another but
they require a large number of ad hoc assumptions and therefore does not lend itself
well to large scale Medical Imaging-based features. These problems become even more
important when trying to use Machine Learning techniques because the predictive power
of the machine learning model decreases as the number of dimensions increases when
the number of dimensions becomes too large. The question is then about how to se-
lect the "correct" features to maximize predictive power. This paper leverages existing
sparsifying machine learning techniques with temporal priors[100], built specically for
progressive disease models, such as AD, together with multivariate tensor-based morpho-
metric (mTBM) features [101] of the Hippocampus to try and predict AD progression up
to 48 months from the baseline MRI measurement. The goal is to evaluate the predictive
power of mTBM against those of cortical thickness and other FreeSurfer-based features,
demographic information (sex and age) as well as genetic information (ApoE-4 Copies).
145
There are a variety of machine learning methods that can be used to predict cognitive
outcomes using large-scale imaging features. A number of single task techniques such
as Ridge and Lasso have been used to predict dierent future time points from baseline
data. These single task operations, however ignore, useful temporal priors such as tem-
poral smoothness in cognitive scores. Multi-task learning oers an opportunity to include
temporal information but only perform well with a small number of features. To this
end, Zhou et al 2013[100] formulated a multi-task learning framework that incorporated
temporal smoothness of the output estimates of cognitive score changes in time as well
as encodes sparsity in it's feature selection by employing a fused Lasso penalty that is
smooth in time.
M06 M12 M24 M36 M48
0
5
10
15
rMSE

without mTBM
with mTBM
Figure 14.1: Bar Chart of the rMSE of predictions with and without mTBM features by
time point
3 convex Fused Sparse Group Lasso (cFSGL)
Zhou et al 2013 [100] has proposed a powerful multi-tasked learning technique that incor-
porates sparsity as well as temporal smoothing for modeling a progressive disease model.
146
In their formulation, each tasked can be though of a single forward predictor from baseline
measurement to a measurement at a certain future time point. In their case, they used
the ADNI dataset and predicted ADAS cognitive scores 6 months after baseline (M06),
12 months after baseline (M12), 24 months after baseline (M24), 36 months after base-
line (M36) and 48 months after baseline (M48). In our study we aim to use the same
ADNI dataset but also incorporate mTBM hippocampus features and compare it to fea-
tures used in their study. We also attempt to combine the dierent feature sets to try to
evaluate the predictive power of each set of features.
The proposed cFSGL can be considered a multi-task regression problem with t time
points and fromn subjects each withd features, wherefx
1
;x
2
;:::;x
n
g represents each of
thed input features for each subject at baseline (i.e. x
i
2R
d
). Similarly,fy
1
;y
2
;:::;y
N
g
represents the target cognitive scores for each subject atN time points (i.e. y
i
2R
N
). For
a single subject (n) each task can be seen as a projection of MR / demographic / genetic
baseline measurements att = 0 represented atx
n
to a future cognitive score measurement
at time t = t
1
(e.g. at 48 months) given by y
n
(t
1
). We can extend this formulation to
a multi-task one by performing projections of all time points simultaneously. In other
words, each set of baseline measurements at t = 0 given by x
n
is projected to a vector
(R
N
withN time points) given byy
1
. The entire mapping can be summarized as a linear
operation using matrices X and Y . X and Y is formed by arranging the patient feature
space row-wise, each row being x
n
or y
N
, and yields a R
nd
X matrix and a R
nN
Y
matrix. Since this is a linear model, a set of weights W (R
dN
) is trained to map x
n
to y
n
or X to Y . To achieve a set of weights that encodes both sparsity and temporal
smoothness. The following cost function is minimized during training.
147
min
W
kXWYk
2
F
+
1
kWk
1
+
2

RW
T

1
+
3
kWk
2;1
(14.1)
wherekWk
1
is the L1-norm or lasso penalty that encodes for sparsity, kWk
2;1
=
P
d
i=1
q
P
t
j=1
W
2
ij
is the group Lasso penalty that encodes for temporal grouping of fea-
tures,

RW
T

1
is the fused lasso penalty, R = H
T
and H2R
t(t1)
where H
ij
= 1 if
i =j,H
ij
=1 ifi =j +1, andH
ij
= 0 otherwise that encodes for temporal smoothness.
Table 14.1: Results showing rMSE and R for M06 through M48 with and without mTBM
features. (* P < 10
3
)
without mTBM with mTBM
rMSE R rMSE R
M06 8.4452 0.6124* 6.5315 0.6499*
M12 8.2058 0.6989* 6.7600 0.6948*
M24 12.3285 0.6975* 6.7292 0.7704*
M36 10.4104 0.6225* 7.1673 0.7323*
M48 7.2023 0.6223 7.1232 0.6075
148
4 Multivariate Tensor-based Morphometry (mTBM) fea-
tures
Previous work by Wang et al 2013[101] has provided a strong framework for extracting
parametric surface-based features for sensitive detection of pathological changes across
multiple subjects. In this study we apply this framework but on the hippocampal surface
instead of the cortical surface.
5 Results
The results shown are from 2 simulation experiments where data from ADNI was used to
to both train and test the cFSGL model. Experiment 1 used demographic information
(Age and Gender), Freesurfer volumes and cortical thicknesses (326 features), the number
of ApoE- 4 alleles as well as a baseline MMSE score as features used in the model.
Experiment 2 added the mTBM features from each of the vertices of the hippocampus
segmentation using Freesurfer. The vertex information from mTBM was scaled down by a
factor 10 using bicubic interpolation to yield a total of 2100 features. 90 percent of the 624
subjects were used for training and the remaining 10 percent were used for testing. The
results shown are from the 10 percent of our dataset allocated for testing. We calculated
the root mean square error rMSE(y; ^ y) =
q
ky^ yk
2
2
n
as well as a the correlation coecient
between the pairs of predicted values and actual values at each of the time points.
Table 14.1 shows how predictive performance has improved by incorporating mTBM
features into our dataset.
149
6 Discussion and Conclusion
Current results are from simple single iteration experiments. The performance of cFSGL
is sensitive to model parameters and more work needs to be done in carefully choosing
optimizing model parameters. However, this work goes to show that mTBM features
has the potential to improve predictive power of cFSGL models especially for later time
points.
Whether or not this approach provides more statistical power in disease progression
research requires careful validation for each application. More importantly, we anticipate
that the mTBM and multi-task learning framework may provide new insights on imaging
biomarker and machine learning in medical imaging research. In the future, we plan to
combine them for MCI or other types of disease progression studies. If our proposed
combinational framework helps improve classication accuracy, it would then support the
further use and development of mTBM measurements together with multi-task learning
for predictive applications in the medical imaging eld.
150
−20 0 20 40
−20
−10
0
10
20
30
40
M06
Actual
Predictions
−20 0 20 40
−20
−10
0
10
20
30
40
M12
Actual
Predictions
−20 0 20 40 60
−20
−10
0
10
20
30
40
50
M24
Actual
Predictions
−20 0 20 40
−20
−10
0
10
20
30
40
M36
Actual
Predictions
−20 −10 0 10 20
−15
−10
−5
0
5
10
15
20
M48
Actual
Predictions
Figure 14.2: Results without mTBM Features - Scatter plots of Predictions versus Actual
ADAS Cognition Scores at M06 through M48
−20 0 20 40
−20
−10
0
10
20
30
40
M06
Actual
Predictions
−20 0 20 40
−20
−10
0
10
20
30
40
M12
Actual
Predictions
−20 0 20 40 60
−20
−10
0
10
20
30
40
50
M24
Actual
Predictions
−20 0 20 40
−20
−10
0
10
20
30
40
M36
Actual
Predictions
−20 −10 0 10 20
−15
−10
−5
0
5
10
15
20
M48
Actual
Predictions
Figure 14.3: Results with mTBM Features - Scatter plots of Predictions versus Actual
ADAS Cognition Scores at M06 through M48
151
Chapter 15
Mapping of ApoE4 Related White Matter
Damage using Diusion MRI
Sinchai Tsao
a;b
Darryl Hwa Hwang
a
Niharika Gajawelli
a;b
Stephen Kriger
c
Meng Law
a
Helena Chui
a
Michael Weiner
c
and Natasha Lepore
a;b
a
University of Southern California, Los Angeles, California, USA
b
Children's Hospital Los Angeles, Los Angeles, California, USA
c
Center for Neurodegenerative Disease, San Francisco VA Medical Center
San Francisco, California, USA
Submitted SPIE Medical Imaging 2014
152
1 Abstract
ApoliopoproteinE 4 (ApoE-4) polymorphism is the most well known genetic risk fac-
tor for developing Alzheimers Disease. The exact mechanism through which ApoE 4
increases AD risk is not fully known, but may be related to decreased clearance and in-
creased oligomerization of A. By making measurements of white matter integrity via
diusion MR and correlating the metrics in a voxel-based statistical analysis with ApoE-
4 genotype (whilst controlling for vascular risk factor, gender, cognitive status and age)
we are able to identify changes in white matter associated with carrying an ApoE 4 al-
lele. We found potentially signicant regions (P
uncorrected
< 0:05) near the hippocampus
and the posterior cingulum that were independent of voxels that correlated with age or
clinical dementia rating (CDR) status suggesting that ApoE may aect cognitive decline
via a pathway in dependent of normal aging and acute insults that can be measured by
CDR and Framingham Coronary Risk Score (FCRS).
Keywords: DTI, Diusion MRI, MRI, Alzheimer's Disease, Dementia, ApoE, Neu-
roimaging
2 Introduction
ApoliopoproteinE4 (ApoE4) polymorphism is the most well known genetic risk factor
for developing Alzheimers Disease[109]. The exact mechanism through which4 increases
AD risk is not fully known, but may be related to decreased clearance and increased
oligomerization of A. By making measurements of white matter integrity via diusion
MR and correlating the metrics in a Voxel-based Statistical Analysis with 4 genotype
153
(whilst controlling for vascular risk factor, gender, cognitive status and age) we are able
to identify changes in white matter associated with carrying an ApoE 4 allele.
3 Methods
59 subjects with ages ranging from 70 86 years were recruited from 3 dierent sites in
a prospective longitudinal study of vascular contributions to aging and cognitive impair-
ment, alone or in combination with AD. 4 positive (29 subjects carrying at least one 4
allele) and 4 negative (30 subjects). Vascular risk was assessed using the 10 year Fram-
ingham Coronary Risk Score (FCRS) and the subjects were categorized either as high or
low vascular risk. Dementia staging was performed using the Clinical Dementia Rating
(CDR) system. MR imaging was performed on four>3T scanners. A rigorous image har-
monization routine was performed to ensure all sequences produced comparable images
from the dierent scanners. T1 weighted anatomical images (SPGR / MPRAGE) were
used to coregister all the images to a common space using SPM's DARTEL Algorithm.
b0 maps from the Diusion images were registered to the T1 images using an ane trans-
formation. Both the ane as well as non-ane transformations were concatenated to
achieve coregistration for FA across all the subjects. Statistical Parametric Mapping was
performed using SPM8. 4 status, CDR scores, FRCS Category, Age as well as gender
were used as covariates and factors in the Design Matrix.
4 Results
ApoE-4 related loss of FA was found most prominently in the parahippocampal regions
(gure 15.1) and did not overlap with regions that correlated with Age and CDR (gure
154
0 2 4 6 8 10 12 14 16 18
0
1
2
Number of Subjects
CDR Score

No E4
E4
0 5 10 15 20 25
Low
High
Number of Subjects
FRCS

No E4
E4
0 5 10 15 20
Male
Female
Number of Subjects

No E4
E4
Figure 15.1: Demographic Distribution of the Subjects
15.3). CDR scores correlated most with the posterior fornix tracts (gure 15.3) in blue
whereas the signicant voxels for both ApoE-4 (red) as well as Age (green) resided
more in the anterior regions surrounding the fornix and corpus collosum. A hemispheric
asymmetry was observed with more age-related signicant voxels on the left hemisphere,
consistent with previous studies[109].
5 Conclusion
Our study shows promising results for separating the eects of 4, Normal Aging and
dementia severity when performing large scale MR imaging studies using Diusion MR.
Both the ApoE 4 and CDR eects (parahippocampal gyrus and fornix) are integral
components of the episodic memory system, which is highly vulnerable to neurobrillary
degeneration in AD. The involvement of the medial temporal diencephalic system in
155
Figure 15.2: SPM of ParaHippocampal FA values with ApoE-4 < no ApoE-4 with
P
uncorrected
< 0:01 and extent threshold < 10 voxels (all other covariates were set as
nuisance variables and controlled)
Vascular Cognitive Impairment is less prominent and more variable. Future work in
analyzing the complete N>200 dataset will give us the statistical power to investigate the
anatomical eects of vascular factors as well as dierentiate the cognitive eects of WM
damage by correlating individual ROIs with cognitive scores.
Acknowledgements
This work has been supported by USC ADRC's National Institutes of Aging Program
Grant 5P01AG012435-18 as well as National Institute of Biomedical Imaging and Bio-
engineering grant 5R21EB013456-02.
156
Figure 15.3: SPM of t-scores of the Posterior Cingulum (left) and Hippocampus (right)
using FA values for ApoE4 < no ApoE4 (Red), Age (Green) and CDR (Blue) with
P
uncorrected
< 0:05 and extent threshold < 20 voxels.
157
Chapter 16
Insulin sensitivity and brain reward activation in
overweight Hispanic girls: a pilot study
Tanja C. Adam
1
, Sinchai Tsao
2
, Kathleen A. Page
3
, Houchun Hu
4;5
, Rebecca E.
Hasson
6
, Michael I. Goran
7
1
Department of Human Biology,
Maastricht University, 6200 MD Maastricht
2
Department of Biomedical Engineering,
University of Southern California, Los Angeles, CA 90089
3
Department of Endocrinology, Keck School of Medicine,
University of Southern California, Los Angeles, CA 90033
4
Department of Radiology,
Childrens Hospital Los Angeles, Los Angeles, CA 90027
5
Department of Electrical Engineering,
University of Southern California, Los Angeles, CA 90089
6
Department of Family and Community Medicine,
University of California San Francisco, San Francisco, CA 94118
7
Department of Preventive Medicine,
University of Southern California, Los Angeles, CA 90033
Submitted April 2013, Under Revision in Pediatric Obesity
158
1 Abstract
Context: The alarming prevalence of childhood obesity poses a signicant health prob-
lem for the pediatric population and future generations. Insulin resistance is thought to
be a link between obesity and the associated increased disease risk. In addition to its role
as an energy regulatory signal to the hypothalamus, insulin acts to decrease food reward
possibly through a direct alteration of dopaminergic signaling.
Aim: In the current study we examined the association of insulin sensitivity (SI) with
cerebral activation in response to food and non-food cues.
Design: Twelve Hispanic girls (age: 8-11; BMI (kg/m2): 29.9 5.7) participated in two
study visits. The rst visit included a frequently sampled intravenous glucose tolerance
test. The second visit included a functional neuroimaging (fMRI) session (GE HDxt 3.0
Tesla)) with visual stimulation tasks. Blocks of images (high calorie (HC), low calorie
(LC) and non-food (NF)) were presented in randomized order.
Results: For the comparison HC with NF, SI was inversely associated with activation
in the anterior cingulate (r
2
= 0.65; p < 0.005), the insula (r
2
= 0.69; p < 0.005), the
orbitofrontal cortex (r
2
= 0.74; p < 0.005), and the frontal and rolandic operculum (r
2
= 0.76; p < 0.001). Associations remained signicant after adjustment for body adipos-
ity. No dierences in brain activation were observed comparing LC and NF. Conclusion:
Peripheral insulin sensitivity is associated with cerebral activation in response to food
cues. Insulin resistance may lead to overeating and the exacerbation of the development
of obesity through over-activation of brain reward circuits.
Keywords: insulin sensitivity, pediatrics, brain reward, childhood obesity, functional
imaging
159
2 Introduction
The alarming prevalence of childhood obesity is not only a signicant health problem in
the pediatric population, but poses a risk factor for adult morbidity and mortality as well
(Must and Strauss 1999). One of the most common complications associated with obesity
is type 2 diabetes caused in part by the eects of adiposity on insulin resistance (Boyko,
Fujimoto et al. 2000). Insulin resistance is thought to be a link between obesity and the
associated increased disease risk (Lee 2006). While body adiposity and insulin resistance
are strongly associated in the adult population, the pediatric population of overweight
and obese children still present with a wide variety of insulin sensitivity (Goran, Lane
et al. 2008). Converging evidence shows that insulin as well as sensitivity to insulin is
of major importance for homeostatic, but also for the hedonic, mesolimbic regulation of
hunger and satiety (Figlewicz, Bennett et al. 2006). In addition to being an energy regu-
latory signal to the hypothalamus, insulin acts within the central nervous system (CNS)
to decrease food reward possibly through a direct alteration of dopaminergic signaling
(Figlewicz, Evans et al. 2003). In mouse cells with intact insulin reception, insulin was
shown to increase the ring frequency of dopaminergic neurons in the ventral tegmental
area (VTA), a prominent structure of the brain reward circuit (Konner, Hess et al. 2011).
The eect was abolished, when insulin reception was inactivated. While CNS insulin
production cannot be ruled out, converging evidence suggests that the majority of insulin
enters the brain through active transport from the periphery (Banks 2004). It is therefore
not surprising that manipulations of peripheral insulin concentrations were re
ected in
CNS insulin alterations (Schwartz, Sipols et al. 1990). One hypothesis referred to as the
Central Resistance Model is that genetic or acquired resistance to adiposity-regulating
160
hormones occurs quite commonly and undermines the ability of those hormones to bio-
logically protect against the development of obesity (Schwartz and Niswender 2004). In
human research, the existence of brain insulin resistance along with peripheral insulin
resistance was suggested after the observation of a reduction in insulin induced changes
in the global cerebral metabolic rate for glucose in insulin resistant research participants
compared to insulin sensitive participants (Anthony, Reed et al. 2006). Since insulin
resistance is so closely intertwined with body fat, it has been dicult to tease apart the
individual contribution of either one to alterations in brain reward activity. More detailed
study designs comparing lean and obese individuals found specic associations of either
body mass index (BMI) or insulin sensitivity with activity in food related resting state
networks (Kullmann, Heni et al. 2012). Results show that, especially for reward related
processing, insulin sensitivity might be of particular importance. Considering the range
of insulin sensitivity despite obesity, children appear to be well suited for a more specic
investigation of the individual contribution of insulin sensitivity and body adiposity to
altered food related brain processing, specically highly palatable foods. Central insulin
resistance may contribute to altered food related reward perception even in the pediatric
population (Adam, Toledo-Corral et al. 2009). An uninhibited, overactive reward circuit
would then lead to overeating, setting children up for a chronic weight issue in the future.
Purpose of the current study was to assess the importance of insulin sensitivity for brain
reward activation in response to visual food cues in a pediatric population. It was hypoth-
esized that lower insulin sensitivity in the periphery is associated with stronger activation
of brain reward areas in response to high calorie food images compared to control images.
The association was furthermore expected to be independent of body fat mass.
161
3 Methods
3.1 Participants
Twelve Hispanic, overweight girls were recruited through
yers and presentations at
schools and boys and girls clubs in the greater Los Angeles area. Inclusion criteria for
participants were female gender, Hispanic ethnicity, dened by all four grandparents be-
ing of Hispanic descent, Tanner stage 1 or 2, a body mass index above the 85th percentile
for age and gender (Kuczmarski, Ogden et al. 2000) and right handedness. Participants
were excluded from the study if they had diabetes, suered from a major illness since
birth, had a condition known to aect insulin sensitivity or had a neurological illness.
Written informed consent was obtained from parents and youth assent from participants.
The study was approved by the USC IRB committee.
3.2 Procedures
TAfter obtaining informed consent, participants came to the laboratory for two separate
visits. The rst visit included anthropometric and body composition measurements as
well as an overnight stay in the USC Clinical trials unit, followed by a frequently sam-
pled intravenous glucose tolerance test (FSIVGTT). The FSIVGTT took place at 8 am
following a 10-h overnight fast. The second visit included a fMRI session with visual
stimulation tasks. Visits were scheduled one week apart. All imaging sessions took place
at 3pm with the participants being instructed to not eat anything after 12 noon.
162
3.3 Anthropometric measures
A pediatric health care provider conducted a medical history and physical examination
and determined maturation using the criteria of Tanner (Tanner 1981). Height was mea-
sured to the nearest 0.1cm by a wall-mounted stadiometer and weight was recorded to
the nearest 0.1kg by a balance beam medical scale. BMI and BMI percentiles for age
and gender were determined based upon established CDC normative curves using EpiInfo
2000, Version 1.1 (Flegal, Ogden et al. 2004).
Table 16.1: Anthropometric measures describing mean standard error of the means
(SE) and range (n = 10).
!
Table 1. Anthropometric measures
Mean ± SE Range
Age (years) 9.9 ± 1.1 8-11
Weight (kg) 59.3 ± 13.4 38.7 - 74.8
Height (cm) 144.06 ± 7.11 129.3 - 152.1
BMI (kg/m
2
) 29.9 ± 5.7 23.3 - 36.8
Waist (cm) 87.7 ± 11.93 71 - 103
Hip (cm) 90.8 ± 8.1 82 - 101
Fat mass (kg) 24.21 ± 8.2 13.8 -34.5
Lean mass (kg) 33.2 ± 5.8 22.8 - 41.11

3.4 Body Composition
Body composition (total body fat and total lean mass) for all participants was determined
with dual-energy x-ray absorptiometry (DXA) using a Hologic Discovery A model (82702,
Hologic, Bedford, MA).
163
3.5 Measures of Insulin Sensitivity
All girls participated in a FSIVGTT, assessing insulin sensitivity (SI), the acute insulin re-
sponse (AIR), and disposition index (DI). After a 12h fasting period intravenous catheters
were placed in an anticubital vein of both arms. Two fasting blood samples (-15 and -5
minutes) were followed by a 0.3g/kg body weight intravenous (iv) glucose administration
at timepoint 0. At 20 minutes, insulin (0.02 U/kg body weight, Humulin R; Eli Lilly,
Indianapolis, IN) was administered iv. Blood samples were taken at 2, 4, 8, 19, 22, 30,
40, 50, 70, 100 and 180 minutes. SI, AIR and DI as an indicator for -cell capacity was
calculated with MINMOD MILLENIUM 2003 (Bergman, Phillips et al. 1981). Blood
samples from the FSIVGTT were centrifuged for 10 minutes at 2500rpm and plasma
aliquots were frozen at -80 C until further analysis.
3.6 Analysis of blood parameters
Blood samples from the FSIVGTT were centrifuged and aliquots were stored at -80 C
until further analyses. Blood for glucose analysis was collected in
uoride tubes and
blood for insulin analysis was collected in lithium heparin tubes. Plasma was assayed in
duplicate for glucose using the glucose oxidase method and a Yellow Springs Instrument
2700 analyzer (YSI Inc., Yellow Springs, OH). Insulin was assayed in duplicate using an
ELISA kit from Linco (St. Charles, MO).
3.7 Food picture exposure paradigm
Design of the study was a block design consisting of two visual activation blocks and
control blocks presented in randomized order. All images were chosen based on previous
studies (Page, Seo et al. 2011) or through the International Aective Picture System
164
(Lang 1995). Activation task blocks consisted of 10 pictures of either high calorie (i.e.
pizza, cheeseburger, candy) or low calorie food items (i.e. asparagus, broccoli, carrots).
Before, between and after activation task blocks, control blocks of 10 non-food items were
presented in randomized order (i.e. book, bus, bicycle). Compared to stimulation task
items, control items were selected based on similarity in object complexity, as well as
color, texture and shape variety. All visual stimuli were presented for 2.5 seconds with
a 0.5 second inter-stimulus interval. Pictures were projected from a Macintosh computer
onto a screen placed near the foot of the scanner, where they were viewed through a mirror
mounted on the head coil. Participants were instructed to view the series of pictures with
the intention to recall them after the scanning procedure.
3.8 Imaging and statistical analysis
MRI acquisition was conducted on a 3.0 Tesla scanner (GE HDxt) equipped with an eight-
channel head coil. Foam padding and surgical tape were used to reduce motion of the sub-
jects head. A functional echo-planar T2*(BOLD)-weighted sequence (TR: 3000 ms, TE:
25 ms,
ip angle: 90

, FOV: 192 mm, matrix: 6464, voxel size: 3:43753:43755mm)
was used to acquire 50 volumes during the visual stimulus presentation. A T1-weighted
anatomical volume (TR: 500 ms, TE: 9.88 ms,
ip angle: 90

, FOV: 192 mm, matrix:
256256, voxel size: 0:85940:85945mm) was also acquired at the same slice locations
as the functional sequence to allow for functional-anatomical co-registration. Visual stim-
uli were presented by PsychToolbox for Matlab (Brainard 1997). The functional images
were corrected for motion in SPM8 (Wellcome Trust Centre for Neuroimaging, London,
UK) smoothed and co-registered to the standard Montreal Neurological Institute (MNI)
template for comparison. Through the co-registration process images were sliced into 2
165
x 2 x 2 mm isotropic voxel space using sinc interpolation. A statistical parametric map
was generated for each subject using the general linear model (GLM) with the stimulus
as the primary parameter of the model. To eliminate activation due to movement, 6
movement parameters were also included in the model. A box-car waveform was used as
the reference paradigm. For the statistical analysis, one participant was excluded due to
the inability to participate in the IVGTT. Another participant was excluded due to too
much head motion during the scanning process, leaving a total of 10 participants for the
nal analysis. Contrast images of each subject were used for group statistics calculated as
random eects analysis at the second level, which takes variances between subjects into
account. The statistical threshold used to report group activations was set at p< 0:005.
All activated regions were identied by the Automated Anatomical Labelling software
(Tzourio-Mazoyer, Landeau et al. 2002). Variables were tested for normality and log
transformed if necessary. To assess the eect of SI and fasting insulin on brain activation
in response to high calorie, low calorie and control images, conditions were contrasted in
an ANCOVA using SI and fasting insulin as a covariate. Regression analysis was utilized
to assess the association between brain activation, body adiposity, SI and fasting insulin
concentrations. First, we computed Spatial Parametric Maps produced by either high
or low calorie activation tasks versus the non-food control condition. In a next step we
performed a 2nd level analysis by comparing maps produced by high and low calorie ac-
tivation tasks. Regions of interest were established through their correlation with SI and
fasting insulin (p< 0:005; uncorrected) and a cluster threshold of 10 voxels.
166
4 Results
Mean values (SD) of anthropometric measures and range are described in Table 1.
Three of the participating girls were in Tanner stage 1 and seven were in Tanner stage
2 at the time of study completion. All girls were of Hispanic descent (as dened by all
four grandparents identifying as Hispanic) and overweight (body mass index above the
85th percentile for age and gender). Mean fasting glucose concentrations were 91.4 9.5
mg/dl (ranging from 79109.5 mg/dl). Three participants had impaired fasting glucose
(IFG) according to the standards of the American Diabetes Association and therefore
qualied as pre-diabetic (2000). Fasting insulin concentrations were on average 26.69
19.6 U/ml (ranging from 3.5250.90 U/ml).
4.1 Brain responses to visual stimulation and SI
Signicant dierences in BOLD activation in response to high calorie foods and non-
food pictures (Table 2) were observed in the left caudate, left orbitofrontal cortex, the left
insula, the frontal and rolandic operculum, and the left anterior cingulate. SI was inversely
associated with activation in insula (r
2
= 0.69; p< 0.005, Figure 1A), orbitofrontal cortex
(r
2
= 0.74; p< 0.005), anterior cingulate (r
2
= 0.65; p< 0.005, Figure 1B), frontal, and
rolandic operculum (r
2
= 0.76; p < 0.001). When comparing high calorie food images
with low calorie food images, signicant dierences in BOLD signal were observed for the
putamen and the insula. More specically, SI was negatively correlated with activation in
putamen (r
2
= 0.74; p< 0.005, Figure 2A) and insula (r
2
= 0.66; p< 0.005, Figure 2B).
Body adiposity was signicantly associated with activation in all of the areas previously
listed, but was not associated with SI (p > 0.05). In a next step body adiposity was
167
considered as a covariate in the regression analysis. With body adiposity as a covariate,
SI stayed independently associated with activation in all areas listed above. No dierences
in brain activation were observed comparing low calorie food with non-food images.
4.2 Brain responses and fasting insulin
Signicant dierences in BOLD activation were observed comparing high calorie food
images with non-food control images. Fasting insulin concentrations were related to ac-
tivation in the caudate ((r
2
= 0.70; p < 0.005; Figure 3A) and the insula (r
2
= 0.69; p
< 0.005; Figure 3B). No dierences in activation were observed in areas associated with
reward comparing high calorie food images and low calorie food images, nor did we nd
dierences comparing low calorie food images with non-food control images. Body fat was
not associated with activation in caudate and insula comparing high calorie food images
and non-food control images.
Table 16.2: Comparison of regional brain activation in response to high calorie food >
non food stimuli in 10 overweight Hispanic girls.
!
Table 2. Increases in regional brain activity in response to high calorie pictures vs. control pictures
Reference coordinates
x y z kb Hemisphere Max z Effect size P < 0.005
Caudate -20 22 16 10 L 3.27 1.03 0.002
Insula -32 -24 22 12 L 3.12 0.98 0.002
Orbitofrontal cortex (OFC) -48 24 -4 34 L 3.45 1.09 0.001
Insula 38 -32 6 82 R 3.15 0.99 0.001
Insula -46 4 -6 37 L 3.02 0.95 0.002
Rolandic operculum -56 10 4 17 L 3.18 1 0.001
Anterior Cingulate -4 42 4 36 L 2.92 0.92 0.004
168
!
!
!
!
!
!
!
!
Figure 16.1: Association of insulin sensitivity (SI), insula activation (parameter estimates,
PE) (Figure 1A (top): r2 = 0.69; p< 0.005) and left anterior cingulate activation (Figure
1B (bottom): r
2
= 0.65; p < 0.005), contrasting cerebral activation in response to high
calorie visual stimuli > non-food visual stimuli.
5 Discussion
Results of the current study show that even in children peripheral insulin sensitivity ap-
pears to be associated with the cerebral response to food cues. Stronger cerebral activation
was associated with lower insulin sensitivity and higher fasting insulin concentrations, par-
ticularly in brain areas associated with motivation and reward. Insulin sensitivity was also
inversely associated with activation of gustatory (frontal operculum) and somatosensory
(rolandic operculum) regions encoding food taste and texture. Our results lend support
169
!
!
!
!
!
!
!
!
!
Figure 16.2: Association of Insulin Sensitivity, putamen activation (parameter estimates,
PE) (Figure 2A: r
2
= 0.74; p < 0.005) and insula activation (Figure 2B: r
2
= 0.66; p
< 0.005), contrasting cerebral activation in response to high calorie visual stimuli > low
calorie visual stimuli.
to the idea of a co-occurrence of peripheral and central insulin resistance, which was pre-
viously suggested by other authors (Anthony, Reed et al. 2006). Insulin has been long
known as an important signal for homeostatic food intake regulation. With the increasing
utilization of brain imaging technology it became apparent, that the disruption of insulin
signaling may play an important role for overeating and the development of obesity, i.e.
through a disruption of insulin signaling in the brain reward circuit.
170
!
!
!
!
!
!
!
!
!
!
Figure 16.3: Association of fasting insulin concentrations, caudate activation (parameter
estimates, PE) (Figure 3A: r
2
= 0.70; p < 0.005) and insula activation (Figure 3B: r
2
=
0.69; p< 0.005), contrasting cerebral activation in response to high calorie visual stimuli
> non-food visual stimuli.
In rodent studies the intraventricular administration of insulin and leptin decreased
sucrose self-administration, a task frequently utilized to measure food reward (Figlewicz,
Bennett et al. 2006). After receiving a high fat diet for ve weeks, the ability of both
adiposity signals to decrease food reward was abolished, supporting the idea of central
resistance to the same.
Strikingly, the high fat animals did not gain weight dierent than the chow-fed control
rats, excluding weight as the responsible variable for the results.
171
It has been dicult in the past to tease apart whether associations between obesity and
brain reward activity are due to obesity or are due to the insulin resistance accompanying
obesity. Insulin sensitivity is strongly related to body weight and insulin resistance is
often thought of as the linking factor between obesity and disease risk (Cruz, Shaibi et al.
2005). There are a few recent studies that support the idea of insulin sensitivity aecting
the brain reward response to food cues, rather than obesity alone (Kullmann, Heni et
al. 2012; Jastrebo, Sinha et al. 2013). Given the fact that children present with a
wide range of insulin sensitivity despite obesity, a pediatric group of research participants
seems to be an ideal group to investigate the importance of insulin sensitivity, obesity, or
both for brain reward signaling.
Body adiposity was also positively related to activity in all of the listed areas of the
brain. However, when statistically considering body adiposity in the relationship between
insulin sensitivity and brain reward activity, we nd the relationship to be independent
of body adiposity.
It is an ongoing discussion if obesity is associated with increased or decreased activity
in response to food cues and the question if obese individuals eat for compensation of
reward deciency or due to over-activity of reward signaling, remains unanswered.
Traditionally, insulin resistance has been approached as a consequence of obesity.
Studies demonstrating the consequences of disrupted insulin signaling suggest however,
that overeating and obesity might very well be a consequence of insulin resistance, forming
a vicious cycle that becomes increasingly hard to break.
Body mass and body adiposity have been well investigated regarding their association
with cerebral activity in response to food cues. However, reports on the direction of the
172
association have been controversial. While one group described a positive BMI depen-
dent activation of regions associated with food reward comparing visual high calorie with
neutral stimuli (Rothemund, Preuschhof et al. 2007), others nd negative associations of
BMI and brain reward area activation (Stice, Spoor et al. 2008). So far, controversial re-
sults often have been explained by the concept of sensitivity to reward, a personality trait
that is strongly implicated in the risk for addiction and overeating (Davis, Strachan et al.
2004). Lower cerebral reward sensitivity may be an inherent predisposition prompting
the individual for constant overstimulation in order to reach average stimulation levels.
It may however also be a long-term consequence of overstimulation through an abun-
dance of high calorie, highly palatable food, resulting in neuroadaptations, i.e. receptor
downregulation. This hypothesis nds support in studies investigating dopamine, a major
brain reward transmitter. Findings from animal studies suggest that midbrain dopamine
neurons are direct targets of insulin and leptin, and that they participate in mediating
the eects of these hormones on reward-seeking behavior (Figlewicz, Evans et al. 2003).
Several studies demonstrated a lower availability of dopamine 2 receptor (D2) in striatal
regions of obese research participants (Wang, Volkow et al. 2002) or polymorphisms asso-
ciated with the D2 receptor gene (Stice, Spoor et al. 2008). In recent studies insulin was
found to play an important role for dopaminergic signaling (Konner, Hess et al. 2011),
possibly through protein kinase B (Speed, Saunders et al. 2011). Insulin receptors are
widely distributed throughout the limbic brain. We were not able to measure dopamine
activity in our research participants, but results from other studies suggest that dier-
ences in dopamine activity and reception may in part explain the results from the current
study.
173
Functional imaging studies in children are often criticized due to the diculties of
children to not move during the scanning process. We only had to exclude one participant
due to too much head movement. Other studies support that large scale fMRI studies
are feasible in children as young as ve years (Byars, Holland et al. 2002). Therefore,
the advantage that young age oers for research regarding the cause and eect of health
issues weighs out practical issues of the study execution. Research studies in adults show
that abnormal neural responses to food cues persist in participants after weight loss and
that successful dieters are characterized by a strong prefrontal activation in response to
food cues (DelParigi, Chen et al. 2004; DelParigi, Chen et al. 2007; Bruce, Hancock
et al. 2011). The prefrontal cortex is involved in self-control and behavioral monitoring,
emotion regulation and salience attribution. In children, the implementation of inhibitory
choices regarding food intake is in part dependent on the morphological and physiological
development of the brain during childhood and adolescence (Gomez-Perez, Ostrosky-
Solis et al. 2003) and a progressively greater focal activation in the prefrontal cortex
during adolescent development has been observed regarding tasks that require inhibition
of behavioral responses (Tamm, Menon et al. 2002). Results from our study suggest
that chronic activation of the brain reward system may develop before the prefrontal
cortex is developed enough to inhibit the detrimental behavior of overeating, establishing
neurological patterns that predispose children for overweight and lifelong weight struggle.
Our study shows the association of peripheral insulin sensitivity with cerebral activation
in response to food cues independent of adiposity. The results therefore demonstrate the
importance of treatment of insulin resistance in addition to weight loss as a measure to
regulate neural responses to food cues to prevent excessive weight gain through overeating
and obesity.
174
Chapter 17
Conclusion
1 Bone Age Assessment
In my earlier work on computer vision-based Bone Age Assessment (Chapter 12), I studied
the development of the skeletal system as it progresses from infancy to adult. I used the
qualitative and descriptive information from the Greulich and Pyle Atlas, Tanner and
Whitehouse methods and translated it into extracting imaging features. In this case, the
skeletal system was relative simple and features were based on either (1) the existence
or non-existence of a particular bone (2) the size and shape of growing bones through
development and lastly, (3) the appearance of the growth plate as it starts to fuse at the
end of development. The contrast in this case was from X-ray so image intensity is related
to bone density but most of the image processing was performed on binary images. The
conclusion from this work was that the Radius bone did provide some improvement in
detecting age changes at later age ranges. However, it is not sucient since the system
still under performs at the later age ranges. Possible solutions include (1) Adding the
Ulna as another bone of interest and (2) Use the overall size of all the bones as a feature
to measure bone growth.
175
2 Obesity
The next system we probed was the interaction between the nervous system and the
endocrine system and how that aects obesity. The hypothesis was that with obese
individuals their body does not give enough positive feedback when stimulated with low
calorie foods. This feedback is believed to be attenuated by the endocrine system. The
exact mechanism of attenuation is not known but believed to be related to insulin and
blood sugar levels.
In our initial work with Tanja Adam PhD, we designed a fMRI experiment that pre-
sented visual stimulus in the form of low calorie and high calorie foods and applied fMRI
tools to the neurobiological problem of Obesity. We measured how well the T2* BOLD
signal correlated in time with the presentation paradigm convolved with the hemodynamic
response function.
The project with Kathleen Page MD showed more understanding of how this could
in
uence or how the reward pathways could in
uence functional connectivity. Further-
more, this experiment allowed us to probe whether gut response could be dierent from
a visual response. With functional connectivity we are at the forefront of looking at
down-regulation instead of up-regulation and looking at overall connectivity instead of
correlation with a visual paradigm which gives us an new viewpoint / dimension of brain
function.
We then looked at how well our fMRI experiment with visual stimulus correlated with
endocrinological factors such as insulin resistance. We found that insulin as neurotrans-
mitter in the endocrine system could have a downstream eect of how the reward pathway
reacts to low and high calorie foods.
176
3 Aging and Dementias
We then turn our attention to white matter damage and dementia. This project was
particular challenging because there are no standard methods for characterizing neuronal
damage in vivo. The main methods of imaging white matter damage has been White
Matter Hyperintensities (WMHs) using FLAIR MRI. White matter is particular dicult
because of the structured nature of the ber bundles and neurons. Histology of White
Matter post-mortem has driven how we look at in vivo ways of imaging white matter.
The histological understanding, together with our understanding of aging allows us to
postulate what may be contaminating our white matter diusion signal.
In our CSF contamination chapter, we were able to design a contamination correction
technique based on the biological phenomenon of brain atrophy where CSF partial volume
increases as the brain tissue atrophies.
Similar to the idea with insulin modulating the reward system, we postulated that
a genetic factor could be modulating white matter damage. Using a similar framework
but with the diusion MRI metric instead of the fMRI statistical parametric map, we
correlated each voxel with the subject's genetic factor status- in this case, we looked
at whether the subject is a (1) non-carrier (2) homozygote or (3) heterozygote of the
ApoE-4 gene. This allowed us to ask whether this genetic factor modulates white matter
decline. In this case we assumed that the process can be approximately linear since we
are using the general linear model but the modulation could very well be non-linear and
an extension of this work could be to try to model this interaction non-linearly.
For voxel-wise methods, detection becomes very computationally intensive since we are
correlating each voxel with each of the co-factors. To make the problem computationally
177
viable we mainly assumed linearity in modulation of voxel values by other factors such
as insulin or genetics, but we believe this a reasonable assumption since our domain
knowledge of most biological systems informs us that tissue metrics measured via imaging
tend to have monotonic modulation functions. For example, the deterioration of FA values
should be approximately linearly modulated by the genetic factor ApoE.
With features from the bone age assessment work, we have much lower dimensionality
since we are not doing voxelwise analysis, therefore we were able to use higher order
classication / correlation models such as fuzzy logic. I believe that this may be also
possible with some of the voxelwise MRI work if we are able to do some reasonable
dimension reduction using domain knowledge. However, I also believe that most changes
are dictated by monotonic modulation functions and therefore can be detected by linear
models. Moreover, even if you use higher order methods for high dimensional feature
spaces, the t is only as good as the quality of the feature space and high order methods
do not substitute for better SNR or better domain knowledge-based denition of the
feature space.
4 Generating Meaningful Imaging Contrast
Throughout this thesis we've been designing metrics or image contrasts that correlates
well with either brain function or some pathological state in the tissue. In the imaging
of obesity, we assume that the T2* Blood Oxygen Level Dependent (BOLD) activity
is correlated with brain activity but we know that there are vascular eects that could
displace the signal away from the exact location of brain activity. However, with this type
of study it is reasonable just to deconvolve the Hemodynamic Response Function (HRF)
178
which is essentially the point spread function of the BOLD eect. We then correlate
the deconvolved T2* BOLD time sequence with the time sequence of the presentation
paradigm.
If we assume that the HRF is uniform across the entire brain, we can then do the
analysis without deconvolving the hemodynamic response function and do voxelwise cor-
relations to look at functional connectivity. There still needs to be work done on correla-
tion due to vascular changes and how this aects brain connectivity. Obviously vascular
changes are part of the response to brain activity and whether that is something that
we consider actual brain activity. The actual brain activity that we want to detect is
action potentials going the neuron causing the release of neurotransmitters. However the
surrogate event is the vascular response to the need for more energy as modeled by the
HRF that is measured by the T2* BOLD signal. The question is how well this surrogate
event correlates with actual action potential based brain activity. There could be global
vascular eects that can cause this assumption to be violated. Moreover, local vascular
eects due to dierences in "plumbing" could also detect dierences that are not due to
brain activity. The connectivity to seem connected as seen by correlation but may be
purely due to vascular eect.
An extension to the fMRI BOLD experiment, especially relating to the high and
low calorie foods, another novel contrast would be the correlation or the second level
correlation of the t-scores from the T2* BOLD signal's correlation with the presentation
paradigm. The correlation of the correlation would then be the second level analysis,
where these t scores are correlated with a subject's amount of insulin sensitivity. This
ties back to the rst goal of using demonstrating how domain knowledge can be used to
build hypothesis. Our understanding of how the brain works tell us that the dierence
179
of BOLD signal correlation with the presentation of low calorie foods versus with high
calorie foods could be modulated by insulin sensitivity and we could then use this two
level statistical approach to detect this dierence.
Moving on to white matter imaging contrast with diusion MRI, once again our under-
standing of the domain knowledge allowed us to denite what meaningful contrast in this
case means. We realized that meaningful contrast in the context of white matter imaging
is the an accurate measurement of the actual bers. Certainly, the amount of CSF partial
volume may be correlated with how much white matter atrophy exists at a single voxel.
These atrophic eects tend to be non-focal and widespread in the brain. However, what
we really need in our FA maps are measurements without the CSF signal. The atrophic
eects of the brain are much better measured by morphometric methods, which looks at
are how much the brain deforms as it ages. Although we haven't looked at CSF partial
volume as a measure of brain atrophy because we aren't able to accurately measure the
CSF partial volume accurately yet. Certainly, the changes based on CSF partial volume
as measured by DTI are not useful because they are non-focal. Therefore the DTI signal
without CSF partial volume should give us more interesting information about the more
focal white matter damage. This extends to our voxelwise analysis with the genetic factor
ApoE4 which mirrors our second level fMRI analysis with insulin sensitivity, where we
would like to look for focal changes and not these brain wide eects.
This general model framework also allows us to look at multiple covariates and not
only ApoE4. With our poster in ISMRM we showed that we can separate voxelwise
correlations of white matter tissue measures such as FA with multiple covariates such
as ApoE4 and Age, giving us the possibility of separating the eects of each of these
covariates on white matter deterioration.
180
5 Pattern Detection
Last but not least we evaluate our goal of designing computation techniques for pattern
detection. With the bone age assessment, we applied a number of machine learning
techniques to the bone age assessment system. We applied ANN techniques but we found
that the fuzzy logic system was quite adequate in weighting the dierent features to give
an accurate prediction of bone age. However, we did evaluate applying Support Vector
Machine (SVM) on highly non-linear feature spaces and may be worth trying since it is
quite robust. But the common theme here is that the computation technique is as good
as the feature space and the feature space has to very robust and nely tuned for the
pattern that you are looking for. In the case of bone age assessment, we found that the
improvements in prediction power either ANN or fuzzy logic was comparable and there
is simply not enough information in the radius bone to accurate predict the last couple
of years of development. Therefore beyond better computation techniques we need more
information in our feature space.
There has also been attempts at using Massive Articial Neural Network using a whole
image as an input with many hidden layers. Not only is training such massive articial
neural network (MANN) dicult, and computationally inecient but the results often
run into problems with over tting (citation). The bone age assessment project was a
good rst step at applying computational techniques for pattern detection in medical
imaging since we have both the pediatric radiologist's reading of the bone age as well
as the normal subject's chronological age as targets. The advantage of the computation
technique in this case is also much obvious since the computer vision is known to be more
181
consistent than inter-reader variation. Moreover, computer vision does not suer from
fatigue that a human reader would.
We were also able to apply the general linear model to both the obesity imaging
fMRI work as well as the diusion MRI white matter imaging work to investigate how
non-imaging factors modulate the imaging measures. The main computational technique
capstone project was the application of convex fused sparse group lasso in predicting
disease progression. Although the technique was developed by our collaborating at Ari-
zona State University, our understanding of domain knowledge (in this case dementia)
denitely allows us to have a better idea of how to ne tune the system for maximum
performance. Moreover, it is interesting to look at how the lasso portion of the algorithm
behaves in dimension reduction in this high dimensional features space.
6 Future Work
The RadSearch project is not yet fully complete. We have not explored how the system
can be implemented whilst being HIPAA compliant.
The Obesity Neuroimaging work unfortunately is no longer on-going at USC. I still
believe that it is extremely interesting to look at the dierences between gut-based stim-
ulus response and taste-based stimulus response. I believe the gut, which is an often
neglected part of the nervous system, may hold the keys to understanding obesity and
our perception of food-based stimuli. This becomes even more compelling, especially,
when you look at development and how the gut develops by separating from the brain
stem. Furthermore, the interaction between the gut and the brain is not well known and it
would be interesting to develop fMRI / EEG techniques that could sensitive to activity in
182
that part of the nervous system. With the on-going obesity epidemic, the idea of obesity
as a pathology has been more accepted and more work needs to be done on nding the
origins of this pathology. This is predicated on seeing how the brain reacts dierently to
dierent stimulus based on dierent obesity statuses. I believe that information will help
support the increasingly accepted view that obesity is not just a lack of motivation but a
pathology of the reward pathways and thus has a lot of socioeconomic and public health
implications. Endocrine-Nervous interactions are also of interest, especially in clarifying
the causal relationship between insulin sensitivity developed and reward pathway changes.
With the diusion MRI work, our lab's ICA has the potential of changing what white
matter DTI information can be acquired during a routine clinical scan. Moreover, promis-
ing model-based reconstruction techniques using a dictionary techniques as well as work
with higher order might be necessary to understanding what is going on. Moreover, met-
rics that are more than tensor-based metrics such as FA and MD may also be helpful.
However, whatever new metrics are developed should maintain the advantages of the
tensor model; The model should be orientation invariant and scanner agnostic.
Moreover, I would still like to model the diusion signal as Wide Sense Stationary
signal (WSS) and leverage the pool of probabilistic signal processing techniques to model
the diusion signal. Moreover, Van Wedeem's work at getting gold standard high angular
resolution scans of the brain might allow us to build better model-based diusion MRI
methods.
I am also still interested in looking at functional Diusion MRI. During my PhD we
tried for 6 months to use the diusion MRI signal as the surrogate signal for neuronal
activity. Our work suggests that this might be possible although better signal processing
183
techniques are still needed to determine a robust point spread function relationship be-
tween the diusion eect and neuronal activity. Moreover, I don't believe that we have
fully ne tuned the diusion MR pulse sequence to look at the diusion time resolution
necessary for signal changes from neural activity. I also believe that although high b-value
(greater than 4000) diusion scans have very low SNR, a lot of interesting information
exists in measuring small diusion distances. Moreover, we are not unfamiliar with this
low SNR environment and is not unlike the SNR of T2* BOLD images. Furthermore, I
believe that the diusion MRI image reconstruction engine from our MRI scanners may
have ltered out some of the meaningful signal when bringing the signal from the fre-
quency to the spatial domain. Therefore, I believe that analysis using the raw k-space
data may yield better results.
With the machine learning / computational technique works, we have yet to see any
techniques in the clinical space suggests that more work done is necessary in translating
these techniques into the clinic. It is telling that the bone age assessment system has not
been applied even though we have a very low dimension feature space in this case.
184
References
[1] N. Tzourio-Mazoyer, \Automated anatomical labeling of activations in spm using
a macroscopic anatomical parcellation of the mni mri single-subject brain," Neu-
roImage, vol. 15, no. 1, pp. 273{289, 2002.
[2] D. L. Sackett, \Evidence-based medicine," in Seminars in perinatology, vol. 21,
pp. 3{5, Elsevier, 1997.
[3] V. Yaghmai, S. Kuppuswami, J. W. Berlin, and S. A. Salehi, \Evaluation of personal
digital assistants as an interpretation medium for computed tomography of patients
with intracranial injury," Emergency Radiology, vol. 10, no. 2, pp. 87{89, 2003.
[4] D.-K. Kim, S. K. Yoo, and S. H. Kim, \Instant wireless transmission of radiological
images using a personal digital assistant phone for emergency teleconsultation,"
Journal of telemedicine and telecare, vol. 11, no. suppl 2, pp. 58{61, 2005.
[5] W. M. Tellis and K. P. Andriole, \Integrating multiple clinical information systems
using the java message service framework to enable the delivery of urgent exam
results at the point of care," Journal of Digital Imaging, vol. 18, no. 4, pp. 316{325,
2005.
[6] B. Raman, R. Raman, L. Raman, and C. F. Beaulieu, \Radiology on handheld
devices: Image display, manipulation, and pacs integration issues1," Radiographics,
vol. 24, no. 1, pp. 299{310, 2004.
[7] U. Tachinardi, M. de Sa Rebelo, P. de Magalhaes Oliveira Jr, and P. Pilon, \Real
time information from bedside monitors as part of a web-based patient record.,"
in Proceedings of the AMIA Symposium, p. 691, American Medical Informatics
Association, 2001.
[8] B. J. Liu, L. Documet, J. Documet, H. Huang, and J. Muldoon, \Wireless remote
control clinical image work
ow: utilizing a pda for osite distribution," in Medical
Imaging 2004, pp. 178{184, International Society for Optics and Photonics, 2004.
[9] J. Documet, B. J. Liu, L. Documet, and H. Huang, \Wireless remote control of
clinical image work
ow: using a pda for o-site distribution and disaster recovery,"
Journal of the American College of Radiology, vol. 3, no. 7, pp. 520{527, 2006.
185
[10] J. Documet, R. Shretha, K. Wang, and L. B.J., \A vendor agnostic real-time patient
dashboard to monitor patient work
ow in a radiology department.," in Presented
at Radiological Society of North America's Annual Meeting, 2007.
[11] A. Pyrros, P. Nikolaidis, V. Yaghmai, S. Zivin, J. I. Tracy, and A. Flanders, \A
bayesian approach for the categorization of radiology reports," Academic radiology,
vol. 14, no. 4, pp. 426{430, 2007.
[12] C. Friedman, P. O. Alderson, J. H. Austin, J. J. Cimino, and S. B. Johnson, \A gen-
eral natural-language text processor for clinical radiology," Journal of the American
Medical Informatics Association, vol. 1, no. 2, pp. 161{174, 1994.
[13] Y. Huang and H. J. Lowe, \A grammar-based classication of negations in clinical
radiology reports," in AMIA Annual Symposium Proceedings, vol. 2005, p. 988,
American Medical Informatics Association, 2005.
[14] Y. Huang and H. J. Lowe, \A novel hybrid approach to automated negation de-
tection in clinical radiology reports," Journal of the American Medical Informatics
Association, vol. 14, no. 3, pp. 304{311, 2007.
[15] B. Desjardins and R. C. Hamilton, \A practical approach for inexpensive searches
of radiology report databases," Acad Radiol, vol. 14, pp. 749{56, Jun 2007.
[16] V. Bashyam and R. K. Taira, \A study of lexical behavior of sentences in chest
radiology reports," in AMIA Annual Symposium Proceedings, vol. 2005, p. 891,
American Medical Informatics Association, 2005.
[17] P. Szolovits, \Adding a medical lexicon to an english parser," in AMIA Annual Sym-
posium Proceedings, vol. 2003, p. 639, American Medical Informatics Association,
2003.
[18] C.-H. Hsiao, T.-C. Hsu, J. N. Chang, S. J. Yang, S.-T. Young, and W. C. Chu,
\Developing a medical image content repository for e-learning," Journal of Digital
Imaging, vol. 19, no. 3, pp. 207{215, 2006.
[19] R. Jakobovits, J. F. Brinkley, C. Rosse, and E. Weinberger, \Enabling clinicians,
researchers, and educators to build custom web-based biomedical information sys-
tems.," in Proceedings of the AMIA Symposium, p. 279, American Medical Infor-
matics Association, 2001.
[20] W. Ceusters, P. Spyns, and G. De Moor, \From natural language to formal language:
when multitale meets galen," Studies in health technology and informatics, pp. 396{
400, 1997.
[21] W. Ceusters, P. Spyns, and G. De Moor, \From syntactic-semantic tagging to
knowledge discovery in medical texts," International journal of medical informatics,
vol. 52, no. 1, pp. 149{157, 1998.
186
[22] C. E. Kahn and C. Thao, \Goldminer: a radiology image search engine," American
Journal of Roentgenology, vol. 188, no. 6, pp. 1475{1478, 2007.
[23] Y. Huang, H. J. Lowe, D. Klein, and R. J. Cucina, \Improved identication of noun
phrases in clinical radiology reports using a high-performance statistical natural lan-
guage parser augmented with the umls specialist lexicon," Journal of the American
Medical Informatics Association, vol. 12, no. 3, pp. 275{285, 2005.
[24] V. Bashyam and R. K. Taira, \Indexing anatomical phrases in neuro-radiology
reports to the umls 2005aa," in AMIA Annual Symposium Proceedings, vol. 2005,
p. 26, American Medical Informatics Association, 2005.
[25] P. Spyns, N. T. Nh an, E. Baert, N. Sager, and G. De Moor, \Medical language
processing applied to extract clinical information from dutch medical documents.,"
Studies in health technology and informatics, vol. 52, p. 685, 1998.
[26] P. Spyns, \Natural language processing," Methods of information in medicine,
vol. 35, no. 4, pp. 285{301, 1996.
[27] B. I. Reiner, N. Knight, and E. L. Siegel, \Radiology reporting, past, present, and
future: the radiologist's perspective," Journal of the American College of Radiology,
vol. 4, no. 5, pp. 313{319, 2007.
[28] P. S. Cho, R. K. Taira, and H. Kangarloo, \Syntactic parsing of medical reports us-
ing evolutionary optimization," in AMIA Annual Symposium Proceedings, vol. 2005,
p. 920, American Medical Informatics Association, 2005.
[29] G. Hripcsak, C. Friedman, P. O. Alderson, W. DuMouchel, S. B. Johnson, and
P. D. Clayton, \Unlocking clinical data from narrative reports: a study of natural
language processing," Annals of internal medicine, vol. 122, no. 9, pp. 681{688,
1995.
[30] G. Hripcsak, J. H. Austin, P. O. Alderson, and C. Friedman, \Use of natural lan-
guage processing to translate clinical information from a database of 889,921 chest
radiographic reports1," Radiology, vol. 224, no. 1, pp. 157{163, 2002.
[31] P. J. Kenny, \Reward mechanisms in obesity: new insights and future directions,"
Neuron, vol. 69, pp. 664{79, Feb 2011.
[32] K. Stanek, J. Smith, and J. Gunstad, \Structural and functional neuroimaging in
obesity," Brain Imaging in Behavioral Medicine and Clinical Neuroscience, pp. 193{
200, 2011.
[33] E. Stice, S. Spoor, C. Bohon, and D. M. Small, \Relation between obesity and
blunted striatal response to food is moderated by taqia a1 allele," Science, vol. 322,
pp. 449{52, Oct 2008.
187
[34] E. Stice, S. Yokum, and K. Burger, \Youth at risk for obesity show greater activation
of striatal and somatosensory regions to food," The Journal of Neuroscience, 2011.
[35] G.-J. Wang, N. D. Volkow, and J. S. Fowler, \The role of dopamine in motivation
for food in humans: implications for obesity," Expert Opin Ther Targets, vol. 6,
pp. 601{9, Oct 2002.
[36] W. D. S. Killgore and D. A. Yurgelun-Todd, \Body mass predicts orbitofrontal
activity during visual presentations of high-calorie foods," NeuroReport, vol. 16,
pp. 859{863, May 2005.
[37] W. D. Killgore and D. A. Yurgelun-Todd, \Aect modulates appetite-related brain
activity to images of food," International Journal of Eating Disorders, vol. 39,
pp. 357{363, July 2006.
[38] W. D. S. Killgore and D. A. Yurgelun-Todd, \Positive aect modulates activity in
the visual cortex to images of high calorie foods," Int J Neurosci, vol. 117, pp. 643{
53, May 2007.
[39] W. A. Marshall and J. M. Tanner, \Variations in pattern of pubertal changes in
girls," Arch Dis Child, vol. 44, pp. 291{303, Jun 1969.
[40] D. H. Brainard, \The psychophysics toolbox," Spat Vis, vol. 10, no. 4, pp. 433{6,
1997.
[41] W. D. S. Killgore, A. D. Young, L. A. Femia, P. Bogorodzki, J. Rogowska, and D. A.
Yurgelun-Todd, \Cortical and limbic activation during viewing of high- versus low-
calorie foods," Neuroimage, vol. 19, pp. 1381{94, Aug 2003.
[42] T. C. Adam, C. Toledo-Corral, C. J. Lane, M. J. Weigensberg, D. Spruijt-Metz,
J. N. Davies, and M. I. Goran, \Insulin sensitivity as an independent predictor of
fat mass gain in hispanic adolescents," Diabetes Care, vol. 32, pp. 2114{5, Nov 2009.
[43] T. C. Adam, R. E. Hasson, E. E. Ventura, C. Toledo-Corral, K.-A. Le, S. Mahurkar,
C. J. Lane, M. J. Weigensberg, and M. I. Goran, \Cortisol is negatively associ-
ated with insulin sensitivity in overweight latino youth," J Clin Endocrinol Metab,
vol. 95, pp. 4729{35, Oct 2010.
[44] T. C. Adam, R. E. Hasson, C. J. Lane, J. N. Davis, M. J. Weigensberg, D. Spruijt-
Metz, and M. I. Goran, \Fasting indicators of insulin sensitivity: eects of ethnicity
and pubertal status," Diabetes Care, vol. 34, pp. 994{9, Apr 2011.
[45] B. Bennett, D. E. Larson-Meyer, E. Ravussin, J. Volaufova, A. Soros, W. T. Cefalu,
S. Chalew, S. Gordon, S. R. Smith, B. R. Newcomer, M. Goran, and M. Sothern,
\Impaired insulin sensitivity and elevated ectopic fat in healthy obese vs. nonobese
prepubertal children," Obesity (Silver Spring), vol. 20, pp. 371{5, Feb 2012.
188
[46] P. A. M. Smeets, C. de Graaf, A. Sta
eu, M. J. P. van Osch, and J. van der
Grond, \Functional mri of human hypothalamic responses following glucose inges-
tion," Neuroimage, vol. 24, pp. 363{8, Jan 2005.
[47] P. A. M. Smeets, S. Vidarsdottir, C. de Graaf, A. Sta
eu, M. J. P. van Osch, M. A.
Viergever, H. Pijl, and J. van der Grond, \Oral glucose intake inhibits hypothalamic
neuronal activity more eectively than glucose infusion," Am J Physiol Endocrinol
Metab, vol. 293, pp. E754{8, Sep 2007.
[48] S. Vidarsdottir, P. A. M. Smeets, D. L. Eichelsheim, M. J. P. van Osch, M. A.
Viergever, J. A. Romijn, J. van der Grond, and H. Pijl, \Glucose ingestion fails to
inhibit hypothalamic neuronal activity in patients with type 2 diabetes," Diabetes,
vol. 56, pp. 2547{50, Oct 2007.
[49] Y. Liu, J. H. Gao, H. L. Liu, and P. T. Fox, \The temporal response of the brain
after eating revealed by functional mri," Nature, vol. 405, pp. 1058{62, Jun 2000.
[50] G. Musen, D. C. Simonson, N. R. Bolo, A. Driscoll, K. Weinger, A. Raji,
J. Th eberge, P. F. Renshaw, and A. M. Jacobson, \Regional brain activation during
hypoglycemia in type 1 diabetes," J Clin Endocrinol Metab, vol. 93, pp. 1450{7,
Apr 2008.
[51] C. J. Holmes, R. Hoge, L. Collins, R. Woods, A. W. Toga, and A. C. Evans, \En-
hancement of mr images using registration for signal averaging," J Comput Assist
Tomogr, vol. 22, no. 2, pp. 324{33, 1998.
[52] J. L. Lancaster, L. H. Rainey, J. L. Summerlin, C. S. Freitas, P. T. Fox, A. C.
Evans, A. W. Toga, and J. C. Mazziotta, \Automated labeling of the human brain:
a preliminary report on the development and evaluation of a forward-transform
method," Hum Brain Mapp, vol. 5, no. 4, pp. 238{42, 1997.
[53] J. L. Lancaster, M. G. Woldor, L. M. Parsons, M. Liotti, C. S. Freitas, L. Rainey,
P. V. Kochunov, D. Nickerson, S. A. Mikiten, and P. T. Fox, \Automated talairach
atlas labels for functional brain mapping," Hum Brain Mapp, vol. 10, pp. 120{31,
Jul 2000.
[54] S. Tsao, T. C. Adam, M. I. Goran, and M. Singh, \Application of fmri to obesity
research: dierences in reward pathway activation measured with fmri bold during
visual presentation of high and low caloric foods," in Proceedings of SPIE Medical
Imaging, San Diego, CA, 2012.
[55] A. Z. Snyder and M. E. Raichle, \A brief history of the resting state: The washington
university perspective," Neuroimage, Jan 2012.
[56] S. Vidarsdottir, P. Vlug, F. Roelfsema, M. Fr olich, and H. Pijl, \Orally disintegrat-
ing and oral standard olanzapine tablets similarly elevate the homeostasis model
189
assessment of insulin resistance index and plasma triglyceride levels in 12 healthy
men: a randomized crossover study," J Clin Psychiatry, vol. 71, pp. 1205{11, Sep
2010.
[57] D. S. Tuch, T. G. Reese, M. R. Wiegell, N. Makris, J. W. Belliveau, and V. J.
Wedeen, \High angular resolution diusion imaging reveals intravoxel white matter
ber heterogeneity," Magnetic Resonance in Medicine, vol. 48, no. 4, pp. 577{582,
2002.
[58] M. Singh and C.-W. Wong, \Independent component analysis-based multiber
streamline tractography of the human brain," Magn Reson Med, vol. 64, pp. 1676{
84, Dec 2010.
[59] M. Singh and C.-W. Wong, \Recovery of multiple bers per voxel by ica in dti
tractography," Conf Proc IEEE Eng Med Biol Soc, vol. 1, pp. 735{8, 2006.
[60] V. Wedeen, R. Wang, J. Schmahmann, T. Benner, W. Tseng, G. Dai, D. Pandya,
P. Hagmann, H. D'Arceuil, and A. De Crespigny, \Diusion spectrum magnetic
resonance imaging (dsi) tractography of crossing bers," Neuroimage, vol. 41, no. 4,
pp. 1267{1277, 2008.
[61] V. J. Wedeen, P. Hagmann, W.-Y. I. Tseng, T. G. Reese, and R. M. Weissko,
\Mapping complex tissue architecture with diusion spectrum magnetic resonance
imaging," Magnetic Resonance in Medicine, vol. 54, no. 6, pp. 1377{1386, 2005.
[62] J.-D. Tournier, F. Calamante, D. G. Gadian, and A. Connelly, \Direct estimation
of the ber orientation density function from diusion-weighted mri data using
spherical deconvolution," Neuroimage, vol. 23, no. 3, pp. 1176{1185, 2004.
[63] R. W. Cox et al., \Afni: software for analysis and visualization of functional mag-
netic resonance neuroimages," Computers and Biomedical research, vol. 29, no. 3,
pp. 162{173, 1996.
[64] R. Wang, T. Benner, A. Sorensen, and V. Wedeen, \Diusion toolkit: a software
package for diusion imaging data processing and tractography," in Proc Intl Soc
Mag Reson Med, vol. 15, p. 3720, 2007.
[65] A. Hyv arinen and E. Oja, \A fast xed-point algorithm for independent component
analysis," Neural computation, vol. 9, no. 7, pp. 1483{1492, 1997.
[66] P. Maillard, E. Fletcher, D. Harvey, O. Carmichael, B. Reed, D. Mungas, and
C. DeCarli, \White matter hyperintensity penumbra," Stroke, vol. 42, pp. 1917{22,
Jul 2011.
[67] L. C. Silbert, H. H. Dodge, L. G. Perkins, L. Sherbakov, D. Lahna, D. Erten-Lyons,
R. Woltjer, L. Shinto, and J. A. Kaye, \Trajectory of white matter hyperintensity
190
burden preceding mild cognitive impairment," Neurology, vol. 79, pp. 741{7, Aug
2012.
[68] J. P. Seab, W. J. Jagust, S. T. S. Wong, M. S. Roos, B. R. Reed, and T. F. Budinger,
\Quantitative nmr measurements of hippocampal atrophy in alzheimer's disease,"
Magn. Reson. Med., vol. 8, no. 200-208, 1988.
[69] P. M. Thompson, C. Schwartz, R. T. Lin, A. A. Khan, and A. W. Toga, \Three-
dimensional statistical analysis of sulcal variability in the human brain," J. Neurosci,
vol. 16, pp. 4261{4274, 1996.
[70] J. Ashburner and K. J. Friston, \Nonlinear spatial normalization using basis func-
tions," Human Brain Mapping, vol. 7, no. 4, pp. 254{266, 1999.
[71] K. Juottonen, M. P. Laakso, K. Partanen, and H. Soininen, \Comparative mr anal-
ysis of the entorhinal cortex and hippocampus in diagnosing alzheimer disease,"
AJNR Am J Neuroradiol, vol. 20, no. 139-144, 1999.
[72] I. D. Dinov, M. S. Mega, P. M. Thompson, L. Lee, R. P. Woods, C. J. Holmes, D. W.
Sumners, and A. W. Toga, \Analyzing functional brain images in a probabilistic
atlas: a validation of subvolume thresholding," J. Comput. Assist. Tomogr., vol. 24,
pp. 128{138, 2000.
[73] B. Fischl, D. H. Salat, and E. Busa, \Whole brain segmentation: automated labeling
of neuroanatomical structures in the human brain," Neuron, vol. 33, no. 341-355,
2002.
[74] S. M. Smith, \Fast robust automated brain extraction," Hum. Brain Mapp, vol. 17,
no. 143-155, 2002.
[75] B. Fischl, A. van der Kouwe, C. Destrieux, E. Halgren, F. Segonne, and et al,
\Automatically parcellating the human cerebral cortex," Cereb Cortex, vol. 14,
no. 11-22, 2004.
[76] Y. Kabir, M. Dojat, B. Scherrer, F. Forbes, and C. Garbay, \Multimodal mri
segmentation of ischemic stroke lesions," Conf. Proc. IEEE Eng Med. Biol. Soc,
pp. 1595{1598, 2007.
[77] P. Nachev, E. Coulthard, H. R. Jager, C. Kennard, and M. Husain, \Enantiomorphic
normalization of focally lesioned brains," Neuroimage, vol. 39, pp. 1215{1226, 2008.
[78] A. Klein, J. Andersson, B. Ardekani, J. Ashburner, B. Avants, M. Chiang, G. Chris-
tensen, D. Collins, J. Gee, P. Hellier, J. Song, M. Jenkinson, C. Lepage, D. Rueck-
ert, P. Thompson, T. Vercauteren, R. Woods, J. Mann, and R. Parsey, \Evaluation
of 14 nonlinear deformation algorithms applied to human brain mri registration,"
Neuroimage, vol. 46, no. 3, pp. 786{802, 2009.
191
[79] A. Solodkin, U. Hasson, R. Siugzdaite, M. Schiel, E. Chen, K. Rolf, and S. Small,
\Virtual brain transplantation (vbt): A method for accurate image registration and
parcellation in large cortical stroke," Arch Ital Biol, vol. 148, no. 3, pp. 219{41,
2010.
[80] A. Albanese, C. Hall, and R. Stanhope, \The use of a computerized method of bone
age assessment in clinical practice," Horm Res, pp. 2{7, 1995.
[81] W. Grelich and S. Pyle, Radiographic Atlas of Skeletal Development of Hand Wrist,
Stanford, CA: Stanford University Press. 1959.
[82] A. Gertych, A. Zhang, J. Sayre, S. Pospiech-Kurkowska, and H. K. Huang, \Bone
age assessment of children using a digital hand atlas," to appear in J Comp Med
Imag Graphic, vol. 31, pp. 4{5, May 2007.
[83] H. K. Huang, A. Zhang, B. Liu, Z. Zhou, J. Documet, N. King, and L. W. C. Chan,
\Data grid for large-scale medical image archive and analysis," in Proceedings of the
13th ACM International Conference on Multimedia, pp. 1005{1013, 2005.
[84] E. Pietka, A. Gertych, and K. Witko, \Informatics infrastructure of cad system,"
Comput Med Imaging Graph, vol. 29, pp. 157{69, 2005.
[85] F. Cao, H. K. Huang, E. Pietka, and V. Gilsanz, \Digital hand atlas and web-based
bone age assessment: system design and implementation," Comput Med Imaging
Graph, vol. 24, pp. 297{307, 2000.
[86] E. Pietka, A. Gertych, S. P. E. Kurkowska, F. Cao, H. K. Huang, and V. Gilzanz,
\Computerassisted bone age assessment: graphical user interface for image process-
ing and comparison," J Digit Imaging, vol. 17, pp. 175{88, 2004.
[87] E. Pietka, A. Gertych, S. Pospiech, F. Cao, H. K. Huang, and V. Gilsanz,
\Computer-assisted bone age assessment: image preprocessing and epiphy-
seal/metaphyseal roi extraction," IEEE Trans Med Imaging, vol. 20, pp. 715{29,
2001.
[88] E. Pietka, S. Pospiech-Kurkowska, A. Gertych, and F. Cao, \Integration of com-
puter assisted bone age assessment with clinical pacs," Comput Med Imaging Graph,
vol. 27, pp. 217{28, 2003.
[89] A. G. Zhang and B. J. Liu, \Automatic bone age assessment for young children
from newborn to 7-year- old using carpal bones," J Comp Med Imag and Graphics,
vol. 31, pp. 4{5, May 2007.
[90] P. DuBois, MySQL: the denitive guide to using, programming, and administering
MySQL 4. Pearson Education, 3rd ed., 2003.
192
[91] H. E. Williams, David Lane Web Database Applications with PHP MySQL.
O'Reilly, 2nd ed., 2002.
[92] E. Pietka, A. Gertych, S. Pospiech, F. Cao, H. K. Huang, and V. Gilsanz,
\Computer-assisted bone age assessment: image preprocessing and epiphy-
seal/metaphyseal roi extraction," IEEE Trans. Med. Imaging., vol. 20, pp. 715{729,
Aug 2001.
[93] W. Grelich and S. Pyle, Radiographic Atlas of Skeletal Development of Hand Wrist.
Stanford, CA: Stanford University Press, 1959.
[94] J. M. Tanner and R. H. Whitehouse, Assessment of Skeletal Maturity and Prediction
of Adult Height (TW2 Method). London: Academic Press, 1975.
[95] A. Zhang, A. Gertych, B. J. Liu, H. K. Huang, and S. Kurkowska-Pospiech, \Carpal
bone segmentation and features analysis in bone age assessment of children," in
Proceedings of RSNA Conference, (pp.688Chicago, pp.688), Chicago, 2005.
[96] A. Zhang, A. Gertych, B. J. Liu, and H. K. Huang, \Data mining for 'average'
image based on phalangeal and carpal bone features in a large-scale digital hand
atlas," in Proceedings of RSNA Conference, 2006.
[97] A. Zhang, A. Gertych, and B. J. Liu, Automatic Bone Age Assessment for Young
Children from Newborn to 7-Year- Old Using Carpal Bones. JCMIG, 2007.
[98] A. Zhang, S. Tsao, J. Sayre, A. Gertych, B. J. Liu, and H. K. Huang, \Is greulich
and pyle atlas still a good reference for bone age assessment?," in Proceedings of
SPIE Medical Imaging, February 2007.
[99] A. Gertych, E. Pietka, and B. J. Liu, \Segmentation of regions of interest and post-
segmentation edge location improvement in computer-aided bone age assessment,"
Pattern Anal Applic, vol. 10, pp. 115{123, April 2007.
[100] J. Zhou, J. Liu, V. A. Narayan, and J. Ye, \Modeling Disease Progression via
Multi-task Learning," NeuroImage, 2013.
[101] Y. Wang, L. Yuan, J. Shi, A. Greve, J. Ye, A. W. Toga, A. L. Reiss, and P. M.
Thompson, \Applying tensor-based morphometry to parametric surfaces can im-
prove mri-based disease diagnosis," Neuroimage, vol. 74, pp. 209{30, Jul 2013.
[102] R. J. Caselli, D. E. C. Locke, A. C. Dueck, D. S. Knopman, B. K. Woodru,
C. Homan-Snyder, R. Rademakers, A. S. Fleisher, and E. M. Reiman, \The neu-
ropsychology of normal aging and preclinical alzheimer's disease," Alzheimers De-
ment, Mar 2013.
193
[103] L. S. Elias-Sonnenschein, S. Helisalmi, T. Natunen, A. Hall, T. Paajanen, S.-K.
Herukka, M. Laitinen, A. M. Remes, A. M. Koivisto, K. M. Mattila, T. Lehtim aki,
F. R. J. Verhey, P. J. Visser, H. Soininen, and M. Hiltunen, \Genetic loci associated
with alzheimer's disease and cerebrospinal
uid biomarkers in a nnish case-control
cohort," PLoS One, vol. 8, no. 4, p. e59676, 2013.
[104] S. J. Teipel, M. Grothe, S. Lista, N. Toschi, F. G. Garaci, and H. Hampel, \Rele-
vance of magnetic resonance imaging for early detection and diagnosis of alzheimer
disease," Med Clin North Am, vol. 97, pp. 399{424, May 2013.
[105] G. A. Becker, M. Ichise, H. Barthel, J. Luthardt, M. Patt, A. Seese, M. Schultze-
Mosgau, B. Rohde, H.-J. Gertz, C. Reininger, and O. Sabri, \Pet quantication of
18f-
orbetaben binding to -amyloid deposits in human brains," J Nucl Med, Mar
2013.
[106] K. Blennow and H. Zetterberg, \The application of cerebrospinal
uid biomarkers
in early diagnosis of alzheimer disease," Med Clin North Am, vol. 97, pp. 369{76,
May 2013.
[107] I. Hajjar, L. Brown, W. J. Mack, and H. Chui, \Alzheimer pathology and angioten-
sion receptor blockers," JAMA Neurol, vol. 70, p. 414, Mar 2013.
[108] S. G. Mueller, M. W. Weiner, L. J. Thal, R. C. Petersen, C. R. Jack, W. Jagust,
J. Q. Trojanowski, A. W. Toga, and L. Beckett, \Ways toward an early diagno-
sis in alzheimer's disease: the alzheimer's disease neuroimaging initiative (adni),"
Alzheimers Dement, vol. 1, pp. 55{66, Jul 2005.
[109] B. S. Bagepally, H. N. Halahalli, J. P. John, L. Kota, M. Purushottam, O. Mukher-
jee, P. T. Sivakumar, S. Bharath, S. Jain, and M. Varghese, \Apolipoprotein e4
and brain white matter integrity in alzheimer's disease: tract-based spatial statis-
tics study under 3-tesla mri," Neurodegener Dis, vol. 10, no. 1-4, pp. 145{8, 2012.
[110] M. Schwartz and K. Niswender, \Adiposity signaling and biological defense against
weight gain: absence of protection or central hormone resistance?," J Clin En-
docrinol Metab, vol. 89, pp. 5889{97, 2004.
[111] R. Kuczmarski, C. Ogden, L. Grummer-Strawn, K. Flegal, S. Guo, R. Wei, Z. Mei,
L. Curtin, A. Roche, and C. Johnson, \Cdc growth charts: United states," Adv
Data, pp. 1{27, 2000.
[112] S. Kullmann, M. Heni, R. Veit, C. Ketterer, F. Schick, H. Haring, A. Fritsche, and
H. Preissl, \The obese brain: association of body mass index and insulin sensitiv-
ity with resting state network functional connectivity," Hum Brain Mapp, vol. 33,
pp. 1052{61, 2012.
194
[113] K. Anthony, L. Reed, J. Dunn, E. Bingham, D. Hopkins, P. Marsden, and S. Amiel,
\Attenuation of insulin-evoked responses in brain networks controlling appetite and
reward in insulin resistance: the cerebral basis for impaired control of food intake
in metabolic syndrome?," Diabetes, vol. 55, pp. 2986{92, 2006.
[114] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Del-
croix, B. Mazoyer, and M. Joliot, \Automated anatomical labeling of activations
in spm using a macroscopic anatomical parcellation of the mni mri single- subject
brain," Neuroimage, vol. 15, pp. 273{89, 2002.
[115] J. Bruce, L. Hancock, A. Bruce, R. Lepping, L. Martin, J. Lundgren, S. Malley,
L. Holsen, and C. Savage, Changes in brain activation to food pictures after ad-
justable gastric banding. Surg Obes Relat Dis, 2011.
[116] K. Page, D. Seo, R. Belfort-DeAguiar, C. Lacadie, J. Dzuira, S. Naik, S. Amarnath,
R. Constable, R. Sherwin, and R. Sinha, \Circulating glucose levels modulate neural
control of desire for high-calorie foods in humans," J Clin Invest, vol. 121, pp. 4161{
9, 2011.
[117] T. Adam, C. Toledo-Corral, C. Lane, M. Weigensberg, and D. Spruijt-Metz, \Davies
jn a nd goran mi: Insulin sensitivity as an independent predictor of fat mass gain
in hispanic adolescents," Diabetes Care, vol. 32, pp. 2114{5, 2009.
[118] Y. Rothemund, C. Preuschhof, G. Bohner, H. Bauknecht, R. Klingebiel, H. Flor,
and B. Klapp, \Dierential activation of the dorsal striatum by high-calorie visual
food stimuli in obese individuals," Neuroimage, vol. 37, pp. 410{21, 2007.
[119] D. Figlewicz, S. Evans, J. Murphy, M. Hoen, and D. Baskin, \Expression of receptors
for insulin and leptin in the ventral tegmental area/substantia nigra (vta/sn) of the
rat," Brain Res, vol. 964, pp. 107{15, 2003.
[120] J. Tanner, \Growth and maturation during adolescence," Nutr Rev, vol. 39, pp. 43{
55, 1981.
[121] N. Speed, C. Saunders, A. Davis, W. Owens, H. Matthies, S. Saadat, J. Kennedy,
R. Vaughan, R. Neve, C. Lindsley, S. Russo, L. Daws, K. Niswender, and A. Galli,
\Impaired striatal akt signaling disrupts dopamine homeostasis and increases feed-
ing," PLoS One, 2011.
[122] C. Davis, S. Strachan, and M. Berkson, \Sensitivity to reward: implications for
overeating and overweight," Appetite, vol. 42, pp. 131{8, 2004.
[123] S. Wallner-Liebmann, K. Koschutnig, G. Reishofer, E. Sorantin, B. Blaschitz,
R. Kruschitz, H. Unterrainer, R. Gasser, F. Freytag, C. Bauer-Denk, A. Schienle,
A. Schafer, and H. Mangge, \Insulin and hippocampus activation in response to im-
ages of high-calorie food in normal weight and obese adolescents," Obesity (Silver
Spring), vol. 18, pp. 1552{7, 2010.
195
[124] M. Goran, R. Bergman, M. Cruz, and R. Watanabe, \Insulin resistance and associ-
ated compensatory responses in african-american and hispanic children," Diabetes
Care, vol. 25, pp. 2184{90, 2002.
[125] J. Lee, \Insulin resistance in children and adolescents," Rev Endocr Metab Disord,
vol. 7, pp. 141{7, 2006.
[126] D. Figlewicz, J. Bennett, A. Naleid, C. Davis, and J. Grimm, \Intraventricular in-
sulin and leptin decrease sucrose self-administration in rats," Physiol Behav, vol. 89,
pp. 611{6, 2006.
[127] M. Schwartz, A. Sipols, S. Kahn, D. Lattemann, G. Taborsky, R. Bergman,
S. Woods, and D. Porte, \Jr," : Kinetics and specicity of insulin uptake from
plasma into cerebrospinal
uid, pp. 378{83, 1990.
[128] L. Tamm, V. Menon, and A. Reiss, \Maturation of brain function associated with
response inhibition," J Am Acad Child Adolesc Psychiatry, vol. 41, pp. 1231{8,
2002.
[129] A. Jastrebo, R. Sinha, C. Lacadie, D. Small, R. Sherwin, and M. Potenza, \Neural
correlates of stress- and food cue-induced food craving in obesity: Association with
insulin levels," Diabetes Care, vol. 36, pp. 394{402, 2013.
[130] M. Cruz, G. Shaibi, M. Weigensberg, D. Spruijt-Metz, G. Ball, and M. Goran, \Pe-
diatric obesity and insulin resistance: chronic disease risk and implications for treat-
ment and prevention beyond body weight modication," Annu Rev Nutr, vol. 25,
pp. 435{68, 2005.
[131] A. DelParigi, K. Chen, A. Salbe, J. Hill, R. Wing, E. Reiman, and P. Tataranni,
\Persistence of abnormal neural responses to a meal in postobese individuals," Int
J Obes Relat Metab Disord, vol. 28, no. 370-7, pp. 440{8, 2007.
[132] M. Goran, C. Lane, C. Toledo-Corral, and M. Weigensberg, Persistence of Pre-
Diabetes in Overweight and Obese Hispanic Children: Association With Progressive
Insulin Resistance, Poor Beta-cell Function and Increasing Visceral Fat. Diabetes,
2008.
[133] R. Bergman, L. Phillips, and C. Cobelli, \Physiologic evaluation of factors con-
trolling glucose tolerance in man: measurement of insulin sensitivity and beta-cell
glucose sensitivity from the response to intravenous glucose," J Clin Invest, vol. 68,
pp. 1456{67, 1981.
[134] A. Byars, S. Holland, R. Strawsburg, W. Bommer, R. Dunn, V. Schmithorst, and
E. Plante, \Practical aspects of conducting large-scale functional magnetic reso-
nance imaging studies in children," J Child Neurol, vol. 17, pp. 885{90, 2002.
196
[135] K. Flegal, C. Ogden, and M. Carroll, \Prevalence and trends in overweight in
mexican-american adults and children," Nutr Rev, pp. 144{8, 2004.
[136] E. Stice, S. Spoor, C. Bohon, and D. Small, \Relation between obesity and blunted
striatal response to food is moderated by taqia a1 allele," Science, vol. 322, pp. 449{
52, 2008.
[137] A. Must and R. Strauss, \Risks and consequences of childhood and adolescent
obesity," Int J Obes Relat Metab Disord, pp. 2{11, 1999.
[138] A. Konner, S. Hess, S. Tovar, A. Mesaros, C. Sanchez-Lasheras, N. Evers, L. Ver-
hagen, H. Bronneke, A. Kleinridders, B. Hampel, P. Kloppenburg, and J. Bruning,
\Role for insulin signaling in catecholaminergic neurons in control of energy home-
ostasis," Cell Metab, vol. 13, pp. 720{8, 2011.
[139] D. H. Brainard, \The psychophysics toolbox," Spat Vis, vol. 10, no. 4, pp. 433{6,
1997.
[140] E. Gomez-Perez, F. Ostrosky-Solis, and O. Prospero-Garcia, \The development of
attention, memory and the inhibitory processes: the chronological relation with the
maturation of brain structure and functioning]," Rev Neurol, vol. 37, pp. 561{7,
2003.
[141] P. Lang, \The emotion probe. studies of motivation and attention," Am Psychol,
vol. 50, pp. 372{85, 1995.
[142] G. Wang, N. Volkow, and J. Fowler, \The role of dopamine in motivation for food
in humans: implications for obesity," Expert Opin Ther Targets, vol. 6, pp. 601{9,
2002.
[143] W. Banks, \The source of cerebral insulin," Eur J Pharmacol, vol. 490, pp. 5{12,
2004.
[144] E. Boyko, W. Fujimoto, D. Leonetti, and L. Newell-Morris, \Visceral adiposity and
risk of type 2 diabetes: a prospective study among japanese americans," Diabetes
Care, vol. 23, pp. 465{71, 2000.
197

Abstract (if available)

Abstract The motivation for this work is a vision of widespread adoption of a priori quantitative epidemiological information for clinical decision-making, and can be seen as a quantitative large-scale extension of evidence-based medicine (EBM) [2]. Medical images can be seen as a spatially encoded map of physiological measurements that can be used to predict prognosis and to drive treatment plans. This paradigm can be very powerful and is driven by the recent big data revolution in computer science as well as the increasing availability of medical imaging modalities due to decreases in manufacturing costs. In order to achieve this overarching goal, three practical requirements must be reached and correspond to the parts of this thesis: Part A: Developing IT infrastructure and technology that enables the dataset to be properly collected and organized for analysis. Part B & C: Generation of functional (Part B) and structural (Part C) medical imaging contrast that are optimized for analysis. Part D: Pattern recognition techniques (including both image processing and machine learning techniques) to mine information from the large imaging datasets generated. As part of the thesis, I discuss my contribution to IT infrastructure (Part A) by developing a Short Message Service (SMS)-based system to control the clinically used Picture Archival and Communication System (PACS) (Ch.2) as well as an imaging study tool that categorizes patient imaging data for use in retrospective studies(Ch.3). I then go on to detail my work with functional neuroimaging of obesity using functional magnetic resonance imaging (fMRI)(Ch.4) and (Ch.5). Chapters 6-9 details my efforts at studying abnormal aging versus normal aging using diffusion MRI as well as applications of diffusion MRI to surgical planning. Chapters 10 discusses my work integrating diffusion MR with FLAIR MRI to investigate the properties of white matter lesions and how it can be used in the clinical setting. Chapter 11 then moves on to talk about my work modifying standard brain parcellation techniques to allow them to work with aged brains with large infarcts. Chapters 6-11 altogether represent my efforts in structural neuroimaging using MRI (Part C). The thesis then closes with capstone work in development staging using hand x-rays using fuzzy logic (Ch. 12 & 13). To close the work with Alzheimer's Disease (AD) and aging, we used machine learning techniques to predict disease progression based on a baseline MRI scan as well as higher order analysis of our diffusion MRI dataset by integrating MRI information with other clinical information such as neuropsychological tests, cardiovascular status. This is all in an effort to computationally explore the relationship between MRI measurements and clinical presentation of disease as measured by neuropsychological scores. Similarly with the Obesity work, we related fMRI activation differences between high and low calorie foods with non-imaging information such as insulin resistance (Ch. 16).

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Unveiling the white matter microstructure in 22q11.2 deletion syndrome with diffusion magnetic resonance imaging

PDF

Diffusion MRI of the human brain: signal modeling and quantitative analysis

PDF

Reproducibility and management of big data in brain MRI studies

PDF

Functional models of fMRI BOLD signal in the visual cortex

PDF

Correction, coregistration and connectivity analysis of multi-contrast brain MRI

PDF

Shape, pose, and connectivity in subcortical networks across the human lifespan

PDF

Parametric and non‐parametric modeling of autonomous physiologic systems: applications and multi‐scale modeling of sepsis

PDF

Neuroimaging in complex polygenic disorders

PDF

Neuroimaging markers of risk & resilience to brain aging and dementia

PDF

Age differences in diffusivity in the locus coeruleus and its ascending noradrenergic tract

PDF

Characterization of lenticulostriate arteries using high-resolution black blood MRI as an early imaging biomarker for vascular cognitive impairment and dementia

PDF

3D vessel mapping techniques for retina and brain as an early imaging biomarker for small vessel diseases

PDF

Transparent and lightweight medical image analysis techniques: algorithms and applications

PDF

Toward understanding speech planning by observing its execution—representations, modeling and analysis

PDF

Simulating electrical stimulation and recording in a multi-scale model of the hippocampus

PDF

Functional connectivity analysis and network identification in the human brain

PDF

Quantitative MRI for oxygenation imaging in cerebrovascular diseases

PDF

Hemisphere-specific deficits in the control of bimanual movements after stroke

Asset Metadata

Creator Tsao, Sinchai (author)

Core Title Pattern detection in medical imaging: pathology specific imaging contrast, features, and statistical models

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 10/10/2013

Defense Date 09/11/2013

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Alzheimer's disease,bone age,cFSGL,computed aided diagnosis,convex fused group lasso,diffusion tensor imaging,DTI,fMRI,group lasso,ICA,independent component analysis,machine learning,medical imaging,MRI,neuroimaging,OAI-PMH Harvest,obesity,PACS,Ris,X-ray

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Grzywacz, Norberto M. (committee chair), Lepore, Natasha (committee chair), Gertych, Arkadiusz (committee member), Marmarelis, Vasilis Z. (committee member), Wolf, Walter (committee member)

Creator Email mail@sinchaitsao.com,stsao@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-337019

Unique identifier UC11295256

Identifier etd-TsaoSincha-2091.pdf (filename),usctheses-c3-337019 (legacy record id)

Legacy Identifier etd-TsaoSincha-2091.pdf

Dmrecord 337019

Document Type Dissertation

Format application/pdf (imt)

Rights Tsao, Sinchai

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA