University of Southern California Dissertations and Theses

The neuroimaging of cancer and chemotherapy

(USC Thesis Other)

The neuroimaging of cancer and chemotherapy

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content 1

The Neuroimaging of Cancer and Chemotherapy

Mark S. Shiroishi, M.D.

Conferring Major/Program: Clinical, Biomedical and Translational Investigation

Degree being conferred: Master of Science

USC Graduate School
University of Southern California

Degree conferral date: May 11, 2018

Dedication Page

To my family.

Acknowledgements

I owe deep gratitude to my mentors and thesis committee members Drs. Neda Jahanshad, Paul Thompson and
Meredith Braskie. They have shown tremendous generosity and I have learned so much from them. Similarly, I
am very thankful for the faculty and staff of the SC CTSI – Drs. Jon Samet, Cecilia Patino-Sutton and Melissa
Wilson as well as Jeanne Dzekov and Karen Kim. Drs. Jorge Nieva, Margy Gatz, Greg Ver Steeg, Andy Saykin
and Helena Chui, have also been very giving of their time and guidance.

I would also like to thank the other members of the USC Imaging Genetics Center and Mark and Mary Stevens
Neuroimaging and Informatics Institute including Dr. Arthur Toga, Dr. Vikash Gupta, Joshua Faskowitz, Alyssa
Zhu, Dr. Lauren Salminen, Brandy Riedel, Faisal Rashid, Aggie McMahon and Dan Moyer; my research fellows
Bavrina Bigjahan, Wesley Surento, Dr. Amir Emami, Dr. Francesco D’Amore; the USC Department of Radiology
including Rosy Diaz and my colleagues in the Division of Neuroradiology as well as Drs. Ed Grant, Meng Law
and Bhushan Desai; USC MRI technicians; and Drs. Berislav Zlokovic and Russ Jacobs.

Table of Contents

SECTION Page

PREFACE 1

THE ENIGMA CANCER AND CHEMOTHERAPY WORKING GROUP 2

TECHNICAL DEVELOPMENT 22

CAREER DEVELOPMENT 37
REFERENCES 43

1

PREFACE
This thesis is submitted to the faculty of the University of Southern California in partial fulfillment of the
requirements for the Master of Science in Clinical, Biomedical and Translational Investigations degree.
Members of my thesis committee are: Paul M. Thompson, Ph.D. (Chair), Neda Jahanshad, Ph.D. and Meredith
Braskie, Ph.D.
The research and career development activities in this thesis is a comprehensive overview of work conducted as
part of my NIH KL2 Mentored Research Career Development Award (NIH/NCRR/NCATS grant
KL2TR000131), NIH Loan Repayment Award (NIH 1 L30 CA209248-01, Wright Foundation Pilot Award and
grant #IRG-16-181-57 from the American Cancer Society.

2

THE ENIGMA CANCER AND CHEMOTHERAPY WORKING GROUP
By 2026 there will be 20.3 million US cancer survivors with 14 million surviving at least 5 years (Bluethmann,
Mariotto, & Rowland, 2016; https://www.cancer.gov/about-cancer/treatment/research/understanding-
chemobrain). Therefore, it is critical to better study cancer survivors’ quality of life. Cognitive dysfunction from
non-central nervous system (CNS) cancer and its treatment is called cancer-related cognitive impairment
(CRCI). It is very common, yet its underlying neural substrates are unclear (Low, Kalinski, & Bovbjerg, 2015;
McDonald & Saykin, 2013; Patel et al., 2015; Pomykala et al., 2013; Simo, Rifa-Ros, Rodriguez-Fornells, &
Bruna, 2013; Wefel, Kesler, Noll, & Schagen, 2015). Neuroimaging is an ideal method to study CRCI. However,
performing large neuroimaging studies is difficult for any single site to do on its own, and consequently
neuroscience studies (Button et al., 2013), including neuroimaging are often underpowered. We aim to conduct
the largest brain imaging studies to date of the structural anomalies that may underlie CRCI. Our findings
may ultimately improve disease monitoring, therapy selection, preventative strategies and risk assessment
analysis.
My Primary Mentor Dr. Paul Thompson leads the world’s largest neuroimaging effort, the Enhancing Neuro
Imaging Genetics through Meta-Analysis (ENIGMA) Consortium (Guadalupe et al., 2016; Hibar, Adams, et al.,
2017; Jahanshad et al., 2013; Kochunov et al., 2015; Schmaal, Hibar, et al., 2016; Schmaal et al., 2015; Thompson
et al., 2015; Thompson et al., 2014; van Erp et al., 2015). As recently profiled in Science (Guglielmi, 2018),
ENIGMA crowdsources data worldwide, employs standardized image analysis and meta-/mega-analyzes these
results. This type of collaborative methodology can overcome the problem of low statistical power that is very
common in neuroimaging studies. I am therefore in a unique position to accomplish my primary aim by forming
the ENIGMA Cancer and Chemotherapy Working Group (CCWG). This will result in enough power to
detect the subtle brain effects of CRCI and clear up inconsistencies that no single research site could do on its
own.
CRCI after chemotherapy, known as “chemobrain,” has attack rates of up to 78% and an effective treatment
remains elusive (Treanor et al., 2016; Wefel & Schagen, 2012). The CRCI neuroimaging literature is still in its
3

infancy and most studies have focused on chemobrain in breast cancer survivors. These are mainly small, single-
institution studies using morphometric MRI that have found decreased gray matter (GM) and white matter
(WM) after chemotherapy, generally in the frontotemporal regions (de Ruiter et al., 2012; Inagaki et al., 2007;
Koppelmans et al., 2012; McDonald, Conroy, Ahles, West, & Saykin, 2010); some changes have been associated
with cognitive deficits.(Inagaki et al., 2007). However, results have been variable and contradictory (Simo et
al., 2013; Yoshikawa et al., 2005). Limited evidence also suggests that brain abnormalities and cognitive
dysfunction may be due to cancer even without chemotherapy (Scherling et al., 2012; Wefel, Lenzi, Theriault,
Davis, & Meyers, 2004). Our own pilot data (Shiroishi, Gupta, Bigjahan, et al., 2017) has found decreased
cortical surface area or thickness in those with of a variety of different cancers with and without
chemotherapy as well as in lung cancer and prostate cancer survivors. A few structural diffusion tensor
imaging (DTI) MRI studies focusing on WM integrity have also been performed in CRCI, again mainly in small
cohorts of breast cancer survivors. They have generally found decreased WM integrity in the frontal, temporal
and parietal regions of the brain after chemotherapy, though again results were variable (Abraham et al., 2008;
de Ruiter et al., 2012; Deprez et al., 2011; Deprez, Billiet, Sunaert, & Leemans, 2013).
However, comparison of these morphometric and DTI MRI studies is difficult given the different means of
image analysis. Our proposal can address a critical gap in the CRCI literature by robustly characterizing the
brain structural changes from morphometric and DTI MRI across different cancer types with and without
chemotherapy/other therapies using standardized image analysis. Furthermore, our pilot data in prostate cancer
survivors indicates that evolving methods like machine learning-derived brain predicted age (BPA) can
potentially generate novel imaging biomarkers of CRCI not yet available with traditional analysis. To address
these needs, we propose the following specific aims to be tested within the ENIGMA CCWG (currently 11 sites
worldwide) using very large sample sizes, standardized image processing and mega-analysis:
Specific Aim 1. Compare brain morphometric MRI measures in: (1) All cancers (± chemotherapy) vs control,
(2) Chemotherapy vs control, (3) No chemotherapy vs control, (4) Chemotherapy vs no chemotherapy, and (5)
Most common individual types of non-CNS cancer without chemotherapy vs controls. Subgroup analysis will
4

also compare difference classes of chemotherapy as well as non-chemotherapy treatments. Hypothesis: There
will be significant decreases in the amount of GM and WM of the frontotemporal regions of adult non-CNS cancer
survivors treated with and without chemotherapy in all comparison groups.
Specific Aim 2. Compare white matter integrity with DTI MRI in the same comparison groups as for Aim 1.
Hypothesis: There will be decreased regional brain WM integrity of adult non-CNS cancer survivors treated with
and without chemotherapy in all group comparisons.
Specific Aim 3. This exploratory aim will utilize BPA from machine learning in order to find and validate the
robustness of empirical sub-phenotypes in the above comparisons for Aim 1 using morphometric data.
Hypothesis: Machine learning methods will uncover previously unknown phenotypes, like accelerated BPA,
that may underlie CRCI.
The proposed work will accomplish 3 objectives. First, our ENIGMA study will provide enough power for the
most robust assessment of structural brain anomalies in a variety of cancers and therapies that may be associated
with CRCI. Second, it will provide unparalleled training in Big Data neuroimaging methods, machine learning,
consortium management and cognitive assessment. Finally, the results from this study will provide preliminary
data for independent funding through mechanisms such as NIH R01 grants.

RESEARCH STRATEGY
A. Significance -
A.1. Cancer-related cognitive impairment (CRCI) is common but its neural substrates remain unclear.
Chemotherapy has resulted in greatly improved outcomes for cancer patients. Unfortunately, cognitive
dysfunction after chemotherapy, known as “chemobrain,” has cumulative incidence as high as 78% (Wefel &
Schagen, 2012) and it can adversely affect quality of life (Wefel et al., 2015). In addition, cancer itself without
chemotherapy, may also result in cognitive dysfunction (Wefel et al., 2004). This cancer consequence, along with
chemobrain, has recently been termed cancer-related cognitive impairment (CRCI) (Wefel et al., 2015).
Changes at the molecular/cellular level can affect brain structure and function. Abnormalities in either can be
5

seen in aging and neurodegenerative diseases that may result in decreased cognitive function (Kesler, 2014). The
underlying pathophysiology of CRCI remains unclear but possible etiologies include decreased white matter
integrity, direct toxic effect of chemotherapeutic agents on neural progenitor cells and postmitotic
oligodendrocytes, abnormal neurogenesis, pro-inflammatory states and altered oxidative balance among other
processes (Wefel et al., 2015). Provocative Question 9 from the National Cancer Institute asks: “What are
the molecular and/or cellular mechanisms that underlie the development of cancer therapy-induced severe
adverse sequelae?”(https://provocativequestions.nci.nih.gov/rfa/mainquestions_listview.html) Though not
directly probing cellular/molecular mechanisms or functional neuroimaging changes, this proposal can provide
better insight into the final neural substrates affected in CRCI, namely altered brain structure. CRCI is one of
the few conditions where potential neurological dysfunction can be anticipated and tracked with neuroimaging
(Kesler et al., 2017). Identification of such changes may ultimately improve disease monitoring, therapy
selection, preventative strategies and risk assessment analysis for cancer survivors suffering from CRCI.
A.2. Neuroimaging studies of CRCI have generally used small sample sizes with, at times, variable and
contradictory results. Neuroimaging could provide more sensitive and reliable biomarkers to predict cognitive
performance in CRCI (Kesler, 2014). The vast majority of CRCI neuroimaging studies have used small sample
sizes as part of single-institution studies of structural, morphometric T1-weighted MRI. These have generally
found significantly decreased amounts of gray (GM) and white matter (WM) of the brains of chemotherapy-
treated cancer survivors compared to controls, generally in the frontotemporal regions (de Ruiter et al., 2012;
Inagaki et al., 2007; Koppelmans et al., 2012; McDonald et al., 2010). Some of these changes have been focal
while others have been diffuse. However, not all studies have found significant differences (Yoshikawa et al.,
2005) and, as McDonald and my external advisor Saykin (McDonald & Saykin, 2013) have suggested, it is
possible that there is publication bias where negative results may be not be reported. This type of investigative
approach will result in a slow, incremental advance in knowledge of CRCI (Wefel, Vardy, Ahles, & Schagen,
2011) and is reflective of the concern that the average statistical power in the neurosciences is very low (Button
et al., 2013). Other well-known brain diseases, such as major depressive disorder (MDD), the most common
6

psychiatric disorder, or schizophrenia, have been studied for decades with underpowered neuroimaging studies
and their neuroanatomical substrate remain unclear (Schmaal, Hibar, et al., 2016; van Erp, Hibar, Rasmussen,
Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle, Hartberg, Gruber, Kraemer, Zilles,
Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders, de Haan, Veltman, Satterthwaite,
Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton, Hass, Calhoun, Bockholt,
Sponheim, Shoemaker, van Haren, Pol, et al., 2016).
A.3. There is a large amount of heterogeneity in neuroimaging study methodology. Apart from
underpowered studies, the large variability with regard to study design, image processing methods and analysis
of CRCI neuroimaging studies can also lead to variable and contradictory findings (McDonald & Saykin, 2013;
Simo et al., 2013).
A.4. Little is known about the neuroimaging changes of CRCI across most cancers (McDonald & Saykin,
2013; Simo et al., 2013; Wefel et al., 2015). The vast majority of the neuroimaging literature of CRCI comes
from studies of breast cancer survivors, with little known regarding other cancer types.
A.5. Few neuroimaging investigations have employed multiparametric MRI studies (de Ruiter et al.,
2012; Ferguson, McDonald, Saykin, & Ahles, 2007; Menning et al., 2015) of CRCI. Multiparametric MRI
has the potential to provide even more insight into the neuroanatomical substrates affected in CRCI because
complementary techniques like morphometric and DTI MRI give a more complete picture of the structural
alterations in CRCI by looking at both the size of brain structures as well as WM integrity.
A.6. Little is known about the use of machine learning techniques and its derived brain-predicted age
(Cole, Poudel, et al., 2017) (BPA) in CRCI. The human brain changes as adults grow older. This brain aging
generally parallels the decrease in cognitive ability and with greater age, there is increased risk of dementia and
neurodegenerative disease (Abbott, 2011). The last few years have seen increased use of machine learning
algorithms towards the analysis of neuroimaging data (Dosenbach et al., 2010; Lemm, Blankertz, Dickhaus, &
Muller, 2011). A relatively recent application of machine learning is the determination of BPA (Cole, Poudel, et
al., 2017). A BPA that is greater than chronological age might reflect deleterious age-related alterations to the
7

brain. Very recent work (Cole, Poudel, et al., 2017) has shown that BPA is an accurate and reliable technique
that can be used in multi-center brain MRI studies. Increased BPA has been demonstrated in studies in adults
with mild cognitive impairment who develop Alzheimer’s disease (Gaser et al., 2013), schizophrenia
(Koutsouleris et al., 2014), Down’s syndrome (Cole, Annus, et al., 2017), diabetes (K. Franke, Gaser, Manor,
& Novak, 2013), epilepsy(Pardoe et al., 2017) and traumatic brain injury (Cole, Leech, Sharp, & Alzheimer's
Disease Neuroimaging, 2015), while increased amounts of education and physical exercise (Steffener et al.,
2016) as well as meditation (Luders, Cherbuin, & Gaser, 2016) appear to have decreased BPA. With regard
to CRCI specifically, little is known regarding the application of machine learning methods. A few recent pilot
studies (Kesler et al., 2017; Kesler et al., 2013) in breast cancer survivors have demonstrated the potential of
other types of machine learning algorithms to learn complex interactions amid numerous predictors. However, to
the best of our knowledge, there has not been a study which has applied machine learning-derived BPA to
CRCI.
B. Innovation - The proposed study will provide new information regarding the neuroanatomical substrates of
CRCI by employing the following innovations through the ENIGMA CCWG:
B.1 We will obtain sample sizes not previously reached for a CRCI neuroimaging study and use standardized
study design and image analysis to produce more robust findings. ENIGMA has used its unique methodology to
publish the largest neuroimaging studies to date of schizophrenia (van Erp, Hibar, Rasmussen, Glahn, Pearlson,
Andreassen, Agartz, Westlye, Haukvik, Dale, Melle, Hartberg, Gruber, Kraemer, Zilles, Donohoe, Kelly,
McDonald, Morris, Cannon, Corvin, Machielsen, Koenders, de Haan, Veltman, Satterthwaite, Wolf, Gur, Gur,
Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton, Hass, Calhoun, Bockholt, Sponheim, Shoemaker,
van Haren, Pol, et al., 2016), MDD (Schmaal, Hibar, et al., 2016; Schmaal, Veltman, et al., 2016), and brain
regional volume (Hibar, Adams, et al., 2017; Hibar et al., 2015). Apart from small sample sizes, the large amount
of heterogeneity of prior neuroimaging study methods can result in variable and non-reproducible results (Button
et al., 2013; McDonald & Saykin, 2013; Simo et al., 2013; Thompson et al., 2015). Our standardized methodology
within the ENIGMA CCWG for the mega-analysis of morphometric and DTI MRI data will be a significant step
8

towards addressing these shortcomings. This mega-analytic approach will ensure the preservation of individual
patient-level data. Most of the ENIGMA disease working groups have focused on neuropsychiatric disorders
and there have been no oncology applications up this point. Our proposal will provide enough power to detect
the subtle effects of cancer and chemotherapy on the brain, clearing up inconsistencies to identify the specific
neural substrates involved in CRCI, namely altered brain structure.
B.2. The ENIGMA CCWG will examine a wide variety of cancers treated with and without chemotherapy.
The vast majority of CRCI neuroimaging studies have focused on breast cancer survivors and little has been
studied regarding other types of non-CNS cancers. Our proposal is the first of its kind to study multiple cancers
at once. We will be studying cancer as a whole as well as the most common types of individual cancers on their
own as well as the effects of different types of therapies (i.e. different types of chemotherapy, hormonal, targeted
therapy, etc.).
B.3. We will perform multiparametric structural MRI studies employing morphometric as well as DTI MRI.
By examining brain structure via both alterations in brain structure size and white matter integrity using
standardized methodology, we will be able to derive different but complementary indications of damage to neural
structures that may underlie CRCI.
B.4. Recent developments in machine learning have introduced new training objectives for neural networks
(Cole, Poudel, et al., 2017) that promote learning of features that are both more interpretable and more
generalizable across domains (sometimes called transfer learning or domain adaptation). In an exploratory
Specific Aim, we will measure machine learning-derived BPA in order to determine if we can uncover novel
imaging biomarkers of CRCI that could not be derived using traditional methods of imaging analysis. Previous
CRCI literature have described disparate effects, with differences in both spatial coherence (focal vs. diffuse) and
strength. This may be due to the heterogeneity of cancers observed, as well as the heterogeneity of treatments.
Due to ENIGMA's unique positioning as a large multi-site group in CRCI, we are able to model and possibly
disentangle some of these factors. The discovery of consistent sub-phenotypes would not only aid predictive
methods, but could provide potential insight for future mechanistic experiments.
9

C. Approach -
C.1. Preliminary Studies
C.1.A. The proposed research project is feasible. For this proposal, we have already formed the ENIGMA
CCWG composed of 10 sites from the US, Europe and Asia as well as imaging and clinical data available from
the UK Biobank Imaging Study (https://www.ukbiobank.ac.uk/). This is the world’s largest population biobank
and it includes brain scans of patients with various diseases such as cancer, heart disease and dementia (Table 1).

Table 1. ENIGMA CCWG Sites
Institution Site Collaborator
Keck Medical Center of USC
Los Angeles, California, USA
Mark S. Shiroishi, MD
Stanford University Medical Center
Palo Alto, California, USA
Kristen Yeom, MD
Ronald Regan UCLA Medical Center
Los Angeles, California, USA
Whitney B. Pope, MD
City of Hope National Medical Center
Duarte, California, USA
Bihong Chen, MD, PhD
MD Anderson Cancer Center
Houston, Texas, USA
Rivka R. Colen, MD
Brown University Medical Center
Providence, Rhode Island, USA
Jerrold L. Boxerman, MD
Huntington Medical Research Institute
Pasadena, California, USA
Kevin S. King, MD
Yantai Yuhuangding Hospital
Shandong, China
Bo Gao, MD, PhD
Shuguang Hospital
Shanghai, China
Chun Ying Zhao, MD
University of Gothenburg
Gothenburg, Sweden
Marie Kalm, PhD
UK Biobank Acquired by ENIGMA

I will utilize the infrastructure and techniques already in existence for ENIGMA under the direction of my
Primary Mentor Dr. Thompson. He is the PI of an NIH Big Data to Knowledge (BD2K) grant (1U54EB020403-
01) to support the ENIGMA Center for Worldwide Medicine, Imaging and Genomics at USC. ENIGMA has
recently published some of the largest neuroimaging studies ever performed (n > 30,000) regarding human
10

hippocampal (Hibar Adams, et al., 2017), and intracranial(Stein et al., 2012) and subcortical structural volumes
(Guadalupe et al., 2016; Hibar et al., 2015), MDD (Schmaal, Hibar, et al., 2016; Schmaal et al., 2015),
schizophrenia (van Erp, Hibar, Rasmussen, Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle,
Hartberg, Gruber, Kraemer, Zilles, Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders,
de Haan, Veltman, Satterthwaite, Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton,
Hass, Calhoun, Bockholt, Sponheim, Shoemaker, van Haren, Pol, et al., 2016) and DTI MRI (Jahanshad et al.,
2013; Kochunov et al., 2015). This body of work demonstrates that this proposal is feasible. ENIGMA now has
more than 30 working groups focusing on 12 major brain diseases including: attention deficit hyperactivity
disorder, autism, addiction, 22q11.2 deletion syndrome, HIV, post-traumatic stress disorder, major depressive
disorder, obsessive compulsive disorder, bipolar disorder, schizophrenia and epilepsy. Working groups also focus
on research methods such as DTI MRI (Figure 1). The ENIGMA Consortium is also in partnership with other
large consortia including the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Cohorts for Heart and
Aging Research in Genomic Epidemiology (CHARGE) Consortium and the IMAaging GENetics (IMAGEN)
Consortium.
Figure 1. ENIGMA Map. The ENIGMA consortium now consists of over 30 Working Groups made up of 500 scientists from over
200 institutions and 35 countries; several of these Working Groups have several ongoing secondary projects led by different
investigators. Here we show 12 of the working groups, focusing on specific diseases and methodologies. Centers where individuals
are scanned are denoted with color-coded pins (legend, bottom left).(Thompson et al., 2015)

11

The ENIGMA CCWG will be the first ENIGMA working group to focus on cancer imaging. As DTI MRI is a
primary component of my proposal, it is noteworthy that my co-mentor Dr. Neda Jahanshad leads the ENIGMA
DTI Working Group (http://enigma.ini.usc.edu/ongoing/dti-working-group/). This group has pioneered DTI
processing methods that can result in reproducible quantitative results that account for inter-site acquisition
differences as well as quality control procedures that can be used in Big Data contexts
(http://enigma.ini.usc.edu/ongoing/dti-working-group/; Jahanshad et al., 2013; Kochunov et al., 2015). The
infrastructure for our research efforts is further enhanced by access to the Big Data for Discovery Science
(BDDS) Center at USC from a separate NIH BD2K grant (1U54EB020406-01) to Dr. Arthur Toga.
C.1.B. Morphometry Preliminary Studies. As part of my KL2 award, we recently published (Shiroishi,
Gupta, Bigjahan, et al., 2017) a paper that was presented at the 2017 International Symposium on Medical
Information Processing and Analysis (SIPAIM) meeting. This was a single-institution retrospective cross-
sectional pilot study at USC of a variety of non-CNS cancer patients treated with and without chemotherapy. We
included adult patients with non-CNS cancers who had undergone clinically-indicated brain MRI from January
2017 to June 2017. Because we did not have healthy control patients imaged on the same MRI scanners as the
cancer patients, our non-cancer controls without a history of malignancy were patients being evaluated for
uncomplicated neurological complaints of headache, dizziness/vertigo etc. without a history of significant
neuropsychiatric disease. We excluded patients with histories of intrathecal or intraventricular chemotherapy,
cranial radiation, metastatic brain tumors and documented neuropsychiatric disorders because many
neuropsychiatric disorders are known to have brain structural abnormalities (Magon et al., 2015; Peng, Chen,
Yin, Jia, & Gong, 2016; Schmaal, Veltman, et al., 2016). For neuroimaging data processing, we employed
established workflows including ENIGMA protocols (http://enigma.usc.edu/protocols/imaging-protocols/).
Statistical Methods: We utilized an age-adjusted linear regression model and the residuals after age adjustment
were used for the group comparisons. Independent t-tests were used for group comparisons using data with a
normal distribution. Wilcoxon Rank Sum tests were used when the data distribution was not normal. The
Benjamini and Hochberg false discovery rate (FDR) was used for multiple comparisons correction to control the
12

false positive rate at 5%. The following group comparisons were made: cancer cases treated with chemotherapy
(Ch+) vs control (C); cancer cases without chemotherapy (Ch-) vs C; subgroup analysis of lung cancer cases with
and without chemotherapy (LCa) was also compared vs C. Results: A total of 64 subjects fit inclusion criteria and
passed scan quality control. There were 45 case and 19 C patients. Of the cancer cases, 22 were scanned after
chemotherapy (Ch+) while 23 were scanned without chemotherapy (Ch-). The most common cancer in our cohort
was lung cancer (n = 16). Our analysis detected statistically significant decrease in either cortical surface area or
cortical thickness in multiple ROIs of primarily the frontal and temporal lobes for the comparisons between
Ch+ vs. C, Ch- vs. C, and LCa vs. C. Effect sizes were variable with the greatest seen in the left middle temporal
surface area ROI (Cohen’s d = -0.690, 95% confidence interval (-1.329, -0.051)) in the Ch- vs C group
comparison.
Utilizing similar methodology, we recently submitted work to the 2018 Organization for Human Brain
Mapping (OHBM) Meeting (currently under review) that compared 125 men with a history of prostate cancer
compared to 125 age-matched controls from data derived from another of our ENIGMA CCWG sites, the UK
Biobank. The base statistical model used imaging metrics as the dependent variable and prostate cancer diagnosis
as an independent variable. Covariates were age at scan, smoking status, hypertension, and diabetes, along with
FreeSurfer intracranial volume (ICV) if the imaging variable was a cortical surface measure. False discovery rate
(FDR) correction was applied to account for multiple comparisons. Prostate cancer diagnosis was associated with
a significant decrease in the pars opercularis surface area (p = 0.039, Cohen's d = -0.340).
C.2.B. DTI Preliminary Studies: Using DTI data from the same UK Biobank prostate cancer and control
cohort as the above morphometry analysis submitted to the 2018 OHBM meeting, 122 subjects with a prostate
cancer diagnosis and 122 age-matched healthy controls who passed quality control were analyzed. Diffusion
tensor imaging metrics were calculated using ENIGMA protocols (http://enigma.ini.usc.edu/ongoing/dti-
working-group/) in each region of interest of the JHU atlas. These were regressed against prostate cancer
diagnosis while co-varying for age, diabetes, hypertension, and smoking status. Prostate cancer diagnosis was
13

associated a decreased fractional anisotropy (FA) in the middle cerebellar peduncles (Cohen's d = -0.34, p = 0.03);
these results were promising but did not survive correction for multiple comparisons in the pilot study.
C.3.B. Machine Learning Preliminary Studies:
Using the same UK Biobank prostate cancer and
control cohort as above and FreeSurfer morphometry
outputs, we applied a machine learning algorithm to
determine brain predicted age (BPA) and deviation
from chronological age. This was based on gradient
boosting regression method on collective normalized
brain metrics to predict brain age (Cole, Poudel, et al.,
2017). The training set was composed of 1,382 non-
cancer controls while testing was performed on 126
prostate cancer and 126 controls subjects. Controls were age-matched to the prostate cancer group in both training
and testing. After regressing out intracranial volume, education, and fluid intelligence and deriving brain age
using the residual structural measures for each hemisphere (thickness, volume, surface area), we found that those
with prostate cancer and hypertension had significantly accelerated BPA (p = 0.015, overall model p < 2.2e-
16) (Figure 2).
C.4. Summary of Preliminary Data: Our preliminary data indicates that by examining a cohort of subjects
with a variety of non-CNS cancer in addition to breast cancer, other types of non-CNS cancers treated with
chemotherapy may result in brain morphometric cortical abnormalities. These were mainly in the
frontotemporal regions. We also found similar findings in those not exposed to chemotherapy, for which only
limited data currently exists. DTI analysis in prostate cancer patients showed promising results of abnormal WM
integrity compared to controls. The overall direction of the findings are, in general, consistent with previous
small breast cancer CRCI neuroimaging studies. Additionally, using a novel application of machine learning
methods, we found significantly accelerated BPA in prostate cancer survivors with hypertension compared
Figure 2. Brain Predicted Age from the UK Biobank Prostate
Cancer Cohort. This histogram depicting machine learning
analysis shows that prostate cancer survivors who have
hypertension demonstrate significantly accelerated brain
predicted age compared to controls.
14

controls. Overall, these preliminary data from 2 sites within the ENGIMA CCWG demonstrate the feasibility
and motivation of this proposal to study many different types of cancer and chemotherapies using well-powered
methods within ENIGMA. The completion of this proposal can help to extend and validate these results. It can
provide new insight into the structural neural substrate of various cancer types and therapies in CRCI.
C.5. Research Design and Methods:
Although ENIGMA has been focused upon the use of imaging genetics techniques, for the sake of feasibility,
we have chosen to focus exclusively on neuroimaging measures at this time. Given the technical difficulty in
measuring structural changes in those with space-occupying brain lesions, we are also excluding patients with
brain tumors.
C.5.A. Study Design Overview
ENIGMA uses a rolling framework for incorporating sites into its working groups (van Erp, Hibar, Rasmussen,
Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle, Hartberg, Gruber, Kraemer, Zilles,
Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders, de Haan, Veltman, Satterthwaite,
Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton, Hass, Calhoun, Bockholt,
Sponheim, Shoemaker, van Haren, Hulshoff Pol, et al., 2016). We have already begun to form the ENIGMA
CCWG (Table 1) that will perform a world-wide, large-scale mega-analysis of cross-sectional neuroimaging
data. This will enable us to reach very large sample sizes into the thousands and carry out the largest study ever
of CRCI.
C.5.B. Study Population
All sites that have already agreed to contribute brain scans to the ENIGMA CCWG are major academic medical
centers in the US, Europe and China which have large oncology practices. These include the MD Anderson
Cancer Center, widely regarded as one of the premier cancer hospital in the United States and 4 other NCI-
designated Comprehensive Cancer Centers. In addition to these sites, the CCWG will incorporate existing scans
and clinical data from the UK Biobank Imaging Study (https://www.ukbiobank.ac.uk/). This will give us access
15

to the world’s largest population biobank which includes patients with various diseases such as cancer, heart
disease and dementia, a large proportion of whom undergo brain scans.
Neuroimaging – An overview of neuroimaging
procedures is given in Table 3.We will be using
neuroimaging data that has already been acquired at each
site for reasons unrelated to this study. Mega-analysis
for the ENIGMA CCWG will be performed centrally
(Figure 3) at the USC IGC using established ENIGMA
harmonized analysis and quality-control procedures
(http://enigma.usc.edu/protocols/imaging-protocols/).
Neuroimaging and clinical data will be transferred via
secure file transfer protocol.
- Morphometric MRI: Each CCWG site will send us
high-resolution T1-weighted structural brain scans. The
cortical GM thickness/surface area/volumes, subcortical
volumes, WM volumes and intracranial volume (ICV) will
be calculated in a standardized fashion using FreeSurer (Fischl et al., 2002) and ENIGMA
protocols(http://enigma.usc.edu/protocols/imaging-protocols/) for mega-analysis.
- DTI MRI: Each CCWG will send us their DTI scans. We will then use standardized ENIGMA DTI protocols
for quality control and mega-analysis (http://enigma.ini.usc.edu/ongoing/dti-working-group/). This protocol has
been successfully utilized world-wide to extract robust measures from diffusion-weighted MRI scans. The DTI
measure that we will use first is fractional anisotropy (FA), one of the most commonly used scalar maps of
white matter integrity. Data analysis steps include conversion of Digital Imaging and Communications in
Medicine (DICOM) images to diffusion-weighted imaging and T1-weighted sets. The T1-weighted images will
be run through FreeSurfer as mentioned above and the DTI will follow with correction for eddy current distortions
Table 3. Overview of Imaging Procedures
Specific Aim 1 Morphometric MRI
Specific Aim 2 DTI MRI
Specific Aim 3 Machine Learning
Figure 3. Overview of ENIGMA CCWG Mega-Analysis
16

and movement using affine registration, mask creation, correction of echo-
planar imaging (EPI)-induced susceptibility artifacts and FA calculation.
- Machine Learning: The T1-weighted scans mentioned above will first
be processed using FreeSurfer. Then segmentations of cortical thickness,
cortical surface area and subcortical volumes based on the Desikan-Killiany
atlas will first visually inspected for accuracy following standardized
protocols. We will regress out sex and intracranial volume (ICV) to
normalize and control for confounds in brain region metrics. To determine
brain predicted age (BPA), and the deviation from chronological age, we will
apply a machine learning algorithm (Cole, Poudel, et al., 2017) using
gradient boosting regression on these collective normalized brain measures
to calculate BPA. This will involve use of 3D convolutional neural networks
using MRI volumes as inputs. We will use 5 repeated blocks of 3D
convolution, rectified linear unit (ReLu) 3D convolution, 3D batch normalization, ReLu and max-pooling
operations and a fully connected final layer in order to generate the regression model to output BPA (Figure 4).
Controls will be age-matched to the cancer group in both training and testing. Model accuracy will be expressed
as the correlation between actual age vs BPA (Pearson’s r), total variance explained (R
2
). Validation and
robustness: Because of the variations in treatments between countries, we will use a holdout set from the United
States (which has a majority of our data) in order to validate our estimated sub-phenotypes. If possible we will
also include an international site, though clearly for some regions extrapolation of treatment effects will be
difficult. Using the international nature of our consortium, it may be possible to further extrapolate which sub-
phenotypes are treatment-related (i.e. linked to the specific type of treatment), and which are population related.
Study Measurements: For Specific Aims 1 - 3, an overview of primary outcome and exposure variables are
provided in Table 4.
Figure 4. Overview of BPA Machine
Learning Workflow
17

C.5.D. Data Analyses: To reiterate, based on the known literature and our own preliminary data, our
hypotheses for our 3 aims are:
Specific Aim 1 Hypothesis: There will be significant decreases in the amount of GM and WM of the
frontotemporal regions of adult non-CNS cancer survivors treated with and without chemotherapy in all
comparison groups.
Specific Aim 2 Hypothesis: There will be decreased regional brain WM integrity of adult non-CNS cancer
survivors treated with and without chemotherapy in all group comparisons.
Specific Aim 3 Hypothesis: Machine learning methods will uncover previously unknown phenotypes, like
accelerated BPA, that may underlie CRCI.

Table 4. Study Measures
Primary Outcome Variable Type of Variable
Brain - morphometric MRI –thickness/area/volume Continuous
Brain - white matter integrity - DTI MRI – fractional anisotropy Continuous
Brain - machine learning – brain predicted age Continuous
Exposure Variables
Does the patient have cancer? Dichotomous, yes or no
Did the patient have chemotherapy? Dichotomous, yes or no
Did the patient have non-chemotherapy treatment (surgery,
radiation, immune-, hormone, targeted therapy, stem
cell transplantation)?

Dichotomous, yes or no
Age Continuous
Sex Dichotomous
Menopausal status for females Dichotomous
Education Continuous
Diabetes or hypertension Dichotomous
Cancer stage at the time of imaging Ordinal
Does the patient have non-CNS metastases at the time of imaging? Dichotomous, yes or no
Duration of cancer illness Ordinal: None, short, long
Time since start of chemotherapy Continuous
18

Our data analysis to test these hypotheses is as follows: For specific Aims 1, 2 and 3. We will perform
comparisons of differences in morphometric, DTI and BPA measures for the following groups: (1) All cancer (±
chemotherapy) vs control, (2) Chemotherapy vs control, (3) No chemotherapy vs control, (4) Chemotherapy vs
no chemotherapy and (5) Most common individual types of non-CNS cancer without chemotherapy compared to
controls. Since this study is not a randomized trial, the group comparisons could be confounded by the factors
such age and other covariates/comorbidities listed above in Table 4. Thus, we will conduct propensity score
matched comparisons. Multivariate logistic regression will be used to generate the propensity score. For each of
the 5 comparisons, a corresponding propensity score will be generated. The greedy, nearest neighbor matching
function will be used for the matching. This means once a match is made the case will not be reconsidered. The
cases will be ordered and sequentially matched to the nearest un-matched control. If more than one un-matched
control matches to a treatment case, the control match is selected at random. The random effects model will be
used to estimate the difference between case and control by considering a nested data structure when individuals
are nested within pairs. The data source site will be considered as the third layer of hierarchical data in the random
effects model. There are 4 major classes of chemotherapeutic agents: alkylating agents, antimetabolites, plant
alkaloids and antitumor antibiotic. Subgroup analysis will be conducted to compare each class of
chemotherapy vs control, and vs no chemotherapy. There are 6 types of non-chemotherapy anti-cancer
therapies that we will also consider: surgery, radiation therapy, immunotherapy, hormone therapy, targeted
therapy and stem cell transplantation. Subgroup analysis will be conducted to compare each class of non-
chemotherapy treatment vs. control, and vs. chemotherapy. The Benjamini and Hochberg procedure will be
used to control the multiple comparison error. For data obtained from the UK Biobank, we will also incorporate
cognitive measures recommended by the International Cognition and Cancer Task Force (Wefel et al., 2011), if
available, from Paired Associates Learning, Trail Making, Word Production and Pairs Matching tests. We
will conduct exploratory analysis for the correlation between the three imaging markers (morphometric, DTI and
BPA) and cognitive measures. A hierarchical linear regression model will be used incorporate the nested data
19

structure and assess the correlation. The interaction term will be used to test the difference in correlation between
the 5 sets of conditions in matching pair comparison above.
D. Power Estimation - The power estimation was conducted using the paired t-test which is equivalent to the
random effects modeling with a 1:1 match. We will have 5 main comparisons and 20 exploratory subgroup
analyses. For the 5 main comparisons, we will use an alpha level of 0.01 for conservative power estimation. For
the exploratory subgroup analysis we will use a small alpha level of 0.001. In the 5 year period, we will recruit
5000 patients. Among those 5000 patients, we should be able to create at least 500 matching pairs for each of the
main comparisons. For 500 matching pairs, we can reach 80% power with a very small effect size of 0.16. Even
with the number of matching pairs dropping to 100 pairs, we can still have 80% power to detect a moderate effect
size of 0.35. For the subgroup analysis, we can have 80% power in detecting moderate (0.35) to large (0.6) effect
sizes with an alpha level of 0.001 with the sample size ranging from 50 to 200 matching pairs. For correlation
analysis, assuming we can only obtain 300 subjects with cognitive measures from 2500 UK Biobank samples, we
can still have 82% power to detect a small effect size of r = 0.2 with a penalized alpha value of 0.01 for multiple
comparison correction. PASS
®
15 was used for all power estimations.
E. Expected Outcomes - I expect to establish a new paradigm for the study of CRCI using Big Data approaches
via the ENIGMA Consortium. This research project has the potential to provide the most robust and
comprehensive assessment of the neuroanatomical substrates that might underlie CRCI in non-CNS cancer
patients treated with and without chemotherapy across a wide variety of different cancer and therapy types. Using
novel application of machine learning algorithms to determine BPA, I expect to uncover new imaging biomarkers
of CRCI not available through conventional analysis. Knowledge gained from this may ultimately improve CRCI
disease monitoring, therapy selection, preventative strategies and risk assessment analysis.
F. Limitations/Potential Problems/Alternative Approaches - As with other ENIGMA studies (B. Franke et al.,
2016; Guadalupe et al., 2016; Schmaal, Hibar, et al., 2016; Schmaal, Veltman, et al., 2016; Thompson et al.,
2015; van Erp, Hibar, Rasmussen, Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle,
Hartberg, Gruber, Kraemer, Zilles, Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders,
20

de Haan, Veltman, Satterthwaite, Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton,
Hass, Calhoun, Bockholt, Sponheim, Shoemaker, van Haren, Pol, et al., 2016), our proposal will not use
prospective harmonization of imaging acquisition. Since this could be a source of heterogeneity, standardized
ENIGMA imaging analysis methods and considerable quality control procedures are in place
(http://enigma.usc.edu/protocols/imaging-protocols/). These steps should ameliorate this concern and this method
has been considered scientifically rigorous in prior ENIGMA publications in high-impact journals (B. Franke
et al., 2016; Guadalupe et al., 2016; Schmaal, Hibar, et al., 2016; Schmaal, Veltman, et al., 2016; Thompson et
al., 2015; van Erp, Hibar, Rasmussen, Glahn, Pearlson, Andreassen, Agartz, Westlye, Haukvik, Dale, Melle,
Hartberg, Gruber, Kraemer, Zilles, Donohoe, Kelly, McDonald, Morris, Cannon, Corvin, Machielsen, Koenders,
de Haan, Veltman, Satterthwaite, Wolf, Gur, Gur, Potkin, Mathalon, Mueller, Preda, Macciardi, Ehrlich, Walton,
Hass, Calhoun, Bockholt, Sponheim, Shoemaker, van Haren, Pol, et al., 2016).
Another potential limitation of my approach includes the heterogeneity of chemotherapy regimens and the
inclusion of other modes of medical therapy such as hormonal and biological therapies. However, our analysis
using both conventional as well as exploratory (machine learning) methods will attempt to uncover differences
that could be attributed to differences in therapies. Furthermore, other confounding factors such as hydration and
nutritional status, exercise, and stress associated with severe illness will not be taken into consideration (Duning
et al., 2005; King et al., 2015; Mondelli et al., 2010; Nakamura, Brown, Araujo, Narayanan, & Arnold, 2014).
Our cancer cohort will have undergone brain MRI scanning for staging purposes, excluding metastatic disease or
as part of an unrelated research protocol. Given the very clinical setting of our CCWG sites, we will be using non-
cancer rather than “healthy” controls as previously described. While we would be missing cancer survivors who
never underwent brain MRI or “healthy” controls, performing such a study on this scale would be prohibitively
expensive. We feel that the unprecedented statistical power that our ENIGMA approach provides carries
enough generalizability to be of scientific value.
While previous ENIGMA efforts have focused on using a meta-analytic approach towards pooling of
neuroimaging data, this proposal will utilize a mega-analytic approach. This will avoid the inherent limitation
21

due to loss of individual patient-level data through meta-analysis and this approach has been used in recent
published ENIGMA studies focusing on bipolar disease,(HibarWestlye, et al., 2017) obsessive compulsive
disorder (Boedhoe et al., 2017), autism spectrum disorder (van Rooij et al., 2017) and attention deficit
hyperactivity disorder (Hoogman et al., 2017).
G. Future Directions - The optimal study design that could validate the results of this proposal would be a multi-
institutional, prospective longitudinal study using both neuroimaging methods over time with cognitive correlates.
The training and experience that I gain from this proposal will provide me with the optimal tools to carry out such
a study Doing so would require an R01-type mechanism for adequate funding and would be individual grants
with each focusing on the most common non-CNS cancers("https://www.cancer.gov/about-
cancer/understanding/statistics,") such as breast cancer, lung cancer, prostate cancer, colon and rectum cancer,
bladder cancer, melanoma of the skin, etc.) The proposed project will ideally serve as pilot data for just such an
application to the NIH or equivalent funding agency.
H. Timeline and Benchmarks for Success -

Benchmark Year 1 Year 2 Year 3 Year 4 Year 5
Rolling recruitment of world-wide ENIGMA CCWG
sites
≥1,000
scans
≥2,000
scans
≥3,000
scans
≥4,000
scans
≥5,000
scans
R01 grant submission** See Table 2. Proposed
Training Activities and Benchmarks in ENIGMA CCWG
in Candidate Section for more details.
Submit R01 grants

22

TECHNICAL DEVELOPMENT – Towards a new method of brain morphometry derived from contrast-
enhanced T1-weighted MR images.
For neuroscience research studies, MRI-based brain morphometry relies on T1-weighted (T1w) scans performed
without the use of gadolinium based contrast agents (GBCAs). While many such non contrast-enhanced (nCE)
research brain MRIs are performed yearly, they are undoubtedly far outnumbered by those obtained clinically. In
this clinical context, contrast-enhanced (CE) T1w scans using GBCAs are frequently obtained for many
indications like cancer. These represent a potential wealth of untapped morphometric data that could be used to
answer many important biomedical questions. Therefore, the ability to exploit this unused vast resource could
represent a major paradigm shift in the neurosciences. However, using these scans is currently problematic
because the signal intensity changes from GBCAs confounds image registration, skull stripping, segmentation,
etc. that are critical in current automated morphometry algorithms.
The neural substrates and biological mechanisms of CRCI remain largely unknown. Understanding these
concepts may ultimately improve disease monitoring, therapy selection, preventative strategies and risk
assessment analysis for cancer survivors suffering from CRCI. Neuroimaging is an obvious approach to
answering these questions. However, the CRCI neuroimaging literature is still in its infancy and the vast majority
of these studies have focused almost entirely on breast cancer survivors after chemotherapy (McDonald & Saykin,
2013; Shiroishi et al., 2017; Simo, Rifa-Ros, Rodriguez-Fornells, & Bruna, 2013). These studies invariably use
small sample sizes and little is known about other types of non-central nervous system (CNS) cancers, as well as
the effects of treatments other than chemotherapy or in the absence of therapy at all. Furthermore, these studies
generally have employed traditional research-dedicated MRI scans without GBCAs. The average statistical power
in the neurosciences is very low (Button et al., 2013). Because it is imperative to produce reproducible results, it
is necessary to conduct studies using large, well-powered cohorts (Ioannidis, 2005). It is in this context that the
large amount of unused clinical brain MRI scans could prove useful. Using clinical scans without GBCAs, our
own recent preliminary work suggests that many types of non-CNS cancers in addition to breast cancer have
significant morphometric cortical brain abnormalities, both treated with and without chemotherapy (Shiroishi et
23

al., 2017). However, because most of these scans in institutions world-wide are used to exclude metastatic disease
to the brain, they are typically CE T1w scans. Therefore, the primary objective of this proposal is to develop and
validate a novel machine (deep) learning algorithm designed to produce accurate measures of brain morphometry
from CE T1w MRI scans. This will be accomplished via the following specific aims:
Specific Aim 1. Develop a machine (deep) learning method to remove the effects of GBCAs from CE T1w scans
to allow accurate brain morphometry.
Specific Aim 2. Validate the results of this deep learning technique, at both the voxel-wise and patient-levels, by
comparing results of morphometry derived from conventional post-processing using nCE T1w images for the
same cancer survivor with various non-CNS cancers treated with and without chemotherapy.
Specific Aim 3. Conduct further validation of by comparing the discrimination power in differentiating these
cancer survivors from non-cancer patients using brain morphometry derived from conventionally post-processed
nCE versus CE T1w images using our deep learning approach.
The proposed work will produce a novel deep learning image processing method that could be used across
multiple institutions and MRI scanners. This can be applied to not only studies of CRCI, but also a whole host of
both normal and pathological processes where already-existing CE T1w MRIs could be a robust source of
inexpensive data and thus be a boon to neuroscience researchers.
RESEARCH STRATEGY
A. SIGNIFICANCE
A.1. The average statistical power in the neurosciences is very low, including those in neuroimaging (Button
et al., 2013). Recently, there has been growing concern that most published research findings may not be true
(Ioannidis, 2005). A major contributor to this concern is due to small sample sizes leading to reduced power. Low
statistical power can then decrease the likelihood of not only finding a true, non-null effect, but can also reduce
the chance that a statistically significant result is actually true (Button et al., 2013). This can then result in inflation
of effect sizes and decreased reproducibility of findings. This has been found in the neurosciences, including
neuroimaging (Button et al., 2013). For example, several decades of neuroimaging studies have been focused on
24

neuropsychiatric disorders like major depressive disorder (MDD), the most common psychiatric disorder, or
schizophrenia. However, most of these studies have been underpowered, and thus the true neuroanatomical
substrates that underlie these conditions remain to be definitively identified (Schmaal, Hibar, et al., 2016; van
Erp et al., 2015).
A.2. Morphometric brain MRI studies rely on research-dedicated non-contrast enhanced (nCE) T1w scans.
This is a limited and expensive resource that is likely a primary reason why most morphometric brain MRI
studies are underpowered. However, in clinics and hospitals world-wide, a tremendous number of brain MRI
scans are performed daily for a variety of clinical indications that could help address the limitation of small
sample size. These large databases of already-existing scans represent a potential wealth of inexpensive MRI
data that could help overcome the pitfalls of underpowered studies and empower researchers to answer many
biomedical questions. The ability to use these clinical scans could represent a major paradigm shift for
neuroimaging researchers in any clinical domain. However, for many clinical indications such as cancer, infection
and demyelinating disease, these MRIs are performed as contrast enhanced (CE) T1w scans using gadolinium-
based contrast agents (GBCAs), Figure 1. Using CE T1w images to calculate brain morphometric measures is
currently difficult because the
signal intensity changes
from GBCAs confounds
image registration, skull
stripping, and segmentation in
current automated processing
algorithms.
A.3. Cancer-related cognitive impairment (CRCI) is emerging as a major public health issue (Bluethmann
et al., 2016). The CRCI neuroimaging literature is still in its infancy and the vast majority of these studies have
focused almost entirely on breast cancer survivors after chemotherapy (McDonald & Saykin, 2013; Shiroishi,
Gupta, Bigjahan, et al., 2017; Simo et al., 2013; Wefel et al., 2015). These studies invariably have used small
Figure 1. CE vs. nCE T1w scans. A. Sagittal and coronal CE T1w MRIs of the brain
compared to B., sagittal and coronal nCE T1w MRIs of the brain. Note the presence of
hyperintense signal from the GBCA in the venous structures in A. that are absent in B.
25

sample sizes derived from dedicated research nCE T1w scans. In addition, little is known about other types of
non-CNS cancer, as well as the effects of treatments other than chemotherapy or in the absence of therapy at all.
B. INNOVATION
This proposal to develop and validate a novel deep learning algorithm for obtaining accurate measures of brain
morphometry from clinically-indicated CE T1w MRI scans can help address low statistical power for a wide
variety of future neuroimaging studies. In this particular instance, we will demonstrate the utility of our approach
for better understanding brain structural differences in survivors of various non-CNS cancers compared to
subjects with no cancer history.
C. APPROACH
C.1. Preliminary Studies
C.1.A. Our preliminary work on brain morphometric data derived from clinically-indicated CE T1w scans
(Shiroishi et al., 2016) suggested that, compared to non-cancer controls, there are significant volumetric decreases
in subcortical structures in survivors of various non-CNS cancers treated both with and without chemotherapy.
The validity of these findings is further supported by our subsequent work using clinically-indicated nCE T1w
that found analogous results in the cerebral cortex in a similar cohort of non-CNS cancer survivors (Shiroishi,
Gupta, Bigjahan, et al., 2017). Furthermore, we have already performed work towards creation of a probabilistic
atlas for CE T1w MR images of the brain designed to assess the validity of automated brain morphometry derived
from CE T1w scans (Shiroishi, Gupta, Faskowitz, et al., 2017).
C.1.B. Preliminary Data Methods
Clinically-indicated CE T1w scans (Shiroishi et al., 2016): Study Design: We performed a HIPAA-compliant
single-institution IRB-approved cross-sectional pilot study at the Keck Hospital of USC across 5 MRI scanners.
A waiver of informed consent obtained. Participants: We searched our neuroimaging database for adults ≤ 60
years of age with non-CNS cancers (cases) and those without cancer (controls) who underwent MRI. Cases were
scanned to evaluate for brain metastases and controls to evaluate headache, dizziness or other uncomplicated
neurological histories. We excluded those with histories of intrathecal or intraventricular chemotherapy, cranial
26

radiation, metastatic brain tumors and documented neuropsychiatric disorders including migraine, depression,
anxiety and other neuropsychiatric conditions. Image Processing: Clinical CE T1w scans prevented several
automatic image processing schemes to run successfully. Manual skull stripping was performed using
BrainSuite15c. Adaptive Non-Local Means de-noising (in ANTS) (Avants, Tustison, & Song, 2009) was
performed on the T1w masked images. Twenty publicly available datasets of healthy adults were obtained and
automatic tissue segmentation of gray and white matter and CSF was performed using FSL’s FIRST. Joint label
fusion using these scans was performed to digitally impute areas of poor T1w contrast in study scans due to
image artifact. FreeSurfer (version 5.3) was run on the imputed the T1w images. QC was performed using
ENIGMA protocols and poorly segmented subjects were excluded. ENIGMA subcortical volume protocols were
used to extract volumes of the lateral ventricles, pallidum, putamen, caudate, hippocampus, thalamus, nucleus
accumbens, and amygdala. Volumes were averaged across hemispheres. Statistical Methods: A random-effect
model was used to test differences between case and control while considering clustered data structure (patient
clustered within MRI scanner) and controlling for age, sex, intracranial volume and cerebrovascular risk factors.
We accounted for multiple hypothesis testing using the Bonferrroni correction. The following group comparisons
were made: All cancer ((chemotherapy (Ch+) and cancer cases without chemotherapy (Ch-)) vs control (C); Ch+
vs Ch-; Ch+ vs C; Ch- vs C.
Clinically-indicated nCE T1w scans (Shiroishi, Gupta, Bigjahan, et al., 2017): Study Design: We performed a
HIPAA-compliant, single-institution, IRB-approved, retrospective cross-sectional pilot study performed at the
Keck Hospital of USC. A waiver of informed consent was obtained. Participants: Our inclusion criteria included
adults ≥ 18 years of age with various non-CNS cancers (cases) and those without cancer (controls) who underwent
brain MRI that included volumetric nCE T1-weighted imaging. Cases were scanned to exclude brain metastases
and controls to evaluate a variety of clinical indications including headache, dizziness or vertigo, among others.
Exclusion criteria included patients with histories of intrathecal or intraventricular chemotherapy, cranial
radiation, metastatic brain tumors and documented neuropsychiatric disorders including histories of migraine,
depression, anxiety and other neuropsychiatric conditions. Image Processing: For each subject, structural T1w
27

scans were performed on one of three 3T MRI scanners. Images were analyzed using FreeSurfer (version 5.3).The
segmentations of 141 regions were extracted. Average values of bilateral regions were also calculated. All
segmentations were visually inspected for accuracy following a thorough and standardized quality control
protocol designed by the ENIGMA Consortium (http:// enigma.ini.usc.edu/protocols/imaging-protocols/).
Statistical Methods: For data not normally distributed, a Wilcoxon score transformation was used so that the
comparison was conducted in a non-parametric fashion. We utilized an age-adjusted linear regression model and
the residuals after age adjustment were used for the group comparisons. Independent t-tests were used for group
comparisons using data with a normal distribution. Wilcoxon Rank Sum tests were used when the data distribution
was not normal. The Benjamini and Hochberg false discovery rate (FDR) was used for multiple comparisons
correction to control the false positive rate at 5%. Effect sizes (Cohen’s d) were also determined. SAS 9.4 was
used for the statistical analysis. The following group comparisons were made: cancer cases treated with
chemotherapy (Ch+) vs control (C); cancer cases without chemotherapy (Ch-) vs C; subgroup analysis of lung
cancer cases with and without chemotherapy (LCa) was also compared vs C.
Creating a probabilistic atlas for
CE T1w MR images of the brain
(Shiroishi, Gupta, Faskowitz, et
al., 2017): We have conducted
preliminary work towards creation
of a novel image processing
pipeline designed to assess the
validity of automated brain
morphometry derived from CE
T1w scans. The data set consists of
50 subjects from the above analysis
on clinically indicated nCE T1w scans (Shiroishi, Gupta, Bigjahan, et al., 2017). Each subjected had both nCE
Figure 2. Atlas creation for CE T1w MR images of the brain.

28

and CE T1w images, where the construction of our CE atlas was aided by the accompanying nCE images. The
CE images were rigidly registered with nCE images with 6 degrees of freedom, Figure 2. Steps 2-3 illustrate the
nCE template construction. Step 4 shows the label transfer from the EVE template to the nCE template. The
transformation in step 1 and 2 were combined to allow for aligning the CE images. Step 3 and 4 were repeated
for CE images. The nCE and the CE templates along with probabilistic ROI definitions are shown on the right of
Figure 2. All registrations were computed using ANTS (Avants et al., 2009).
C.1.C. Preliminary Data Results
Figure 2. Atlas creation for CE T1w MR images of the brain.
Clinically-indicated CE T1w scans (Shiroishi et al., 2016): We analyzed imaging data from 104 total subjects,
including 60 cancer cases. Among them, 35 had had medical therapy (chemotherapy ± biological therapy ±
hormonal therapy). Seven treated patients did not have chemotherapy and had various combinations of biological,
local radiation and hormonal therapy. There were a total of 44 non-cancer controls. Four different group
comparisons were made, Table 1.
We evaluated the average volumes of 8 subcortical regions
because we did not expect to detect laterality differences.
These 8 regions were: thalamus, caudate, putamen, pallidum,
hippocampus, amygdala, accumbens and ventricular volumes.
With 8 hypotheses being tested and a desired overall α = 0.05,
after Bonferroni correction, α = 0.05/8 = 0.00625 for each
individual hypothesis. Our results are summarized in Table 1.
We found that average caudate volume was significantly
decreased in the cancer cases overall compared to non-cancer
controls. There was also a trend toward decreased average
volumes of the hippocampus and amygdala. When we
compared cancer cases with treatment compared to cancer cases without treatment, we did not detect any

Table 1. Subcortical Structures Group Comparison
Results

Comparison Subcortical Volume
All CA vs C ↓ Caudate (p = 0.0007)*
↓ Hippocampus (p = 0.0127)
↓ Amygdala (p = 0.0164)
CA + vs CA - ↓ Amygdala (p = 0.0075)
↓ Accumbens (p = 0.0083)
CA + vs C ↓ Caudate (p = 0.0047)*
↓ Amygdala (p = 0.0008)*
CA- vs C ↓ Caudate (p = 0.005)*
* Denotes statistically significant difference. Volumes
refer to the average volume of the bilateral subcortical
structure. Legend: CA = cancer cases, CA + = cancer
cases with treatment, CA - = cancer cases without
treatment, C = non-cancer controls
29

significant differences, although we did detect near significant
decreases in average amygdala and average accumbens
volumes. We found that the average caudate and average
amygdala volumes were significantly smaller in cancer cases
with treatment compared to controls. Finally, the average
volume of the caudate was significantly smaller in cancer
cases without treatment compared to controls.
Clinically-indicated nCE T1w scans (Shiroishi, Gupta,
Bigjahan, et al., 2017): A total of 64 subjects were ultimately
included in our analysis. There were 45 case and 19 C patients. Of the cancer cases, 22 were scanned after
chemotherapy (Ch+) while 23 were scanned without chemotherapy (Ch-). The most common cancer in our cohort
was lung cancer (n = 16). Six cancer cases had developed 2 different types of cancers. Our analysis detected
statistically significant decrease in either cortical surface area or cortical thickness in ROIs of primarily the frontal
and temporal lobes for the comparisons between Ch+ vs. C, Ch- vs. C, and LCa vs. C. An overview of these
ROIs is listed in Table 2. Effect sizes were generally small to medium, but variable, with the greatest seen in the
left middle temporal surface area ROI (Cohen’s d -0.690, 95% confidence interval (-1.329, -0.051)) in the Ch- vs
C group comparison.
C.1.D. Summary of Preliminary Data
Our preliminary data indicate by using clinically-indicated CE and nCE T1w scans, we found abnormalities
in both subcortical and cortical structures of survivors of various non-CNS cancers treated with and without
chemotherapy compared to non-cancer controls. The subcortical brain regions interact with cortical regions to
coordinate movement(Kravitz et al., 2010) and cognitive functions like memory, learning and motivation
(Pessiglione, Seymour, Flandin, Dolan, & Frith, 2006). The overall direction of these findings are, in general,
consistent with previous breast cancer CRCI studies. However, we appear to be the first to show that subcortical

Table 2. Cortical Structures Group Comparison
Results

Comparison # of ROIs*
Ch+ (N = 22) vs C (N = 19) 5

Ch- (N = 23) vs C (N = 19) 21

LCa (N = 16) vs C (N = 19) 5

* Denotes statistically significant difference in either
cortical surface area or cortical thickness.
30

volumes of structures other than the hippocampus were significantly smaller in the cancer survivors, both with
and without treatment, compared to controls.
Overall, these preliminary data demonstrate the feasibility and motivation for this proposal that aims to develop
and validate a novel deep learning algorithm to accurately perform brain morphometry in clinically-indicated
CE T1w MRI scans. These kinds of scans represent a potential wealth of untapped morphometric data that could
be used to answer many important biomedical questions. Therefore, the ability to exploit this unused vast resource
could represent a major paradigm shift in the neurosciences and be a potential solution to underpowered
neuroimaging studies of many processes, including, but not limited to CRCI.
C.2. Research Design and Methods
C.2.A. Study Design Overview: MRI scans that will be used to develop and validate our deep learning technique
of performing brain morphometry from CE T1w images will be provided from clinical MRI cases performed
within the Department of Radiology at the Keck School of Medicine of the University of Southern California.
The USC Department of Radiology is the among the busiest in the nation. We have multiple MRI scanners in the
department including: 3 GE 1.5 T HD Excite MRI scanners, 2 GE 3T HD Excite Twin MR scanners, 1 Siemens
1.5T Ultra MRI scanner, 1 GE 3T HDXt Excite Twin MRI scanner and 1 Siemens 1.5T Quantum MRI scanner.
We are staffed by full time physicists with pulse programming and post-processing capabilities. The USC
Division of Neuroradiology has 11 faculty attending neuroradiologists, 6 neuroradiology fellows, 5 rotating
radiology residents and multiple visiting research scholars. We also have the image post-processing support of
the 4D Quantitative Imaging Lab. There are more than 40 residents in diagnostic radiology residency program –
making it one of the largest and most competitive in the country. Overall, the department has more than 65 full-
time faculty members, 9 fellowship programs and 18 clinical radiology fellows.
C.2.B. Study Population: For Specific Aims 2 and 3, we will validate the results of our deep learning technique
using brain MRIs of survivors with various non-CNS cancers who have undergone scanning, usually for staging
or to exclude metastatic disease. As a comparison group, we will also utilize non-cancer patients who have
undergone scanning for uncomplicated neurological histories such as headache or dizziness.
31

C.2.C. Study Procedures: Beginning 2 years ago, all of our cancer patients have had both nCE and CE T1w
images for their clinical brain MRI examinations. Both of these types of scans from the same subjects will be
utilized for our 3 Specific Aims. Our inclusion criteria are: adults ≥ 18 years of age with various non-CNS cancers
(cases) and those without cancer (controls) who underwent brain MRI that included volumetric non-contrast-
enhanced T1-weighted imaging. Subjects will be found using the Montage neuroimaging database. Cases are
usually scanned to exclude brain metastases and controls to evaluate a variety of clinical indications
uncomplicated neurological histories such as headache or dizziness that are unlikely to result in morphometric
abnormalities. Exclusion criteria include patients with histories of intrathecal or intraventricular chemotherapy,
cranial radiation, metastatic brain tumors and documented neuropsychiatric disorders including histories of
migraine, depression, anxiety and other neuropsychiatric conditions.
C.2.D. Development of deep learning/deep learning algorithm
Specific Aim 1: In this proposal, we aim to introduce two deep learning strategies for making the best possible
use of the abundance of CE T1w images for cancer survivors available in the clinics. Deep learning is becoming
increasingly useful in the medical imaging community for performing low-level tasks including segmentation of
regions of interest, clustering of voxels or fibers into regions or bundles, and even classification of disease. The
two deep learning strategies we will implement in
this study are stacked denoising autoencoders and
3D U-net.
Stacked Denoising Autoencoder(sDA): As the
name suggests, stacked denoising
autoencoders(Gondara, 2016; Vincent, Larochelle,
Lajoie, Bengio, & Manzagol, 2010) are constructed
by stacking multiple layers of denoising autoencoders (dA), which in turn are essentially a stochastic extension
of classical autoencoders. In denoising autoencoders, the model is forced to learn an uncorrupted input image
reconstruction given a biased input image. We will adopt this model for removing the effect of GBCA induced
Figure 3. Schematic representation of a stacked denoising
autoencoder. The autoencoder has a contracting and an expanding
path. The neural network learns the latent representation of the
image and uses that information to reconstruct the input. It can also
be trained for bias correction, where the input image is a CE T1w
image (left) and the output image is a nCE T1w image (right).

32

intensity bias in the T1w images. As shown in Figure 3, the input to sDA is a CE T1w image and the output is
the corresponding nCE T1 image. The network will be trained on 50 CE and nCE T1w images. As an alternative
approach we could also train the network using 2D image patches. The reason for such a choice is that we believe
that the intensity bias is uniform across the whole image except in the regions of cerebrospinal fluid (CSF), where
the effects are pronounced. Drawing inspiration from Vincent et al. (Vincent et al., 2010) where they tested the
sDA on masked noise and salt and pepper noise, we believe that the sDA will be able to remove the intensity bias
in the CE T1w images. We would like to point out that both masked noise and salt and pepper noise drastically
corrupt small regions of the image leaving parts of the image untouched. The basic premise behind such
autoencoders is that the models are learning a “good representation” of the data. The assumption is that higher-
level representations are relatively stable and robust to local corruption of the input data. The problem presented
in this proposal is in principle the same, i.e., we believe that even though the CE T1w images have local intensity
biases, the higher level representations remain uncorrupted. These learned higher level representations can be
further used for reconstructing the uncorrupted input, the nCE T1w images in our case.
3D U-net: Another deep learning strategy we would like to explore is called 3D U-net (Cicek, Abdulkadir,
Lienkamp, Brox, & Ronneberger, 2016). It is used for generating dense volumetric segmentation from sparsely
annotated datasets. The 3D U-Net was
inspired by the earlier U-
Net,(Ronneberger, Fischer, & Brox,
2015) which deals with multiclass
image segmentation for 2D images.
3D U-Net consists of layers of
convolutional feature maps which are
a set of learnable filters. These filters
learn spatial features like edges,
patterns and orientation. As it can be
Figure 4. 3D U-Net architecture. The green boxes denote feature maps and the
numbers above them shows the number of feature maps. The architecture has a
contracting path on the left and an expanding path on the right. The contracting path
aids in object localization while the expanding path generate the precise high
resolution segmentation.

33

seen in Figure 4, the feature maps are arranged in a contracting and expanding manner with a bottleneck in the
center. The former captures the contextual information, while the latter precisely localizes the segmentation. The
network generalizes reasonably well even with few training examples. This is because each 3D image is
comprised of repetitive structures with little variation. For training, a weighted cross-entropy loss is minimized
using a stochastic gradient descent solver. It is possible to train the network for multilabel segmentation as was
shown by Cicek et al. (Cicek et al., 2016).
Implementation details: Based on our recent experience (Shiroishi, Gupta, Faskowitz, et al., 2017), our training
data set will consist of 50 nCE T1w and their corresponding CE T1w images. All the images will be aligned to
the MNI space. The CE image will be rigidly registered to the nCE image. FreeSurfer segmentations and brain
mask will be generated for the nCE images and the labels will be transferred to the CE images using the
transformations from the previous step. We will then perform data augmentation by flipping, rotating and scaling
as well as applying small distortions to the image using random deformation fields. The image intensity will be
normalized to a range of 0 to 1000. The ROI labels will then be checked for quality by a neuroanatomy expert
and edited if needed. For sDA, the training images consist of co-registered CE T1w images as input and nCE T1w
images as output. For the 3D U-net, CE T1w images and the ROI labels act as input and output, respectively. Our
previous experience with similar problems shows that with a NVidia K40 GPU, the training period should be
between 4 to 6 days. We regularly use TensorFlow and Keras libraries, which has an extensive user base and
support for implementing our deep learning architectures.
C.2.E. Data Analyses
Specific Aim 2: In this validation step, we will use the intraclass correlation ((ICC), two-way mixed with absolute
agreement) to assess the agreement between 1) morphometry measures derived from conventional post-
processing using nCE and 2) CE T1w scans post-processed with our deep learning technique in the same subject.
This will be performed at both voxel-wise and patient levels for the same cancer survivor with various non-CNS
cancers treated with and without chemotherapy. As a sensitivity analysis to examine the robustness of our signal
filtration algorithm, the agreement analysis will be conducted by different scenarios (e.g. by different scanners or
34

by different cancer types), and then explore the distribution of ICC across different scenarios. The ICC and 95%
confidence interval by each scanner will be calculated. We will detect the potential outlier with outstandingly low
ICC.
Specific Aims 3: In this next validation step, we will develop two prediction models using 150 subjects with both
nCE and CE T1w scans. Then we will compare the equivalency of model discrimination power using the area
under the curve (AUC) between the prediction models from nCE vs. CE. Multivariate Adaptive Regression
Spline (MARS) (Friedman, 1991) (Salford Systems, CA) will be used to develop the prediction model. MARS is
a supervised machine learning method. It can process a large amount of candidate predictors without specifying
the model structure, and its output is easier to interpret. To prevent over-fitting, generalized cross validation
(GCV, 10-fold by default) will be performed, which minimizes the influence of outliers and reduces model
complexity to achieve an optimal prediction model with the best generalizability. For the equivalency test, since
the AUC has a Z distribution, it can be conducted using the two one-sided test (TOST) procedure with the
equivalence boundary of the difference in AUC between [-0.05, 0.05]. The robustness of discrimination power
from CE T1w scans will be further validated using 700 independent cases (350 cancer and 350 non-cancer). Since
the 700 independent testing sample has mixed conditions, e.g. different cancer type for the cancer cases, and
different comorbidities for the non-cancer cases, exploratory analysis will be conducted within different
conditions to detect potential scenarios with outstandingly low prediction accuracy.
C.2.F. Sample Size Consideration
Specific Aim 2: We will validate the signal filtration algorithm developed in Aim 1 by assessing the agreement
between 1) brain morphometry measures derived from conventional post-processing using nCE and 2) CE T1w
scans post-processed with our deep learning technique in the same subject. A sample size of 300 scans will
produce a very narrow 95% confidence interval. For example when ICC=0.8, the half width of 95% CI will be
only 0.04.
Specific Aim 3: The equivalence test can be conducted using the two one-sided test (TOST) procedure with the
equivalence boundary of the difference in two paired AUCs between [-0.05, 0.05]. Using a sample size of 300
35

with both nCE and CE (150 non-cancer, 150 cancer) we can achieve 79% power in rejecting the two one-sided
null hypotheses under an alpha level of 0.05 using a very narrow boundary of [-0.05, 0.05], thus claim the
equivalence in the two paired AUCs by assuming the true difference is 0 with a standard division of 0.2.(Phillips,
1990) For further validation of the accuracy of the prediction model based on CE T1w images, an independent
testing sample of 700 (350 non-cancer, 350 cancer) can produce a very narrow 95% CI of AUC with a half width
of 0.02 when AUC is high as 0.9. The half width of 95% CI will increase when the AUC decreases, but even
with a very low AUC as 0.6, the half width of 95% CI is still as narrow as 0.035. PASS15 was used for all power
calculations.
D. EXPECTED OUTCOMES
We expect to establish a novel machine-learning algorithm that will allow accurate brain morphometry measures
to be calculated from clinically-indicated CE-T1w images. Development of this method of brain morphometry
can allow researchers, at marginal expense, to utilize previously untapped clinical brain MRI data that can vastly
boost sample size and statistical power for neuroimaging studies of not only CRCI, but many other types of
biomedical questions.
E. LIMITATIONS/POTENTIAL PROBLEMS/ALTERNATIVE APPROACHES
The deep learning strategies discussed in this proposal have been very successful in the past (Hinton, Osindero,
& Teh, 2006; Krizhevsky, Sutskever, & Hinton, 2012; LeCun, Bengio, & Hinton, 2015; Ren, He, Girshick, &
Sun, 2015; Schmidhuber, 2015; Simonyan & Zisserman, 2014; Szegedy, Ioffe, Vanhoucke, & Alemi, 2016; Xu,
Mo, Feng, & Chang, 2014) and there are no significant limitations to the proposed methods other than a lack of a
huge amount of training data. However, this can be resolved by incorporating different data augmentation
techniques. Training a deep neural network usually requires extensive experimentation with hyper-parameters
(for example: number of feature maps in different layers, choice of the ideal cost function, learning rate, batch
size and activation functions to name a few). Although, an exhaustive search for the optimal parameters for
training the neural network can be difficult, it is usually possible to find a solution. Overall, we feel that the
36

strategies put forth in this proposal are feasible and, if supported, will help us develop further understanding of
CRCI and potentially many other diseases.
F. FUTURE DIRECTIONS
Future work will involve an R01 or equivalent grant application that will be a multi-center study and multiple
different MRI scanners to assess the ability of our deep learning approach to differentiate non-CNS cancer
patients, treated with and without chemotherapy, from non-cancer survivors and healthy controls. We will do
subgroup analysis that will be focused on different types of non-CNS cancer primary types, different classes of
chemotherapy agents as well as other forms of therapy including biological and hormonal therapies.

37

CAREER DEVELOPMENT
My long-term career goal is to do work that can improve the quality of life of non-CNS cancer survivors by
understanding the spectrum of brain structural abnormalities that may underlie cancer-related cognitive
impairment (CRCI). I will be using the statistical power of Big Data neuroimaging methods from the ENIGMA
Consortium, the world’s largest neuroimaging effort. Knowledge gained from this proposal could contribute
towards monitoring CRCI disease progression, improving therapy selection, choosing preventative strategies and
making risk assessments for survivors of cancer. Given the high volume demands of clinical imaging, few
academic radiologists have had the opportunity to obtain formal, rigorous research training. Up to this point, I
have been able to complete most of the coursework and requirements towards the Masters of Science degree in
Clinical, Biomedical and Translational Investigations through the KL2 program. As a result, I have developed
a solid foundation regarding clinical and translational research methods and the responsible conduct of research.
I have also gained some valuable practical experience in morphometric brain MRI acquisition, processing and
analysis. However, in order for me to realize my long-term goal, I have identified a set of short-term
multidisciplinary Learning Goals that cover topics that I will need to master (Table 1). A timeline is included
below (Table 2).
Plan for Career Development/Training Activities:
Mentorship Team: The neuroimaging core (Thompson, Jahanshad) of my superb mentoring team have stayed
constant since my KL2. Drs. Thompson and Jahanshad have worked together for years and have been extremely
successful in obtaining NIH and other substantial extramural funding; publishing in high-impact journals; and
training scores of undergraduate/graduate students and post-doctoral scholars who have gone onto very successful
academic careers. For this project, I have added Drs. Ver Steeg, Gatz and Nieva as co-mentors and Dr. Saykin as
a consultant/external advisor. To bolster the practical experience I will gain by carrying out my proposal, I will
accomplish my short-term Learning Goals through didactic courses, seminars/workshops/scientific meetings
and directed readings suggested by my mentoring team. My short-term Learning Goals include:
38

Consortium Management of the ENIGMA Cancer & Chemotherapy Working Group (CCWG): Until
now, my research experience has been limited to single-institution studies. To achieve my long-term career goal,
I need training in research consortium management. Practical Experience: Building upon the promise of the
preliminary single-center data that I have recently generated, I have begun to assemble my ENIGMA CCWG.
Along with the UK Biobank, I now have 11 sites pledging a total of at least 1,000 MRI during scans the first year.
My Primary Mentor, Dr. Paul Thompson, is an internationally-recognized expert in the performance of massive-
scale neuroimaging studies and development of imaging and mathematical/computational tools for neuroimaging
and genomic data at the Keck School of Medicine of USC. He is the PI and co-founder of the ENIGMA
Consortium, an NIH Center of Excellence that performs the world’s largest brain imaging studies. I have
substantial research infrastructure available to me supported by 2 NIH BD2K awards. During the 1
st
year of the
award period, Dr. Thompson will provide guidance regarding the creation of official organizational documents,
including data analysis plans and explanatory memoranda. Over the course of the training period, he will mentor
me in the building and management of a research consortium from initial development to publication, providing
support to new sites joining the CCWG and overseeing publication and authorship issues. Mentorship Meetings:
I have 110 square feet of office space in the USC IGC that is across the hallway from Dr. Thompson. I will meet
with Dr. Thompson for at least 1 hour each week, and often on a daily basis, to receive guidance and feedback
regarding my research and career development progress. I will also attend monthly ENIGMA BD2K workshops,
weekly ENIGMA meetings and seminars in the USC IGC that will cover new techniques of neuroimaging,
statistics and other pertinent areas.
DTI Neuroimaging: While I have had some valuable practical experience in morphometric T1-weighted
brain MRI methods during my KL2 award period, I have had very little experience in the acquisition and analysis
of DTI MRI, particularly within the context of a multi-center study like with the ENIGMA CCWG. In particular,
I intend to gain expertise in the use of the ENIGMA DTI pipeline(http://enigma.ini.usc.edu/ongoing/dti-
working-group/) including: pre-processing of raw DICOM images using quality controlling to detect
artifacts/abnormalities, efficient processing with respect to acquisition parameters, denoising/artifact removal
39

techniques, template and target creation, skeletonization with tract-based spatial statistics, eddy current
correction, field map distortion correction and tensor fitting, among others. My co-mentor for this Learning Goal
is Dr. Neda Jahanshad, a nationally-recognized expert at the Keck School of Medicine of USC who has expertise
in DTI MRI, large-scale data processing, structural connectivity mapping and novel statistical methods. She has
worked with Dr. Thompson since 2007 and she leads the world-wide ENIGMA DTI Working Group, which aims
to harmonize DTI analysis from disparate acquisition methods.(Jahanshad et al., 2013; Kochunov et al., 2015)
Mentorship Meetings: Dr. Jahanshad’s office is in the USC IGC across the hallway from mine. During my award
period, I will meet with Dr. Jahanshad for at least 1 hour per week, and often on a daily basis. Coursework:
During Year 1 of my award period, I will enroll in Fundamentals of Human Neuroimaging (NIIN 510, 3 units),
a USC graduate course that covers principles and applications of neuroimaging approaches, in particular DTI.
Dr. Jahanshad is a teaching faculty for this course.
Machine Learning: In recent years, machine learning algorithms have become an important approach towards
data mining in areas such as neuroimaging. As a clinical neuroradiologist, I have thus far not had the opportunity
to use machine learning. During my time at the USC IGC under Drs. Thompson and Jahanshad, I have gained
some exposure to machine learning in the last few months. Leveraging their expertise in machine learning with
my background in clinical neuroradiology, we recently applied for a multi-PI NIH application which aims to
develop and validate a machine learning algorithm that will remove the signal from gadolinium-based contrast
agents (GBCA) used in contrast-enhanced (CE) T1-weighted MRI scans. Development of such a technique could
potentially allow neuroimaging researchers to use vast amounts of previously unused clinical CE T1-weighted
MRI scans for morphometric analysis to study a wide variety of conditions. A different application for machine
learning in neuroimaging has been to accurately determine brain-predicted age (BPA) (Cole, Poudel, et al.,
2017). The use of machine learning could identify novel imaging biomarkers to better understand CRCI. My
co-mentor for this Learning Goal is Dr. Greg Ver Steeg, a nationally-recognized expert at the USC Information
Sciences Institute who has expertise in machine learning algorithms of MRI. We will explore the viability of
unsupervised and semi-supervised methods from machine learning in order to find and validate the robustness of
40

empirical sub-phenotypes. These include classical methods such as modularity clustering and spectral clustering,
as well as recently developed methods such as CorEx (https://github.com/gregversteeg) which was developed by
Dr. Ver Steeg. The discovery of consistent sub-phenotypes would not only aid predictive methods but could
provide potential insight for future mechanistic experiments. Mentorship Meetings: Dr. Ver Steeg’s office is
located in the same building as the USC IGC. I will meet with Dr. Ver Steeg ≥1 hour twice each month during
the award period and more frequently during manuscript and grant preparation. Coursework: During Year 1 of
my award period, I will enroll in Machine Learning (CSCI 567, 4 units), a USC graduate course that teaches the
principles and application of machine learning methods in data analysis.
Biology and treatment of cancer: From my medical training and clinical work, I have a good understanding
of the biology and of treatment of cancers that affect the CNS. However, my understanding of non-CNS cancer
biology and treatment is rather limited. Developing this expertise will be very important for my career
development. My co-mentor for this Learning Goal is Dr. Jorge Nieva, MD, a nationally-recognized medical
oncologist at the Keck School of Medicine of USC. He will be a primary resource regarding the pathophysiology
of different cancer types as well as chemotherapies, hormonal and biological therapies and other modes of
treatment. He has had a successful track record of extramural funding including PI of an NIH R01 focusing on
lung cancer. Mentorship Meetings: Dr. Nieva’s office is located on the Keck School of Medicine a short walk
from my office in the Department of Radiology. I will meet with Dr. Nieva ≥1 hour every month during the award
period and more frequently during manuscript and grant preparation. Coursework: During Year 2 of my award
period, I will enroll in Molecular Biology of Cancer (INTD 504, 4 units), a USC graduate course in cancer biology
which covers cancer epidemiology, pathobiology, carcinogenesis, tumor biology, immunology and treatment.
Evaluation of Cognition: As a clinical neuroradiologist, I have not had significant experience regarding the
quantitative evaluation of cognition. Given the nature of this study design, assessment of cognition will not be a
major focus. However, understanding how cognitive functions are related to neuroanatomy will be critical for my
career development. My Co-mentor for this Learning Goal will be Dr. Margaret Gatz, an internationally
recognized clinical psychologist at USC with expertise in the evaluation of cognition. Under her guidance, I will
41

develop expertise in the evaluation of a subject’s neurodevelopmental history and premorbid functioning,
potential malingering, examinee burden and functional level. Particular attention will focus on the types of testing
that can be performed as a neuropsychological screen as well as for comprehensive neuropsychological
evaluation. We will discuss journal articles about cancer and cognition, including: what assessment instruments
were chosen and what domains each assesses; the tradeoffs of comparing scores to norms versus measuring
cognition longitudinally within each individual; and other design considerations when incorporating cognition
into a study of cancer. In addition, I will receive additional mentorship regarding cognitive evaluation from my
external advisor/consultant Dr. Saykin (see below), who has multi-disciplinary expertise in neuroimaging,
neuropsychology and CRCI. A future study will involve a prospective longitudinal design of non-CNS cancer
patients and will incorporate cognitive testing along with neuroimaging; I will be working on this design with
Drs. Gatz and Nieva. Mentorship Meetings: Dr. Gatz’s office is located on the main USC campus, which is a
20-minute drive from Keck School of Medicine. I will meet with Dr. Gatz in her office ≥1 hour every month
during the award period. Coursework: During Year 3 of my award period, I will enroll in Seminar in Clinical
Psychology: Clinical Neuropsychology (PSYCH 660, 4 units), a USC graduate course which covers principles of
neuropsychology and the evaluation of cognitive decline.
External Advisor/Consultant: An external advisor will provide me with a different perspective and offer
additional mentoring regarding my proposal and career development. My external advisor is Dr. Andrew Saykin,
an internationally-recognized neuropsychologist and neuroimaging expert at the Indiana University School of
Medicine. He is the Raymond C. Beeler Professor of Radiology and Imaging Sciences and Director of the Center
for Neuroimaging there. He has a long track record of NIH funding focusing on Alzheimer’s disease, imaging
genomics and CRCI. His background in neuroimaging, neuropsychology and CRCI will be a particular asset for
me as I carry out this proposal. He will also be very helpful by introducing me to other CRCI neuroimaging
researchers who may become collaborators in my ENIGMA CCWG. Dr. Saykin has been a long-time collaborator
with Drs. Thompson and Jahanshad of my mentoring team (Adams et al., 2016; B. Franke et al., 2016; Hibar et
al., 2015; Thompson et al., 2014). Throughout the award period, we have budgeted for monthly 1-hour meetings
42

with Dr. Saykin either via Skype or in person at a meeting such as the Society for Neuroscience (SFN) or the
Organization for Human Brain Mapping (OHBM). These meetings will total 10 hours per year. In addition, I will
also make annual 1-week visits to Dr. Saykin’s laboratory at the Indiana University School of Medicine to have
more personalized mentoring and learn new methods from his research group.
Table 2 provides an overview of the proposed meetings, workshops, coursework, publications and grant
submissions during my award period.
Table 2. Proposed Training Activities and Benchmarks
ACTIVITY YEAR 1 YEAR 2 YEAR 3 YEAR 4 YEAR 5
Meetings/Workshops Monthly BDDS/ENIGMA BD2K Workshops, Weekly ENIGMA meetings (big data neuroimaging analyses,
informatics, computational algorithms, workflows); Annual meetings such as: International Cognition and
Cancer Task Force (multidisciplinary workshop on cognition and cancer), Society for Neuroscience (world’s
largest neuroscience meeting), Organization for Human Brain Mapping (multi-modality human brain mapping),
IEEE Conferences on Big Data (big data collection, distribution, analysis, security), SIPAIM (biomedical imaging,
informatics), MICCAI (medical image computing, acquisition, processing), Radiological Society of North
America (clinically-oriented radiology), International Society for Magnetic Resonance in Medicine (innovation,
development and application of MRI), American Society of Clinical Oncology (clinical oncology), American
Association for Cancer Research (clinical and basic science oncology)
Coursework
NIIN 540 -
Neuroimag. Data
Proc. Methods
CSCI 567 - Machine
Learning

INTD 504 - Mol.
Bio. Cancer

PSYCH 660 - Clin.
Neuropsych.

.

Specialized conferences and workshops
Abstracts &
Manuscripts

Abstract - interim
analysis of
morphometric MRI
analysis

Abstract - interim
analysis of DTI MRI
analysis

Abstract – interim
analysis of
machine learning-
derived brain
predicted age
analyses

Manuscript -
morphometric
analysis of all
cancer and
chemotherapy
comparisons
Manuscripts - morphometric and DTI analysis of common
individual cancers without chemotherapy

Manuscripts – machine learning-derived brain predicted age
analyses

Manuscript - DTI MRI analysis of all cancer and chemotherapy
comparisons

Grants Write and submit multiple R01 proposals for: prospective, multi-center studies of the most
common individual non-CNS cancers (i.e. breast, lung, prostate, etc.) incorporating
multi-parametric MRI (morphometric, DTI, machine learning, resting-state functional MRI),
serum inflammatory cytokine levels, neuropsychological testing; imaging genetics
proposals within CCWG for hypothesis generation.

43

REFERENCES

Abbott, A. (2011). Dementia: a problem for our age. Nature, 475(7355), S2-4. doi:10.1038/475S2a
Abraham, J., Haut, M. W., Moran, M. T., Filburn, S., Lemiuex, S., & Kuwabara, H. (2008). Adjuvant
chemotherapy for breast cancer: effects on cerebral white matter seen in diffusion tensor imaging. Clin Breast
Cancer, 8(1), 88-91.
Adams, H. H., Hibar, D. P., Chouraki, V., Stein, J. L., Nyquist, P. A., Renteria, M. E., . . . Thompson, P. M.
(2016). Novel genetic loci underlying human intracranial volume identified through genome-wide association.
Nat Neurosci. doi:10.1038/nn.4398
Avants, B. B., Tustison, N., & Song, G. (2009). Advanced normalization tools (ANTS). Insight j, 2, 1-35.
Bluethmann, S. M., Mariotto, A. B., & Rowland, J. H. (2016). Anticipating the "Silver Tsunami": Prevalence
Trajectories and Comorbidity Burden among Older Cancer Survivors in the United States. Cancer Epidemiol
Biomarkers Prev, 25(7), 1029-1036. doi:10.1158/1055-9965.EPI-16-0133
Boedhoe, P. S., Schmaal, L., Abe, Y., Ameis, S. H., Arnold, P. D., Batistuzzo, M. C., . . . van den Heuvel, O. A.
(2017). Distinct Subcortical Volume Alterations in Pediatric and Adult OCD: A Worldwide Meta- and Mega-
Analysis. Am J Psychiatry, 174(1), 60-69. doi:10.1176/appi.ajp.2016.16020201
Button, K. S., Ioannidis, J. P., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S., & Munafo, M. R. (2013).
Power failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci, 14(5), 365-
376. doi:10.1038/nrn3475
Cicek, O., Abdulkadir, A., Lienkamp, S. S., Brox, T., & Ronneberger, O. (2016). 3D U-Net: learning dense
volumetric segmentation from sparse annotation. Paper presented at the 19th International Conference on
Medical Image Computing & Computer Assisted Intevention - MICCAI Athens, Greece.
Cole, J. H., Annus, T., Wilson, L. R., Remtulla, R., Hong, Y. T., Fryer, T. D., . . . Holland, A. J. (2017). Brain-
predicted age in Down syndrome is associated with beta amyloid deposition and cognitive decline. Neurobiol
Aging, 56, 41-49. doi:10.1016/j.neurobiolaging.2017.04.006
Cole, J. H., Leech, R., Sharp, D. J., & Alzheimer's Disease Neuroimaging, I. (2015). Prediction of brain age
suggests accelerated atrophy after traumatic brain injury. Ann Neurol, 77(4), 571-581. doi:10.1002/ana.24367
Cole, J. H., Poudel, R. P. K., Tsagkrasoulis, D., Caan, M. W. A., Steves, C., Spector, T. D., & Montana, G.
(2017). Predicting brain age with deep learning from raw imaging data results in a reliable and heritable
biomarker. NeuroImage, 163, 115-124. doi:10.1016/j.neuroimage.2017.07.059
de Ruiter, M. B., Reneman, L., Boogerd, W., Veltman, D. J., Caan, M., Douaud, G., . . . Schagen, S. B. (2012).
Late effects of high-dose adjuvant chemotherapy on white and gray matter in breast cancer survivors:
converging results from multimodal magnetic resonance imaging. Hum Brain Mapp, 33(12), 2971-2983.
doi:10.1002/hbm.21422
Deprez, S., Amant, F., Yigit, R., Porke, K., Verhoeven, J., Van den Stock, J., . . . Sunaert, S. (2011).
Chemotherapy-induced structural changes in cerebral white matter and its correlation with impaired cognitive
functioning in breast cancer patients. Hum Brain Mapp, 32(3), 480-493.
Deprez, S., Billiet, T., Sunaert, S., & Leemans, A. (2013). Diffusion tensor MRI of chemotherapy-induced
cognitive impairment in non-CNS cancer patients: a review. Brain Imaging Behav, 7(4), 409-435.
doi:10.1007/s11682-012-9220-1
Dosenbach, N. U., Nardos, B., Cohen, A. L., Fair, D. A., Power, J. D., Church, J. A., . . . Schlaggar, B. L.
(2010). Prediction of individual brain maturity using fMRI. Science, 329(5997), 1358-1361.
doi:10.1126/science.1194144
Duning, T., Kloska, S., Steinstrater, O., Kugel, H., Heindel, W., & Knecht, S. (2005). Dehydration confounds
the assessment of brain atrophy. Neurology, 64(3), 548-550. doi:10.1212/01.WNL.0000150542.16969.CC
Ferguson, R. J., McDonald, B. C., Saykin, A. J., & Ahles, T. A. (2007). Brain structure and function differences
in monozygotic twins: possible effects of breast cancer chemotherapy. J Clin Oncol, 25(25), 3866-3870.
44

Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., . . . Dale, A. M. (2002). Whole brain
segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron, 33(3), 341-355.
Franke, B., Stein, J. L., Ripke, S., Anttila, V., Hibar, D. P., van Hulzen, K. J., . . . Sullivan, P. F. (2016). Genetic
influences on schizophrenia and subcortical brain volumes: large-scale proof of concept. Nat Neurosci, 19(3),
420-431. doi:10.1038/nn.4228
Franke, K., Gaser, C., Manor, B., & Novak, V. (2013). Advanced BrainAGE in older adults with type 2 diabetes
mellitus. Front Aging Neurosci, 5, 90. doi:10.3389/fnagi.2013.00090
Friedman, J. H. (1991). Multivariate adaptive regression splines. The Annals of Statistics, 19(1), 1-67.
Gaser, C., Franke, K., Kloppel, S., Koutsouleris, N., Sauer, H., & Alzheimer's Disease Neuroimaging, I. (2013).
BrainAGE in Mild Cognitive Impaired Patients: Predicting the Conversion to Alzheimer's Disease. PLoS One,
8(6), e67346. doi:10.1371/journal.pone.0067346
Gondara, L. (2016). Medical image denoising using convolutional denoising autoencoders. Paper presented at
the 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW), Barcelona, Spain.
Guadalupe, T., Mathias, S. R., vanErp, T. G., Whelan, C. D., Zwiers, M. P., Abe, Y., . . . Francks, C. (2016).
Human subcortical brain asymmetries in 15,847 people worldwide reveal effects of age and sex. Brain Imaging
Behav. doi:10.1007/s11682-016-9629-z
Guglielmi, G. (2018). In thousands of brain scans, group seeks clues to diseases. Science, 359(6374), 377-
378. doi:10.1126/science.359.6374.377
Hibar, D. P., Adams, H. H., Jahanshad, N., Chauhan, G., Stein, J. L., Hofer, E., . . . Ikram, M. A. (2017). Novel
genetic loci associated with hippocampal volume. Nat Commun, 8, 13624. doi:10.1038/ncomms13624
Hibar, D. P., Stein, J. L., Renteria, M. E., Arias-Vasquez, A., Desrivieres, S., Jahanshad, N., . . . Medland, S. E.
(2015). Common genetic variants influence human subcortical brain structures. Nature, 520(7546), 224-229.
Hibar, D. P., Westlye, L. T., Doan, N. T., Jahanshad, N., Cheung, J. W., Ching, C. R. K., . . . Andreassen, O. A.
(2017). Cortical abnormalities in bipolar disorder: an MRI analysis of 6503 individuals from the ENIGMA Bipolar
Disorder Working Group. Mol Psychiatry. doi:10.1038/mp.2017.73
Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm for deep belief nets. Neural Comput,
18(7), 1527-1554. doi:10.1162/neco.2006.18.7.1527
Hoogman, M., Bralten, J., Hibar, D. P., Mennes, M., Zwiers, M. P., Schweren, L. S. J., . . . Franke, B. (2017).
Subcortical brain volume differences in participants with attention deficit hyperactivity disorder in children and
adults: a cross-sectional mega-analysis. Lancet Psychiatry, 4(4), 310-319. doi:10.1016/S2215-0366(17)30049-
4
http://enigma.ini.usc.edu/ongoing/dti-working-group/ (Producer).
http://enigma.usc.edu/protocols/imaging-protocols/ (Producer).
https://github.com/gregversteeg.
https://provocativequestions.nci.nih.gov/rfa/mainquestions_listview.html (Producer).
https://www.cancer.gov/about-cancer/treatment/research/understanding-chemobrain (Producer).
https://www.cancer.gov/about-cancer/understanding/statistics.
https://www.cancer.gov/types/common-cancers. Retrieved from https://www.cancer.gov/types/common-
cancers
https://www.ukbiobank.ac.uk/ (Producer).
Inagaki, M., Yoshikawa, E., Matsuoka, Y., Sugawara, Y., Nakano, T., Akechi, T., . . . Uchitomi, Y. (2007).
Smaller regional volumes of brain gray and white matter demonstrated in breast cancer survivors exposed to
adjuvant chemotherapy. Cancer, 109(1), 146-156.
Ioannidis, J. P. (2005). Why most published research findings are false. PLoS Med, 2(8), e124.
doi:10.1371/journal.pmed.0020124
45

Jahanshad, N., Kochunov, P. V., Sprooten, E., Mandl, R. C., Nichols, T. E., Almasy, L., . . . Glahn, D. C.
(2013). Multi-site genetic analysis of diffusion images and voxelwise heritability analysis: a pilot project of the
ENIGMA-DTI working group. NeuroImage, 81, 455-469.
Kesler, S. R. (2014). Default mode network as a potential biomarker of chemotherapy-related brain injury.
Neurobiol Aging, 35 Suppl 2, S11-19. doi:10.1016/j.neurobiolaging.2014.03.036
Kesler, S. R., Rao, A., Blayney, D. W., Oakley-Girvan, I. A., Karuturi, M., & Palesh, O. (2017). Predicting Long-
Term Cognitive Outcome Following Breast Cancer with Pre-Treatment Resting State fMRI and Random Forest
Machine Learning. Front Hum Neurosci, 11, 555. doi:10.3389/fnhum.2017.00555
Kesler, S. R., Wefel, J. S., Hosseini, S. M., Cheung, M., Watson, C. L., & Hoeft, F. (2013). Default mode
network connectivity distinguishes chemotherapy-treated breast cancer survivors from controls. Proc Natl Acad
Sci U S A, 110(28), 11600-11605. doi:10.1073/pnas.1214551110
King, J. A., Geisler, D., Ritschel, F., Boehm, I., Seidel, M., Roschinski, B., . . . Ehrlich, S. (2015). Global cortical
thinning in acute anorexia nervosa normalizes following long-term weight restoration. Biol Psychiatry, 77(7),
624-632. doi:10.1016/j.biopsych.2014.09.005
Kochunov, P., Jahanshad, N., Marcus, D., Winkler, A., Sprooten, E., Nichols, T. E., . . . Van Essen, D. C.
(2015). Heritability of fractional anisotropy in human white matter: a comparison of Human Connectome Project
and ENIGMA-DTI data. NeuroImage, 111, 300-311. doi:10.1016/j.neuroimage.2015.02.050
Koppelmans, V., de Ruiter, M. B., van der Lijn, F., Boogerd, W., Seynaeve, C., van der Lugt, A., . . . Schagen,
S. B. (2012). Global and focal brain volume in long-term breast cancer survivors exposed to adjuvant
chemotherapy. Breast Cancer Res Treat, 132(3), 1099-1106. doi:10.1007/s10549-011-1888-1
Koutsouleris, N., Davatzikos, C., Borgwardt, S., Gaser, C., Bottlender, R., Frodl, T., . . . Meisenzahl, E. (2014).
Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders.
Schizophr Bull, 40(5), 1140-1153. doi:10.1093/schbul/sbt142
Kravitz, A. V., Freeze, B. S., Parker, P. R., Kay, K., Thwin, M. T., Deisseroth, K., & Kreitzer, A. C. (2010).
Regulation of parkinsonian motor behaviours by optogenetic control of basal ganglia circuitry. Nature,
466(7306), 622-626. doi:10.1038/nature09159
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks.
Paper presented at the Neural Information Processing Systems, Lake Taho, NV.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444.
doi:10.1038/nature14539
Lemm, S., Blankertz, B., Dickhaus, T., & Muller, K. R. (2011). Introduction to machine learning for brain
imaging. NeuroImage, 56(2), 387-399. doi:10.1016/j.neuroimage.2010.11.004
Low, C. A., Kalinski, P., & Bovbjerg, D. H. (2015). Neurocognitive Impairment as One Facet of Cancer-Related
Sickness Behavior Symptoms. J Natl Cancer Inst, 107(8). doi:10.1093/jnci/djv176
Luders, E., Cherbuin, N., & Gaser, C. (2016). Estimating brain age using high-resolution pattern recognition:
Younger brains in long-term meditation practitioners. NeuroImage, 134, 508-513.
doi:10.1016/j.neuroimage.2016.04.007
Magon, S., May, A., Stankewitz, A., Goadsby, P. J., Tso, A. R., Ashina, M., . . . Sprenger, T. (2015).
Morphological Abnormalities of Thalamic Subnuclei in Migraine: A Multicenter MRI Study at 3 Tesla. J
Neurosci, 35(40), 13800-13806. doi:10.1523/JNEUROSCI.2154-15.2015
McDonald, B. C., Conroy, S. K., Ahles, T. A., West, J. D., & Saykin, A. J. (2010). Gray matter reduction
associated with systemic chemotherapy for breast cancer: a prospective MRI study. Breast Cancer Res Treat,
123(3), 819-828. doi:10.1007/s10549-010-1088-4
McDonald, B. C., & Saykin, A. J. (2013). Alterations in brain structure related to breast cancer and its
treatment: chemotherapy and other considerations. Brain Imaging Behav, 7(4), 374-387.
46

Menning, S., de Ruiter, M. B., Veltman, D. J., Koppelmans, V., Kirschbaum, C., Boogerd, W., . . . Schagen, S.
B. (2015). Multimodal MRI and cognitive function in patients with breast cancer prior to adjuvant treatment--the
role of fatigue. Neuroimage Clin, 7, 547-554. doi:10.1016/j.nicl.2015.02.005
Mondelli, V., Pariante, C. M., Navari, S., Aas, M., D'Albenzio, A., Di Forti, M., . . . Dazzan, P. (2010). Higher
cortisol levels are associated with smaller left hippocampal volume in first-episode psychosis. Schizophr Res,
119(1-3), 75-78. doi:10.1016/j.schres.2009.12.021
Nakamura, K., Brown, R. A., Araujo, D., Narayanan, S., & Arnold, D. L. (2014). Correlation between brain
volume change and T2 relaxation time induced by dehydration and rehydration: implications for monitoring
atrophy in clinical studies. Neuroimage Clin, 6, 166-170. doi:10.1016/j.nicl.2014.08.014
Pardoe, H. R., Cole, J. H., Blackmon, K., Thesen, T., Kuzniecky, R., & Human Epilepsy Project, I. (2017).
Structural brain changes in medically refractory focal epilepsy resemble premature brain aging. Epilepsy Res,
133, 28-32. doi:10.1016/j.eplepsyres.2017.03.007
Patel, S. K., Wong, A. L., Wong, F. L., Breen, E. C., Hurria, A., Smith, M., . . . Bhatia, S. (2015). Inflammatory
Biomarkers, Comorbidity, and Neurocognition in Women With Newly Diagnosed Breast Cancer. J Natl Cancer
Inst, 107(8). doi:10.1093/jnci/djv131
Peng, W., Chen, Z., Yin, L., Jia, Z., & Gong, Q. (2016). Essential brain structural alterations in major
depressive disorder: A voxel-wise meta-analysis on first episode, medication-naive patients. J Affect Disord,
199, 114-123. doi:10.1016/j.jad.2016.04.001
Pessiglione, M., Seymour, B., Flandin, G., Dolan, R. J., & Frith, C. D. (2006). Dopamine-dependent prediction
errors underpin reward-seeking behaviour in humans. Nature, 442(7106), 1042-1045.
doi:10.1038/nature05051
Phillips, K. F. (1990). Power of the two one-sided tests procedure in bioequivalence. J Pharmacokinet
Biopharm, 18(2), 137-144.
Pomykala, K. L., Ganz, P. A., Bower, J. E., Kwan, L., Castellon, S. A., Mallam, S., . . . Silverman, D. H. (2013).
The association between pro-inflammatory cytokines, regional cerebral metabolism, and cognitive complaints
following adjuvant chemotherapy for breast cancer. Brain Imaging Behav, 7(4), 511-523. doi:10.1007/s11682-
013-9243-2
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards real-time object detection with region
proposal networks. arXiv:1506.01497.
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image
segmentation. Paper presented at the Medical Image Computing and Computer-Assisted Intervention –
MICCAI, Munich, Germany.
Scherling, C., Collins, B., MacKenzie, J., Lepage, C., Bielajew, C., & Smith, A. (2012). Structural brain
differences in breast cancer patients compared to match controls prior to chemotherapy. International Journal
of Biology
4, 3-25.
Schmaal, L., Hibar, D. P., Samann, P. G., Hall, G. B., Baune, B. T., Jahanshad, N., . . . Veltman, D. J. (2016).
Cortical abnormalities in adults and adolescents with major depression based on brain scans from 20 cohorts
worldwide in the ENIGMA Major Depressive Disorder Working Group. Mol Psychiatry. doi:10.1038/mp.2016.60
Schmaal, L., Veltman, D. J., van Erp, T. G., Samann, P. G., Frodl, T., Jahanshad, N., . . . Hibar, D. P. (2015).
Subcortical brain alterations in major depressive disorder: findings from the ENIGMA Major Depressive
Disorder working group. Mol Psychiatry.
Schmaal, L., Veltman, D. J., van Erp, T. G., Samann, P. G., Frodl, T., Jahanshad, N., . . . Hibar, D. P. (2016).
Subcortical brain alterations in major depressive disorder: findings from the ENIGMA Major Depressive
Disorder working group. Mol Psychiatry, 21(6), 806-812. doi:10.1038/mp.2015.69
Schmidhuber, J. (2015). Deep learning in neural networks: an overview. Neural Netw, 61, 85-117.
doi:10.1016/j.neunet.2014.09.003
47

Shiroishi, M. S., Faskowitz, J., D'Amore, F., Emami, A., Hwang, D. H., Cen, S., . . . Jahanshad, N. (2016).
Neuroanatomical changes of non-CNS cancer - a cross sectional study using contrast-enhanced clinically-
indicated MRIs. Paper presented at the Society for Neuroscience, San Diego.
Shiroishi, M. S., Gupta, V., Bigjahan, B., Cen, S. Y., Rashid, F., Hwang, D. H., . . . Jahanshad, N. (2017). Brain
cortical structural differences between non-central nervous system cancer patients treated with and without
chemotherapy
compared to non-cancer controls: a cross-sectional pilot MRI study using clinically-indicated scans. Paper
presented at the SIPAIM - 13th International Symposium on Medical Information Processing and Analysis, San
Andres Island, Colombia.
Shiroishi, M. S., Gupta, V., Faskowitz, J., Bigjahan, B., Cen, S. Y., Hwang, D. H., . . . Jahanshad, N. (2017).
Towards creating a probabilistic atlas for contrast-enhanced T1-weighted MR images of the brain: a pilot study.
Paper presented at the Society for Neuroscience, Washington, D.C.
Simo, M., Rifa-Ros, X., Rodriguez-Fornells, A., & Bruna, J. (2013). Chemobrain: a systematic review of
structural and functional neuroimaging studies. Neurosci Biobehav Rev, 37(8), 1311-1321.
Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition.
arXiv:1409.1556.
Steffener, J., Habeck, C., O'Shea, D., Razlighi, Q., Bherer, L., & Stern, Y. (2016). Differences between
chronological and brain age are related to education and self-reported physical activity. Neurobiol Aging, 40,
138-144. doi:10.1016/j.neurobiolaging.2016.01.014
Stein, J. L., Medland, S. E., Vasquez, A. A., Hibar, D. P., Senstad, R. E., Winkler, A. M., . . . Thompson, P. M.
(2012). Identification of common variants associated with human hippocampal and intracranial volumes. Nat
Genet, 44(5), 552-561.
Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2016). Inception-v4, Inception-ResNet and the impact of
residual connections on learning. arXiv:1602.07261.
Thompson, P. M., Andreassen, O. A., Arias-Vasquez, A., Bearden, C. E., Boedhoe, P. S., Brouwer, R. M., . . .
Consortium, E. (2015). ENIGMA and the individual: Predicting factors that affect the brain in 35 countries
worldwide. NeuroImage. doi:10.1016/j.neuroimage.2015.11.057
Thompson, P. M., Stein, J. L., Medland, S. E., Hibar, D. P., Vasquez, A. A., Renteria, M. E., . . . Drevets, W.
(2014). The ENIGMA Consortium: large-scale collaborative analyses of neuroimaging and genetic data. Brain
Imaging Behav, 8(2), 153-182.
Treanor, C. J., McMenamin, U. C., O'Neill, R. F., Cardwell, C. R., Clarke, M. J., Cantwell, M., & Donnelly, M.
(2016). Non-pharmacological interventions for cognitive impairment due to systemic cancer treatment.
Cochrane Database Syst Rev(8), CD011325. doi:10.1002/14651858.CD011325.pub2
van Erp, T. G., Hibar, D. P., Rasmussen, J. M., Glahn, D. C., Pearlson, G. D., Andreassen, O. A., . . . Turner,
J. A. (2016). Subcortical brain volume abnormalities in 2028 individuals with schizophrenia and 2540 healthy
controls via the ENIGMA consortium. Mol Psychiatry, 21(4), 547-553. doi:10.1038/mp.2015.63
van Erp, T. G., Hibar, D. P., Rasmussen, J. M., Glahn, D. C., Pearlson, G. D., Andreassen, O. A., . . . Turner,
J. A. (2015). Subcortical brain volume abnormalities in 2028 individuals with schizophrenia and 2540 healthy
controls via the ENIGMA consortium. Mol Psychiatry.
van Erp, T. G., Hibar, D. P., Rasmussen, J. M., Glahn, D. C., Pearlson, G. D., Andreassen, O. A., . . . Turner,
J. A. (2016). Subcortical brain volume abnormalities in 2028 individuals with schizophrenia and 2540 healthy
controls via the ENIGMA consortium. Mol Psychiatry, 21(4), 585. doi:10.1038/mp.2015.118
van Rooij, D., Anagnostou, E., Arango, C., Auzias, G., Behrmann, M., Busatto, G. F., . . . Buitelaar, J. K.
(2017). Cortical and Subcortical Brain Morphometry Differences Between Patients With Autism Spectrum
Disorder and Healthy Individuals Across the Lifespan: Results From the ENIGMA ASD Working Group. Am J
Psychiatry, appiajp201717010100. doi:10.1176/appi.ajp.2017.17010100
48

Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., & Manzagol, P.-A. (2010). Stacked denoising autoencoders:
learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning
Research, 11, 3371-3408.
Wefel, J. S., Kesler, S. R., Noll, K. R., & Schagen, S. B. (2015). Clinical characteristics, pathophysiology, and
management of noncentral nervous system cancer-related cognitive impairment in adults. CA Cancer J Clin,
65(2), 123-138. doi:10.3322/caac.21258
Wefel, J. S., Lenzi, R., Theriault, R. L., Davis, R. N., & Meyers, C. A. (2004). The cognitive sequelae of
standard-dose adjuvant chemotherapy in women with breast carcinoma: results of a prospective, randomized,
longitudinal trial. Cancer, 100(11), 2292-2299.
Wefel, J. S., & Schagen, S. B. (2012). Chemotherapy-related cognitive dysfunction. Curr Neurol Neurosci Rep,
12(3), 267-275. doi:10.1007/s11910-012-0264-9
Wefel, J. S., Vardy, J., Ahles, T., & Schagen, S. B. (2011). International Cognition and Cancer Task Force
recommendations to harmonise studies of cognitive function in patients with cancer. Lancet Oncol, 12(7), 703-
708. doi:10.1016/S1470-2045(10)70294-1
Xu, Y., Mo, T., Feng, Q., & Chang, E. I. C. (2014). Deep learning of feature representation with multiple
instance learning for medical image analysis. Paper presented at the IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP).
Yoshikawa, E., Matsuoka, Y., Inagaki, M., Nakano, T., Akechi, T., Kobayakawa, M., . . . Uchitomi, Y. (2005).
No adverse effects of adjuvant chemotherapy on hippocampal volume in Japanese breast cancer survivors.
Breast Cancer Res Treat, 92(1), 81-84.
Button, K. S., Ioannidis, J. P., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S., & Munafo, M. R. (2013).
Power failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci, 14(5), 365-
376. doi:10.1038/nrn3475
Ioannidis, J. P. (2005). Why most published research findings are false. PLoS Med, 2(8), e124.
oi:10.1371/journal.pmed.0020124
McDonald, B. C., & Saykin, A. J. (2013). Alterations in brain structure related to breast cancer and its
treatment: chemotherapy and other considerations. Brain Imaging Behav, 7(4), 374-387.
Shiroishi, M. S., Gupta, V., Bigjahan, B., Cen, S. Y., Rashid, F., Hwang, D. H., . . . Jahanshad, N. (2017). Brain
cortical structural differences between non-central nervous system cancer patients treated with and without
chemotherapy compared to non-cancer controls: a cross-sectional pilot MRI study using clinically-indicated
scans. Paper presented at the SIPAIM - 13th International Symposium on Medical Information Processing and
Analysis, San Andres Island, Colombia.
Simo, M., Rifa-Ros, X., Rodriguez-Fornells, A., & Bruna, J. (2013). Chemobrain: a systematic review of
structural and functional neuroimaging studies. Neurosci Biobehav Rev, 37(8), 1311-1321.
Thompson, P. M., Andreassen, O. A., Arias-Vasquez, A., Bearden, C. E., Boedhoe, P. S., Brouwer, R. M., . . .
Consortium, E. (2015). ENIGMA and the individual: Predicting factors that affect the brain in 35 countries
worldwide. NeuroImage. doi:10.1016/j.neuroimage.2015.11.057

Abstract (if available)

Abstract Given improved survival from cancer, it is critical that we understand the morbidity associated with cancer survival that can impact quality of life. One such issue is cancer‐related cognitive impairment (CRCI). CRCI refers to cognitive dysfunction due to cancer and its treatment. Neuroimaging is an ideal method to study CRCI and we are just beginning to understand the neural substrates that might underlie CRCI. This thesis documents our progress using morphometric MRI to study CRCI using data derived from clinically‐indicated MRIs.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Neuroimaging markers of risk & resilience to brain aging and dementia

PDF

Survival trends and related outcomes of survivors of childhood and young adult cancer

PDF

Vascular contributions to brain aging along the Alzheimer's disease continuum

PDF

Magnetic resonance imaging (MRI) staging for breast cancer in a diverse population

PDF

Novel multi-site brain imaging approaches to map HIV-related neuropathology

PDF

Reproducibility and management of big data in brain MRI studies

PDF

Model development of breast cancer detection and staging via rare event enumeration from a liquid biopsy: a retrospective descriptive clinical research study

PDF

Clinical trial availability for adolescents and young adults with cancer

PDF

Blood-brain barrier pathophysiology in cognitive impairment and injury

PDF

A multi-center prospective cohort study of the diagnostic yield and patient experience of multiplex gene panel testing for hereditary cancer risk

PDF

Neuroimaging in complex polygenic disorders

PDF

Investigation of the neural mechanisms of sensorimotor integration in children with autism

PDF

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

PDF

Instability of heart rate and rating of perceived exertion during high-intensity interval training in breast cancer patients undergoing anthracycline chemotherapy

PDF

Eribulin in advanced bladder cancer patients: a phase I/II clinical trial

PDF

Cytoreductive radical prostatectomy for prostate cancer patients with oligometastatic disease

PDF

Decoding the neurological and genetic underpinnings of chronic pain

PDF

The effects of hormonal exposures on ovarian and breast cancer risk

PDF

Refined screening guidelines for radiation-related late effects in childhood cancer survivors: a dose-volumetric guided approach

PDF

Examination of the effect of oleandrin on head and neck cancer cells

Asset Metadata

Creator Shiroishi, Mark Susumu (author)

Core Title The neuroimaging of cancer and chemotherapy

School Keck School of Medicine

Degree Master of Science

Degree Program Clinical, Biomedical and Translational Investigations

Publication Date 05/01/2020

Defense Date 03/23/2018

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

Tag cancer,cancer‐related cognitive impairment,chemobrain,cognition,complications and late effects of therapy,MRI,neuroimaging,OAI-PMH Harvest,psychological/behavioral oncology

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Thompson, Paul M. (committee chair), Braskie, Meredith (committee member), Jahanshad, Neda (committee member)

Creator Email Mark.Shiroishi@med.usc.edu,mshiroishi@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-499083

Unique identifier UC11268340

Identifier etd-ShiroishiM-6297.pdf (filename),usctheses-c40-499083 (legacy record id)

Legacy Identifier etd-ShiroishiM-6297.pdf

Dmrecord 499083

Document Type Thesis

Format application/pdf (imt)

Rights Shiroishi, Mark Susumu

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