Thesis081914Eierud
-
Upload
cyrus-eierud -
Category
Documents
-
view
116 -
download
0
Transcript of Thesis081914Eierud
Developing Neuroimaging Methods to Disentangle
Mild Traumatic Brain Injury
A Dissertation Submitted to the Faculty of
The Graduate School
Baylor College of Medicine
In Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy
by
Cyrus Eierud
Houston, Texas
June, 2014
ii
APPROVED BY THE DISSERTATION COMMITTEE
____________________________________________ Stephen M. LaConte, Ph.D (Chairman)
____________________________________________ Steven J. Ludtke, Ph.D
____________________________________________ Wei Ji Ma, Ph.D
____________________________________________ Read P. Montague, Ph.D
____________________________________________ Timothy M. Ellmore, Ph.D
____________________________________________ Ramiro Salas, Ph.D
APPROVED BY THE DEPARTMENT OF STRUCTURAL AND COMPUTATIONAL BIOLOGY AND MOLECULAR BIOPHYSICS (SCBMB)
____________________________________________ Wah Chiu, Ph.D (Director of Graduate Studies)
APPROVED BY THE DEAN OF GRADUATE SCIENCES
____________________________________________
Deborah L. Johnson, Ph.D.
_______________________________________ Date
iii
Acknowledgements
Thank you to Dr. Anders Eklund and Dr. Catherine Ngo-Wilde for enlightening
discussions about computers, statistics, fMRI, and traumatic brain injuries. When I
reminisce about my best research discussions, they invariably involve you.
Upon becoming a remote student at Virginia Tech Carilion Research Institute (VTCRI), I
accumulated extra challenges with commuting and handling forms. Thank you, Amy
Jackson, Lourdes Fernandez, Amber Eakin, Meagan Simpson, LaRaun Lindsey, and Ben
Bachman, for all your help solving my problems when I could not physically be in
Houston. Thanks for great guidance and friendship within the SCBMB goes to Ben
Bachman, Caleb Davis, Merry-Lynn McDonald, Rossi Irobalieva, Chris Myers, Kit
Menlove, and many more. At VTCRI it has been a great support to know Olivia Fitch,
Dr. Quentin Fischer, Dr. Susanna Kiss, and Dr. George Wilson.
Thanks also go to my lab members for being supportive and professional about our
research. Apart from Dr. Eklund and Dr. Charles Muller, who directly helped me with
this thesis, I got help from Dr. Cameron Craddock, Sean Fletcher, Manek Aulakh,
Jonathan Lisinski, Brittany Hamilton, Allison McKinnon, Kate McRoberts, and others in
the lab.
iv
Thanks to Wah Chiu for being the most generous and inspiring program chair, as well as
to my strong and persistent committee! Throughout my PhD and especially as I moved to
Virginia, all of the committee members have been rock solid in their support of my
research.
During my first years in the LaConte lab, Dr. Weiji Ma’s lab was next door and his door
was always open. His openness was further proven when he became my Baylor advisor
upon my move to Virginia. Very open, too, was Dr. Tim Ellmore, whom I always could
consult about neuroimaging matters, which was especially important as his research may
have been the closest to mine among the committee members. Even though Dr. Steven
Ludtke’s research may have been furthest away from mine, his expertise in imaging was
rock solid. I believe Steven is the one who, among all of us, best knows the nitty-gritty
details about Fourier Transforms, filters, and other image processing. Most importantly,
Dr. Ludtke has been immensely supportive and stepped up to become my Baylor advisor
after Dr. Ma got a position at New York University. Also, I had the honor of having Dr.
Read Montague on my committee; he was a major reason why I decided to apply to
Baylor and a big asset at VTCRI. Finally, big thanks go to Dr. Ramiro Salas for his
sharpness and friendliness. I was truly surprised how quickly Dr. Salas grasped my work
when he stepped into my committee at last minute to enhance the Baylor College of
Medicine’s presence therein.
v
The final thank you goes to my ingenious advisor, Stephen LaConte, for letting me work
on interesting and foundational low-level fMRI topics, which is why I stepped into this
PhD in the first place. Also, I appreciate your ever-growing support for me throughout
my PhD process.
Thanks, as well, to everybody else I forgot to thank! I believe my PhD experience has
been the strongest learning curve in my entire life, as a result of all of your help.
vi
Abstract
In the nineteenth century, posthumous histology acknowledged pathologic effects caused
by mild traumatic brain injury (mTBI). Despite contemporary impressive neuroimaging
technology, which images brains non-invasively, mTBI injuries remain obscure and are
often untreatable. Some progress after 1990 included studies that detected mTBI in cross-
sectional studies in accordance with meta-analyses (Eierud et al., 2014). Leading
neuroimaging methods in the mTBI field include diffusion tensor imaging (DTI) and
functional MRI (fMRI) (Belanger et al., 2007; Eierud et al., 2014).
In this thesis, verification of consistency across neuroimaging studies using meta-
analyses found significant DTI abnormalities in frontal regions and fMRI abnormalities
in analogous frontal regions (independently). Frontal lobe vulnerabilities were associated
with mTBI before neuroimaging was used (Gurdjian, 1975). In addition, our meta-
analysis found that the DTI results reversed, temporally with time post injury, to become
significant in the opposite direction for chronic mTBI (decreased anisotropy) compared
to acute mTBI (increased anisotropy). Many leading mTBI experts (Belanger et al., 2007;
Hunter et al., 2012; Niogi & Mukherjee, 2010; Prabhu, 2011; Pulsipher et al., 2011),
believe that neuroimaging may compliment or exceed the sensitivity of
vii
neuropsychiatric testing alone. The next natural step would be to use other neuroimaging
than CT clinically to support clinical mTBI diagnosis.
According to previous literature, support vector machines (SVMs) have been proven to
predict multidimensional data accurately and are well applicable for fMRI (Cox & Savoy,
2003; Fan et al., 2007; LaConte et al., 2005; LaConte et al., 2007; LaConte et al., 2011;
Mitchell et al., 2004). In this thesis we characterized SVMs so that they may become
more competitive and useful. Our SVM characterization results found that SVM weight
maps match general linear model (GLM) contrast images spatially and that support vector
regression (SVR) can decode motor task rates from fMRI data (p < 2*10-16). Our final
step would be to specialize SVMs to track mTBI, but before that, the task of imaging
mTBI and creating a model system for mTBI were explored.
There are several problems specifically with mTBI that may be separate from
neuroimaging in general. Two problems are that the mTBI population is heterogeneous
(Rosenbaum & Lipton, 2012; Saatman et al., 2008) and that mTBI patients develop
unique outcomes as various mental disorders (Rosenbaum & Lipton, 2012; American
Psychiatric Association, 2000). That conventional neuroimaging of many mTBI patients
does not show any visible abnormalities is not helpful for mTBI detection. Also, the
Glasgow coma scale (GCS) indicator (practical to clinically classify patients) is too brief
to be of much help to neuroimaging since mTBI develops so heterogeneously (Saatman
et al., 2008). These problems were pro tempore avoided using a cognitive insults model
viii
system for our mTBI methods, yielding reversible cognitive insults that were
significantly detected both behaviorally by use of the executive function and by neural
response. Decrease of executive function has been significantly associated with mTBI
across many studies (McCrea et al., 2003; Rohling et al., 2011). In addition, the model
system helped us to significantly detect practice effects in our executive function task (p
< 0.017, 1-sided, by behavior and p < 0.05, 2-sided and with multiple comparisons
correction, by neural response). Practice effects were not supposed to exist in the multi-
source interference task, which we used according to previous literature (Bush et al.,
2003; Bush & Shin, 2006; Sheu et al., 2012). Unfortunately, this practice effect altered
our final mTBI analysis model. However, some behavioral results from our longitudinal
mTBI subjects showed that they significantly improved their reaction times with time
post injury. Future analysis will have to disentangle how much that improvement was
from mTBI recovery or from practice effects.
ix
Notations and Abbreviations
X matrix containing feature vectors
x feature vector
w weight vector
y vector of labels
x scalar
t time point (often for a sample)
T total number of t and number of samples
x
Table of Contents
Approvals ........................................................................................................................ ii
Acknowledgements ........................................................................................................ iii
Abstract .......................................................................................................................... vi
Notations and Abbreviations .......................................................................................... ix
Chapter 1: Introduction and Background ........................................................................ 1
Chapter 2: Neuroimaging Meta-Analysis of Mild Traumatic Brain Injury .................. 20
Introduction ............................................................................................................... 21
Materials and Methods .............................................................................................. 23
Results ....................................................................................................................... 27
Discussion ................................................................................................................. 43
Chapter 3: Detecting Motor Rate using Support Vector Machines .............................. 51
Introduction ............................................................................................................... 51
Materials and Methods .............................................................................................. 65
Results ....................................................................................................................... 69
Discussion ................................................................................................................. 87
Chapter 4: mTBI Model System Using Executive Function ........................................ 89
Introduction ............................................................................................................... 89
Materials and Methods .............................................................................................. 92
Results ..................................................................................................................... 100
Discussion ............................................................................................................... 109
Chapter 5 Longitudinal mTBI ..................................................................................... 113
xi
Introduction ............................................................................................................. 113
Materials and Methods ............................................................................................ 114
Results ..................................................................................................................... 115
Discussion ............................................................................................................... 117
Chapter 6: Summary, Significance, and Future Goals ................................................ 118
References ................................................................................................................... 122
xii
List of Figures
Figure 1: Hydrogen net magnetization ............................................................................. 12
Figure 2: Tissue-specific T1 and T2 relaxations ................................................................ 13
Figure 3. Hemodynamic response. .................................................................................... 17
Figure 4: Overview of mTBI and neuroimaging .............................................................. 28
Figure 5: Time course of structural neuroimaging studies ............................................... 30
Figure 6: Time course of functional neuroimaging studies .............................................. 31
Figure 7: Themes studied with different imaging modalities ........................................... 33
Figure 8. Activation likelihood estimate (ALE) analysis of fMRI mTBI publications .... 36
Figure 9. Anisotropy abnormalities mapped spatially in mTBI ....................................... 38
Figure 10. Anisotropy compared temporally in relation with mTBI or performance ...... 42
Figure 11. Supervised learning ......................................................................................... 56
Figure 12. Hypothetical classification and feature space scatterplots .............................. 57
Figure 13: Hyperplanes layout for DAG multi-classification ........................................... 61
Figure 14: Direct acyclic graph (DAG) decision tree ....................................................... 62
Figure 15: Hypothetical support vector regression scatter plot ........................................ 64
Figure 16. Stimuli paradigm for real-time motor rate effect ............................................ 68
Figure 17. Group behavior motor rate mean with standard deviation .............................. 71
Figure 18. Support vector regression prediction variance ................................................ 72
Figure 19: Similarities between GLM, SVC, and SVR statistical brain patterns ............. 73
Figure 20: Support vector regression of single subject ..................................................... 79
Figure 21. Support vector regression errors per subject graphed ..................................... 80
xiii
Figure 22: Comparing support vector classification to regression prediction errors ........ 84
Figure 23: Simulated SVM signal and results .................................................................. 85
Figure 24: Theoretical and empirical response time series for signal variations within
rates ........................................................................................................................... 86
Figure 25: Behavioral improvement after real-time feedback .......................................... 88
Figure 26: Multi-source interference task (MSIT) ............................................................ 95
Figure 27. Sleepiness according to the Stanford sleepiness scale (SSS) .......................... 96
Figure 28: Behavioral response from MSIT cognitive insult treatments ........................ 101
Figure 29: MSIT behavioral practice effects .................................................................. 103
Figure 30: MSIT neural contrast image of control case ................................................. 105
Figure 31: Neural contrast image of MSIT practice effects ........................................... 106
Figure 32: Neural contrast image of MSIT sleep-deprivation treatment ........................ 108
Figure 33: Real-time MSIT feedback ............................................................................. 111
Figure 34: mTBI reaction time improvements with time post-concussion ..................... 116
xiv
List of Tables
Table 1. mTBI vs. control cluster centroids with FDR (p < 0.05) using ALE ................. 37
Table 2. Center of mass of white matter structures in the ICBM-81 atlas ........................ 39
Table 3. Cluster volumes in μ-liters using support vector machine ................................. 76
Table 4. Support vector classification accuracies from 1 to 5 Hz per subject .................. 78
Table 5. Support vector regression correlation and root mean square error (RMSE)
between predicted and true rate ................................................................................ 81
Table 6: Behavioral data from the MSIT sleep-deprived group ....................................... 97
Table 7: Behavioral data from the MSIT pure control group ........................................... 98
Table 8. Multi-source interference task behavior statistics ............................................ 102
1
Chapter 1: Introduction and Background
In the fourth century BC the Hippocratic Corpus (an ancient Greek medical work)
described and named a disorder similar to mild traumatic brain injury (mTBI). “mTBI”
and “concussion” may be used interchangeably and signify a mental state after a blow to
the head. Most mTBI patients recover, but citing the Centers for Disease Control and
Prevention (CDC) (Alexander, 1995; Kushner, 1998), “15% of patients diagnosed with
mTBI may have experienced persistent disabling problems,” which is the same as post-
concussion syndrome (PCS). Although criteria have been established by the Diagnostic
and Statistical Manual of Mental Disorders IV (APA, 2000) and International Statistical
Classification of Diseases and Related Health Problems (ICD-10), PCS is difficult to
diagnose, and its symptoms are nonspecific. Although mTBI has long been considered a
noncritical injury, serious short-term (McCrea et al., 2003; Rohling et al., 2011) and long-
term effects have been documented (Kraus et al., 2007; Geary et al., 2010). In addition,
there is broad acceptance that multiple mTBIs can have serious long-term consequences
(Guskiewicz et al., 2003). It is hypothesized that mTBI includes neural contusions
(Adams et al., 1980; Beaumont and Gennarelli, 2006; Brandstack et al., 2006; Levin et
al., 1992) or axonal injury (Buki and Povlishock, 2006; Gennarelli et al., 1982; Meythaler
et al., 2001; Povlishock et al., 1992).
2
In 2014 the Brain Injury Association of America (BIAA; the largest TBI organization)
wrote on its home page, “Brain injury is not an event or an outcome. It is the start of a
misdiagnosed, misunderstood, under-funded neurological disease” (BIAA, 2014). Even
though the BIAA includes moderate and severe TBI, mTBI is taken very seriously for
several reasons including that mTBI is a more frequent TBI disability compared to
moderate and severe TBI (Cassidy et al., 2004; Langlois et al., 2004).
About a million people are affected by mTBI every year (Cassidy et al., 2004; Langlois et
al., 2004). Until recently, work to track or alleviate the effects of mTBI has been
unfruitful. Many other important injuries are successfully treated using modern
technology, while mTBI injuries are obscure and often untreatable despite impressive
improvements in neuroimaging and other mTBI-related techniques. Neuroimaging,
except for computed tomography (CT), is seldom used to support diagnoses of mTBI or
PCS. Note that the CT used to image mTBI subjects is primarily used to detect
hematomas, but not to diagnose mTBI per se. It would also be interesting if neuroimaging
could help us understand how to cure chronic mTBI disabilities. Still, tools to peek into
the brain have been explored, and studies have found relevant patterns that detect mTBI
at group levels. Currently DTI (Niogi & Mukherjee, 2010) and fMRI (McAllister et al.,
2006; Ptito et al., 2007) are promising in the mTBI research field among the different
imaging modalities. Animal studies suggest that white-matter neural tissue is affected by
mTBI and that DTI may detect this damage (MacDonald et al., 2007). In addition, DTI
studies have found significant differences between mTBI and control groups in 18 of 21
3
studies (Eierud et al., 2014). In addition, seven fMRI-mTBI studies have found
significant differences between mTBI and control as well (Eierud et al., 2014). The PCS
diagnosis has been debated for hundreds of years (Evans, 1992). Today, with increasingly
impressive neuroimaging tools, we still cannot pinpoint a major PCS marker using
neuroimaging and many mTBI disorders are still only confirmed posthumously using
histology or other techniques (Evans, 1992).
A few important challenges of mTBI and PCS, may include following. First, mTBI can
affect any spatial structure in the brain (Guskiewicz et al., 2007; Rosenbaum & Lipton,
2012) and may cause different mental disorders depending on what structure the injury
affects. Secondly, it is hard to pair the subjective information reported by the people
suffering from PCS with the objective image; e.g., it is believed that athletes at times do
not want to miss any games and play even if their physicians recommend them not to.
Conversely, people without medical insurance may exaggerate their symptoms to receive
economic benefits (Paniak et al., 2002).
Based on mTBI and neuroimaging, this thesis chapter focuses on four areas.
(1) Our first focus is that we used the PubMed database for a meta-analysis, finding
significant DTI and fMRI markers that have been hypothesized in literature.
(2) The brain images used to detect mTBI are either 3D or 4D (when time is included),
often with more than 100,000 voxels (3D pixels) per image. It has been shown in many
cases that SVM predicts multi-dimensional data very well, including for fMRI (Cox &
4
Savoy, 2003; Fan et al., 2007; LaConte et al., 2005; LaConte et al., 2007; LaConte et al.,
2011; Mitchell et al., 2004). Using multi-dimensional data for classification is rather new
and may have started in 1936 when Ronald A. Fisher wrote a pioneering article
suggesting pattern recognition, which is closely related with early literature leading to
SVM (Vapnik & Lerner, 1963). To explore whether SVM is suitable to track mTBI is the
second major focus of this thesis; e.g., to explore whether neuroimaging of a person with
mTBI could predict a future of recovery or chronic disability, which may benefit PCS
diagnosis.
(3) The third focus was to create an mTBI model system to improve our mTBI methods.
First, the neuroimaging sensitivity to detect mTBI or its repercussions is almost strong
enough to be supportive for clinical practice. However, in order to validate a novel
method, there is an advantage to using data from an unquestioned source yielding a clear
signal. Although the end goal is to detect mTBI, mTBI is hard to control for, and an
alternate reversible neural condition that affects the executive function was used to model
mTBI at least from a naïve perspective. The naïve perspective signifies that there is no
requirement that the brain pattern match that of mTBI. We found that sleep deprivation
fulfilled our requirements as a naïve mTBI model system. This model system enabled us
to characterize the efficacy and strengths of our methods without having to induce mTBI
in humans.
(4) The fourth part of this thesis was to tie the knowledge from the meta-analysis, SVM,
and model system to analyze our own mTBI subjects that we have collected longitudinal
data of within a year following their concussions.
5
The intermediate goal of this thesis is to characterize SVM and then to evaluate how well
the SVM methods apply to the cognitive insults model. The final goal is to use the
sensitive SVM methods developed to track mTBI.
Mild Traumatic Brain Injury
Pioneering mTBI research has found significant symptomatic and executive function
differences between the mTBI group and a control group within a couple of days after
concussion (McCrea et al., 2003; Rohling et al., 2011). Since the last two decades, many
experts have categorically imaged mTBI subjects in contrast with subjects in a control
group to find significant differences as well. Imaging is not used to diagnose mTBI itself,
but to test for hematomas as well as to rule out head injury complications from more
severe trauma. Various guidelines for diagnosing mTBI exist, most of which rely on the
Glasgow Coma Scale (GCS; Teasdale & Jennett, 1974; Holm et al., 2005; Rosenbaum &
Lipton, 2012) and details of the injury (such as self- and witness-reported descriptions of
the accident, loss of consciousness, and evaluation of sustained trauma; Ruff et al., 2009).
The GCS assesses motor, verbal, and eye responses; while there is some variability in the
categories, a GCS 8 or below is often considered severe, from 9 to 12 is often considered
a moderate TBI, and from 13 to 15 is considered a mild TBI (Jennett, 1998; Parikh et al.,
2007). Ultimately, the diagnosis of TBI and its severity is made by a clinician. Although
mTBI has long been considered a noncritical injury, serious short- and long-term effects
have been documented (Alexander, 1995; Kraus et al., 2007; Kushner, 1998; Lipton et
6
al., 2009). Additionally, there is broad acceptance that multiple mTBIs can have serious
long-term consequences (Guskiewicz et al., 2003). After initial injury, secondary
mechanisms elicit biochemical, metabolic, and cellular changes in the time frame of
minutes, days, and months (Giza and Hovda, 2001; Loane and Faden, 2010; Xiong et al.,
1997). Within the first 15 minutes post-injury, there is an extreme dip in
neuropsychological performance (McCrea et al., 2002), and deficits can often linger for a
week or longer (McCrea et al., 2003; Rohling et al., 2011). The definition of the acute
time frame varies across publications, and some studies report acute periods of up to 1
month post-injury (Landre et al., 2006). This thesis uses the term “acute mTBI” for up to
2 weeks post-injury. Using the term “acute” or “semi-acute” for periods up to 2 weeks
post-injury is common in the literature (Gasparovic et al., 2009; Mac Donald et al., 2011;
Mayer et al., 2010; Messe et al., 2011).
Most mTBI patients recover, but a substantial minority (perhaps 15%) have
persistent disabling problems (Alexander, 1995; Kushner, 1998), known as PCS.
Although criteria have been established by the Diagnostic and Statistical Manual of
Mental Disorders IV (APA, 2000) and ICD-10, PCS is difficult to diagnose, and its
symptoms are nonspecific. PCS also manifests symptoms similar to those of other
disorders such as major depression (Iverson, 2006; Iverson and Lange, 2003), chronic
pain (Smith-Seemiller et al., 2003), and other diseases such as somatization disorder.
Indeed, neuropsychological testing in chronic stages of mTBI (even on the time scale of
months) has been criticized as nonspecific and insensitive (Iverson, 2005; McCrea and
American Academy of Clinical Neuropsychology, 2008), and several studies have
7
questioned the ecological validity of these assessments (Satz et al., 1999; Silver, 2000)
and proposed improved approaches for detecting persisting cognitive deficits and linking
these to neuroimaging results (Geary et al., 2010). Heterogeneity of injury and current
limitations in the sensitivity of imaging are challenges to developing diagnostic tools as
well as predictors of recovery. Thus, the physical size and heterogeneous distribution of
injury in the brain make detection in an individual challenging and further make reliance
on group averages problematic. In addition, since the time course of the injury leads to
lingering post-concussive symptoms in a small number of injuries (Alexander, 1995;
Kushner, 1998), it is a statistically challenging goal to try to predict which individuals
will not recover fully. Finally, longitudinally, the presence or absence of CT findings
does not correlate with long-term outcomes such as PCS (Hanlon et al., 1999; Huynh et
al., 2006; Kurca et al., 2006; Lee et al., 2008; McCullagh et al., 2001; Tellier et al.,
2009). To our knowledge there are no existing publications that claim to significantly
detect mTBI in a single subject using neuroimaging. To summarize, imaging is
challenging at both acute and chronic stages of mTBI, and attempting to characterize the
full time course compounds the level of complexity. Despite the challenges, there has
been a growing research effort to characterize structural and functional effects of mTBI.
A full range of neuroimaging technologies have been brought to bear on this issue,
including CT, positron emission tomography (PET), single-photon emission computed
tomography (SPECT), magnetoencephalography (MEG), electroencencephalography
(EEG), and 12 subtypes of Magnetic resonance imaging (MRI), such as diffusion tensor
imaging (DTI), magnetic resonance spectroscopy (MRS), arterial spin labeling (ASL),
8
and fMRI. This thesis has information about most imaging modalities related to mTBI,
but mostly it will focus on fMRI and DTI.
Magnetic Resonance Imaging
MRI evolved from nuclear magnetic resonance (NMR) spectroscopy, which was first
developed to image chemical compositions. The MRI differs mainly in two senses. First,
the MRI is expanded to be able to independently receive measurements at several spatial
locations as opposed to from a single point. Secondly, it is designed to image a single
atom type, often protium (hydrogen), since it is very abundant in living tissues. Protium
is the most abundant hydrogen (99.98%). Sixty-three percent of all atoms in the human
body are protium (Freitas, 1998), which is also MRI active. Body fluids and fatty tissues
are extremely protium abundant. Further on, this thesis focuses on protium because
signals from any other atoms are negligible in this case. MRI can detect many different
tissue properties, but most importantly it can create an image. Since the first human MRI
study (Damadian, 1974) was published, MRI has evolved into the preferred clinical
imaging method. Quantum physics, cryogenics, superconduction, strong and precise
magnetic fields, Fourier transforms, radio-wave emitters and receivers, and vast
complicated calculations are used to generate MRI. Today MRIs are acquired similarly to
the method used in the first animal MRI study (Damadian, 1971), with the exception of
using magnetic field gradients (Lauterburg, 1973) and Fourier transforms formalism
(Likes, 1981), which have radically improved the speed with which an image could be
9
acquired. The following short description of how MRI is acquired may be read in more
detail (Westbrook et al., 2013). The process includes four steps:
(1) The subject needs to be positioned in a strong and homogenous magnetic field
called the B0 field. Many images presented in this thesis were acquired by a 3-
tesla magnet, which is about 60,000 times stronger than the earth’s magnetic field.
Each hydrogen atom has two magnetization poles, like a bar magnet. A strong
external magnetic field (B0) aligns a majority of the hydrogen atoms parallel to
the magnetic direction of B0, and the rest are antiparallel to B0 (Figure 1B). In
addition, the B0 field sets the hydrogen precession frequency (!) to ! = !×!,
where ! is the gyromagnetic ratio (42.57 MHz/T for hydrogen) and ! is the
external magnetic field (often B0 = 3 T). However, at this point all hydrogen
atoms are precessing at random phase, so there is no net magnetization vector in
the rotating (precessing) transversal (xy) plane. Exposed to B0, this hydrogen
configuration yields the lowest energy level within any measured macroscopic
region, which has a constant net magnetization vector pointing in the same
direction as the B0 field (Figure 1B).
(2) Applying a second magnetic gradient (B1) across the subject so that one part of
the tissue has less than 3T and the other part has more than 3T results in a slice
with exactly 3T between the two tissue parts. After the desired slice position in
space is chosen (arbitrarily controlled by varying B1), a specific radio pulse (RF)
excites the hydrogen atoms only in the 3T-slice (Figure 1C,D). The energy
excitation, which can be described as the net magnetization vector (NMV)
10
shifting from pointing in the same direction as the B0 field (the lowest energy
level) to rotating in the transversal plane (orthogonally to the B0 field). After the
excitation and slice is selected the B1 field is removed leaving just the B0 field.
The NMV rotation speed is the hydrogen’s precession frequency. Hence,
hydrogen atoms are now precessing in phase, as opposed to before the excitation
(when phases were completely random). This excited slice will gradually relax to
its lowest energy level (as it was before the excitation) depicted in Figure 1B. The
relaxation depends on the tissue-specific time constant T1, which is the time it
exponentially will recover 63% of its original NMV (Figure 1B). For example, for
cerebrospinal fluids, T1 may be around 4 s. Before any significant T1-related
relaxation occurs, the NMV rotates in the transversal plane, yielding a measurable
signal. However, another relaxation occurs that is much faster than the T1-related
relaxation, which is that the atoms’ precessions, within the excited slice, will
dephase from each other. This dephasing decays the signal even faster than the
T1-related relaxation, and has a tissue-specific time constant named T2, which
decays to 37% of its maximum at time T2. Tissues affect the T1 and T2 constants
differently. Therefore it is possible to change the image contrast (to enhance
different tissues) by weighting reception from either the T1 or the T2 signal. A few
tissues’ T1 and T2 exponential relaxations are depicted in Figure 2.
(3) The B1 field that was turned off after excitation can again be used, but now to
encode the precession frequency spatially across the excited slice. In addition, the
signal from the spatially encoded slice can then be received and decoded into
11
meaningful images. The process of how to encode and decode the signals into
images is called sequencing and may be done in many different ways to enhance
different tissue properties. Sequencing, that may use other techniques not
mentioned here, is out of scope for this thesis, and may be found in other
literature (e.g. Bernstein et al., 2004). One note is that to use the Fourier transform
to encode and decode MRI-sequences is extremely practical (Likes, 1981).
(4) For most MRIs, the magnitude of the inverted Fourier transform of the signal
yields the raw image that is a projection of the brain that may be displayed. Still,
many corrections may be needed for statistical imaging analysis. We will discuss
a few of those corrections in the section on preprocessing below.
The MRI repetition time (TR) has different meanings depending on different sequences.
However, in this thesis the TR will always refer to the echo planar imaging (EPI) version
of TR that is used for fMRI acquisition. In this case the entire brain volume is imaged
within the time represented by the TR (often 2 s).
12
Figure 1: Hydrogen net magnetization (A) If no significant external magnetic field exists, hydrogen atoms align their magnetic poles randomly. The magnetic vectors point in the direction of the atoms’ magnetic field. (B) In a strong homogenous external magnetic field, B0 (e.g., 3T) atoms’ magnetic vectors are only parallel or antiparallel to B0. However, most atoms, in the macroscopic volume, align in the same direction as B0, generating a static net magnetization vector (NMV). (C) Using a specific radio frequency (RF) pulse, the NMV is rotated 90° into the transversal (xy) plane that also is an excited state (energy-wise). (D) In the transversal plane the NMV rotates in sync with hydrogen’s precession frequency. (E) The NMV is really composed of many hydrogen atoms precessing in phase. However, the synced phases among the atoms deteriorate (F) with time according to the T2 constant, which is also tissue specific.
13
Figure 2: Tissue-specific T1 and T2 relaxations Different T1 and T2 time constants representing the relaxation curves for different tissues. Graph shows the salient intensity differences between tissues at time the MRI receives the signal (the red line). This salient contrast between tissues is a core MRI property leading to the clear images characteristic for MRI. The magnetization represents the signal emitted from the tissues. Acronyms are: white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). The red line depicts the time point to receive the MRI echo from the tissue. (A) depicts the tissue-specific net magnetization in the z-direction that yields the T1 images. Analogously, (B) depicts the net magnetization rotating in the transversal (xy) plane that yields the T2 images. M0 is the net magnetization right after slice excitation.
14
Functional MRI
The T2-related MRI sequence, called T2*, receives stronger signals from tissues that are
highly water abundant. However, this sequence is also very susceptible to magnetic
distortion. Using the T2* sequence to image neural tissues, one receives blood-oxygen-
level dependent (BOLD) images, which are core to fMRI. Since oxygenated blood is
diamagnetic and deoxygenated blood is paramagnetic, it affects the BOLD signal
conveniently to increase the contrast of the BOLD signal. A well-perceived hypothesis of
how the BOLD relates to neural activation has been formulated by Buxton et al. (1998).
Basically, the BOLD is a function (Equation 1) of the blood volume and concentration of
deoxyhemoglobin content. Independently, deoxyhemoglobin (Equation 2) and blood
volume content (Equation 3) are differential functions of the metabolic rate of the neural
tissue. The link between metabolic rate and neural activity hypothesized by Buxton et al.
has been shown to be accurate empirically (Glover, 1999).
Equation 1: !!!" = !![!! ! − ! + !!(! − (!/!)) + !!(! − !)]
, where
!! is the resting blood volume,
!!,!!,!! are constants based on an oxygen-extraction fraction,
! = !/!! (normalized total deoxyhemoglobin), and
! = !/!! (normalized total blood volume)
Additional variables are used: t is for time, fin is for the stimulus paradigm (often tightly
correlated with the metabolic rate of oxygen), E(t) is for approximate transport conditions
(this variable is actually time dependent and may be written as ! !(!) using Equation 4),
E0 is the resting net extraction of oxygen by the capillary bed, fout(v) depends on venous
15
pressure, and τ0 is a scale constant. Equation 4 is an approximation of a wide range of
oxygen-transport conditions.
Equation 2: !"!"= !
!!!!" !
! !!!
− !!"# !! !! !
Equation 3: !"!"= !
!!!!" ! − !!"# !
Equation 4: ! ! = ! − (! − !!)!!
Neurons are thought to induce the hemodynamic response by controlling vasodilation and
vasoconstriction of the blood vessels in the neural tissue in accordance with Figure 3.
Neuroimage Preprocessing
The raw images have many artifacts that are corrected for by preprocessing, which
includes routines such as slice time correction, motion correction, smoothing, and
normalization.
• Slice time correction: a single functional MRI is often acquired during 2 s and
contains several slices that have been acquired during this time. However, the first
and last slices in the raw image are almost 2 s apart in time, which is often
corrected for.
• Motion correction corrects the image for subject movements during imaging.
16
• Normalization is done since everybody’s uniquely shaped brain needs to be in
the same space for group analysis. Often this is begun by choosing a brain
template, which could be the Talairach-Tournoux system. The Talairach template
was originally a 3D grid system for neurosurgery (Talairach and Szikla, 1967).
Then each subject’s neural image is geometrically transformed to match the
template (Cox, 1996). Today there are many algorithms that transform the subject
space into the Talairach template. This normalizes every subject’s neural tissue
into the same space and makes it possible to run statistics across subjects.
• Smoothing is also used to increase agreement across subjects. Even after perfect
normalization, a specific action that activates a specific neural region may be
located slightly differently across subjects. For instance, if two subjects wiggle
their thumbs, their neural responses may differ by a few millimeters even after
motion correction and normalization. If a couple of millimeters of smoothing are
applied, the regions of responses will overlap from both subjects to satisfy
statistical tests. Smoothing is a 3D filter operation, often using a Gaussian
blurring across the entire image, where the Gaussian kernel has a specific full
width at half the maximum in millimeters (Cox, 1996).
The above steps are basic preprocessing steps that are often used in today’s fMRI
software packages, such as AFNI (Cox, 1996) or FSL (Smith et al., 2004).
17
Figure 3. Hemodynamic response. Since neurons lack intrinsic energy storage, an immediate supply of oxygenated blood is essential after neural activation. This is accomplished by neural signaling to dilate blood capillaries in the close vicinity (B), which generates a supply of freshly oxygenated blood to its location. However, the fresh blood with an increased concentration of oxygenated hemoglobin over deoxyhemoglobin is delayed in accordance to the hemodynamic response function (HRF) depicted in (C). When neurons are inactive, capillaries are constricted (A), and usually most oxygen has been consumed, leaving high deoxyhemoglobin concentration. Both blood volume and oxygenated hemoglobin concentration increase the intensity of the blood-oxygen-level dependence (BOLD) independently. Situation (A) includes two factors (less blood volume and lower oxygenated hemoglobin concentration) to decrease the image intensity that often occurs in brain regions that have low neuronal activity. Situation (B) depicts an increase of both independent factors that often occur in regions with high neuronal activity. This is hypothesized to explain the strong relationship between BOLD signal and neuronal activity.
18
Statistical Analysis of fMRI
The general linear model (GLM) has been very popular in assessing the correlation
between regional neural responses and a behavior paradigm. Briefly, GLM is used to
depict the correlation between the blood-oxygen level dependence (BOLD) responses in
each voxel time series, with a time series of a set of regressors. Each time-dependent
regressor is often the time series of something behavioral the subject performs; e.g., if the
regressor is the subject’s eye state, the regressor could be set to 1 at time points when the
subject has his or her eyes open and set to 0 when subject has his or her eyes closed.
Since the neural response is delayed, the regressors are also convolved with the
hemodynamic response function (HRF; Glover, 1999), or similar approximation, before
the correlation is made (Figure 3C).
For each voxel, Equation 5 optimizes bi for each index i, to produce the smallest sum of
errors (ei) using the least squares fit.
Equation 5:
!! = !! + !!!!,! + ⋯!! = !! + !!!!,! + ⋯
+!!!!,! + !!+!!!!,! + !!
⋮ ⋮!! = !! + !!!!,! + ⋯
⋮+!!!!,! + !!
, where !! is the BOLD intensity at each image-acquisition time point t; !! are the beta
values, i = [1,p]; !!,! is the i:th predictor; and et is the residual error at time point t.
Still, the hypothetical test has not been calculated using GLM. The hypothesis of the
study is implemented using Equation 6, using a contrast vector representing the model
19
hypothesis; e.g., if one hypothesizes that regressors 1 and 2 (!!,! and !!,!) affect the brain
analogously, the null hypothesis would be that the first and second beta values are the
same (!!: !! = !!), which is the same as !!: !! − !! = 0. After the beta values are
rewritten so the null hypothesis sum to zero, the coefficients of the beta values represent
the contrast vector, which in this case would be [1, -1]. Then we may calculate a t-value
for each voxel by using the contrast vector in accordance with any arbitrary hypothesis
(e.g. [1, -1]), according to Equation 6. After the t-value is calculated for each voxel of the
neural image, the contrast image is the result.
Equation 6: ! = !!!!"# ! !!(!!!)!!!
, where c is the contrast vector
20
Chapter 2: Neuroimaging Meta-Analysis of Mild Traumatic Brain Injury
Preface
This chapter was published in Neuroimage Clinical in January 2014. A few alterations
were made to fit the thesis. The title was “Neuroimaging after mild traumatic brain
injury: Review and meta-analysis,” and the research article was authored by Cyrus
Eierud, Cameron Craddock, Sean Fletcher, Manek Aulakh, Brooks King-Casas, Damon
Kuehl, and Stephen LaConte.
Dr. Stephen LaConte (advisor) started an mTBI journal club with an aim to write a meta-
analysis, the membership of which club included the author, Dr. Cameron Craddock (a
post-doc in our lab), Sean Fletcher (a medical student in our lab), and Manek Aulakh
(another medical student in our lab). Dr. LaConte added well-acquainted mTBI experts
Dr. Damon Kuehl, M.D., and Dr. Brooks King-Casas to help us write the publication.
Early on, Dr. Craddock informed us about tools we could use to produce a meta-analysis,
including Ginger ALE and LimeSurvey. We made a PubMed search encompassing the
majority of neuroimaging-mTBI literature, which resulted in 298 publications that we
split among ourselves to extract core information to the LimeSurvey database. Dr. Kuehl
and Dr. King-Casas occasionally checked our work and contributed their expertise to the
publication. All figures in this chapter were crafted by the author and Stephen LaConte,
with various input from all other team members.
21
This chapter reviews the literature on neuroimaging of mTBI.
Introduction
Presented in this chapter are the results from a broad review and meta-analysis of mTBI
across the spectrum of neuroimaging modalities. Studies have examined mTBI at both
acute and chronic stages of injury. In reviewing the literature, it is important to note that
the mean time between neural imaging and the mTBI incidents of a study can affect its
participant-exclusion criterion, leading to prospective and symptomatic mTBI groups
(Dikmen et al., 1992). In prospective mTBI studies, the exclusion criteria are independent
of mTBI (e.g., specific age ranges or drug dependences). Symptomatic mTBI studies
recruit chronic participants. Based on estimated recovery rates, this corresponds to
effectively excluding the majority of those who sustain mTBIs. In other words, studies of
symptomatic groups enroll participants because they have lingering complaints caused
(presumably) by their head injuries, whereas prospective studies recruit based on mTBI
records at the time of concussion (before any chronic mTBI is known).
This paper broadly reviews mTBI neuroimaging studies of structure and function to
highlight the tremendous effort that has been made to investigate the spectrum of acute to
chronic time scales. We additionally provide meta-analyses to examine the current utility
of MRI for studying both structure and function. In terms of structure, some reports have
claimed that MRI is more sensitive to detect complicated mTBIs than CT (Mittl et al.,
1994). Similar to other authors, we use complicated mTBIs to include the broad range of
abnormalities that lead to non-negative imaging results (Arciniegas et al., 2005; Iverson,
22
2005; Williams et al., 1990). It should be pointed out that definitions of “mild” have
varied widely among both clinicians and researchers. Thus while many studies have
excluded participants with imaging findings, this is not universally the case. Among the
other neuroimaging methods, MRI is also unique in that it can be used to study both
structure and function. Many physical parameters provide MRI with a wide range of
contrast mechanisms, enabling “traditional” T1- and T2-weighted structural scans, neural
correlates of brain function using fMRI, white-matter microstructure by diffusion MRI,
and biochemistry through MRS.
The meta-analyses presented here are focused on both structure and function and are MRI
specific. The first meta-analysis is motivated by the heterogeneity of fMRI findings and
focuses on the question of anatomical consistency for fMRI. Similarly, the second
analysis examines the issue of white matter vulnerability to mTBI. Looking at
anatomically localized findings, previous neuroimaging data have suggested that anterior
regions of the brain are more vulnerable to abnormalities (Hashimoto and Abo, 2009;
Lipton et al., 2009; McAllister et al., 1999; Niogi et al., 2008a). This agrees with previous
literature on moderate and severe TBI that has found increased vulnerability in ventral,
frontal, and temporal regions (Bigler, 2007). In addition, it is thought that these regions
have both fragile brain structures and hard skull bones close by. For mTBI, however,
published reports have been highly heterogeneous in their findings of regional white
matter changes. Therefore, we examined whether anatomical consistency in mTBI lesions
exists in the literature. Our third meta-analysis examines the apparent inconsistency in
diffusion-based anisotropy findings across studies, which has led to debates about
23
whether or not anisotropy values increase, decrease, or even change at all after mTBI
(Lange et al., 2012) as well as whether anisotropy levels positively or negatively correlate
with performance levels in neuropsychological assessments (FitzGerald & Crosson,
2011). Recent reports have suggested that it is important to consider the time post-injury
in diffusion-weighted imaging (Mayer et al., 2011; Niogi & Mukherjee, 2010). For
example, Niogi & Mukherjee (2010) suggest that anisotropy is increased in the acute
phase and decreased in the chronic phase in symptomatic TBI patients. Similarly, Mayer
et al. (2011) note that anisotropy values can be either reduced or increased in semi-acute
time points, but they tend to be decreased in later, chronic stages of symptomatic mTBI.
Based on these considerations, we tested the hypotheses that anisotropy is increased in
the acute phase and decreased in the chronic phase. Specifically, we performed a meta-
analysis that considered the time post-injury of each study’s mTBI cohort. Our results
support the hypothesis that acute mTBI is associated with elevated anisotropy values and
that chronic mTBI complaints are correlated with depressed anisotropy.
Materials and Methods
We queried PubMed (http://www.ncbi.nlm.nih.gov/pubmed) for articles on imaging-
based mTBI studies. We extracted key information from these articles and entered items
into a LimeSurvey database (http://www.limesurvey.org) that we developed for this
study. LimeSurvey is an open-source web-based survey application that features an
unlimited number of participants in a survey (in our case, each paper constituted a survey
participant) and flexible export functions to multiple text and software formats. The
24
LimeSurvey web interface enabled the authors to collaborate from different locations.
Data from the articles were manually entered into LimeSurvey, and these data were
examined and summarized using custom-built Python parsing scripts.
We examined publications spanning 21 years, from 1990 to 2011, using a query focused
on mTBI and neuroimaging:
(mTBI or “mild traumatic brain injury” or “post concussive syndrome” or “post
concussion syndrome” or “postconcussion syndrome” or “postconcussive syndrome”)
and (neuroimaging or “magnetic resonance imaging” or MRI or “positron emission
tomography” or PET or magnetoencephalography or MEG or electroencephalography or
EEG or “functional magnetic resonance imaging” or fMRI or “diffusion tensor imaging”
or DTI or T2 or “diffusion spectrum imaging” or DSI or “diffusion weighted imaging” or
DWI or SWI or “susceptibility weighted imaging” or “T2*” or “CT” or “computed
tomography” or FLAIR or “diffusion kurtosis imaging” or “diffusional kurtosis imaging”
or DKI or “single photon emission computed tomography” or SPECT or NIRS or “near-
infrared spectroscopy” or fNIRS or “functional near-infrared spectroscopy” or “resting
state” or “functional connectivity” or “default mode network”) and (“1990” [Publication
Date]: “2011” [Publication Date]).
Our search date was November 11, 2011. This query resulted in 298 publications, from
which we excluded 176, using seven exclusion criteria. Specifically, we excluded 85
review articles, 47 nonimaging articles (most of which mentioned CT in relation to
subjects’ TBI status, but without any association with a statistical variable), 13 case
studies, 12 animal studies, 9 articles that mixed mTBI with moderate/severe TBI or
25
posttraumatic stress disorder (PTSD), 7 articles that were not in English, and 3 others (a
video article, a fatigue article, and a mold exposure article).
We extracted the following into the LimeSurvey database: 1) Time after concussion (the
median/mean time between injury and neuroimaging). If the studies were longitudinal or
had multiple groups of subjects, multiple time points were reported. 2) Range of time
after concussion (the minimum and maximum time between injury and neuroimaging). In
publications where only standard deviations were provided, we approximated the range
as the median ± 1.6 standard deviations. Note that this is analogous to the 90%
confidence interval. 3) Number of subjects (sum of mTBI and control groups). Note that
some publications (such as longitudinal studies) included multiple separate time points.
4) Imaging modality (CT, 12 subtypes of MRI including DTI and fMRI, PET, SPECT,
MEG, EEG, and near-infrared spectrometry). 5) Hypothesized effect of mTBI motivating
the imaging study, which we termed the “theme.” For example, several studies
hypothesized that mTBI would lead to white matter abnormalities, and thus to
characterize this they measured anisotropy. In contrast, studies that hypothesized that
connectivity would be affected could address this research “theme” by using not only
diverse measures including fMRI but also diffusion imaging-based tractography. 6) Other
imaging results (anatomical coordinates, ROIs with significant anisotropy, and other
parameters).
26
Statistical Analysis
The fMRI meta-analysis used Ginger ALE (Eickhoff et al., 2009, Eickhoff et al., 2012;
Turkeltaub et al., 2012). Ginger ALE is an activation likelihood estimator based on
results from multiple publications. All fMRI coordinates were converted into Talairach
space. We used the default Ginger ALE parameters (Eickhoff et al., 2009) and added the
number of subjects per experiment. We chose a false discovery rate threshold level of
0.05. We used all available fMRI publications of mTBI, most of which used working
memory tasks, but the tasks also included resting state fMRI, an auditory oddball task,
and a spatial navigation task (Chen et al., 2007; Krivitzky et al., 2011; Mayer et al., 2009,
Mayer et al., 2011; McAllister et al., 2001; McAllister et al., 2011; Slobounov et al.,
2010; Witt et al., 2010). Only coordinates from contrast images using a two-sample test
(mTBI and control) were included. AFNI (Cox, 1996) was used to present fMRI meta-
analysis images in Talairach space.
For the DTI meta-analyses, each reported region of interest (ROI) was coded using the
ICBM-81 atlas (Mori et al., 2008). ROIs that were outside of this atlas were only used in
the analyses involving time post-injury and omitted for the spatial DTI analysis. Most
publications reported fractional anisotropy (FA) values; however, we also included
studies that reported relative anisotropy (RA). Publications found anisotropy differences
in two predominant ways: either categorically, by examining anisotropy means between
mTBI and control groups, or parametrically, through regression analyses correlating
anisotropy values with neuropsychological performance. AFNI (Cox, 1996) was used to
present DTI meta-analysis results in Montreal Neurological Institute (MNI-152) space.
27
As in other publications that account for diverse sets of assessment scores (Bazarian et
al., 2007), we normalized neuropsychological results using algebraic negation
(multiplying the original score with a negative one) such that better performance would
correspond to positive scores. For example, we did not transform results from the
California verbal learning test, in which scores should be higher for normal performance
than for impaired performance, but we did transform (negate) completion time for Trail
Making A test (in this case, better raw scores are lower, corresponding to faster
completion times). Our motivation for doing this was that it would enable us to
unambiguously discuss negative and positive correlations between neuropsychological
results with changes in imaging measures.
Results
The number of mTBI publications has dramatically increased over the past two decades.
A shows the number of publications with respect to time for the neuroimaging articles we
analyzed in the context of mTBI research as a whole. The more general mTBI PubMed
search in Figure 4A excluded our neuroimaging keywords (to include both imaging and
non-imaging studies). Specifically, we used (mTBI or “mild traumatic brain injury” or
“post concussive syndrome” or “post concussion syndrome” or “postconcussion
syndrome” or “postconcussive syndrome”) and (“1990”[Publication Date]:
“2011”[Publication Date]). The figure shows that the rate of mTBI publications has
increased over time, with a notable upsurge occurring after 2007. Figure 4B classifies
28
Figure 4: Overview of mTBI and neuroimaging mTBI publications between 1990 and 2011: The levels of mTBI studies have been increasing over the past two decades, and the focus on neuroimaging has increased proportionately with the field as a whole. (A) The histogram shows 1,314 mTBI articles published from 1990 to 2011. (B) focuses on the 122 imaging-based studies and provides a graphical breakdown of what imaging modalities were used. To date, the predominant imaging modalities have been CT and MRI.
29
imaging modalities of the publications we analyzed over time. Most of the imaging
modalities are MRI based. The number of studies relying (at least in part) on CT is
relatively large, primarily because it is used ubiquitously (and almost exclusively) in
emergency room settings. Thus, CT was used as part of the clinical characterization of
the subjects in many studies. The functional modalities (EEG/MEG, fMRI, SPECT/PET,
and ASL) have been steadily used during the time period shown. Figure 5 and Figure 6
show that structural and functional neuroimaging studies have examined mTBI across
acute to chronic time scales with a wide range of imaging modalities. It is important to
note that the number of subjects in these publications represents the total number of
mTBI and control participants for each experiment, which could include different
participant cohorts. Also note that many experiments have large variability for subjects’
post-injury data collection times. In addition, while the number of subjects includes
control participants, for obvious reasons these controls did not contribute to the post-
injury collection times in Figure 5 or Figure 6. Focusing on some observations from each
figure, Figure 5 shows that CT studies have been performed on both small and large
cohorts, but the majority of these studies occur soon after injury, usually using the
subjects’ clinical CTs. DTI and the other structural modalities have a much better
coverage of the time post-injury. For Figure 6, one observation is that many EEG studies
also occur soon after injury but in contrast to CT, EEG has also been broadly applied to
the entire time course post-injury and to a large number of subjects. Figure 6 also shows
that task-based fMRI has been similarly broadly applied, while resting state fMRI studies
have predominantly
30
Figure 5: Time course of structural neuroimaging studies Structural neuroimaging studies of mTBI vary widely in terms of both the number of subjects studied and the time range after injury. Shown is a depiction of the number of subjects (mTBIs as well as any controls in each experiment) and the time post-injury (for the mTBIs) of data collection for structural imaging studies. The colored backgrounds indicate the time axis scales (days, weeks, and years). The imaging modality is indicated by color, and each line indicates the study’s post-injury scan range (earliest and latest reported times post-injury). The line’s ellipse represents the median time after injury. To keep all data “visible,” overlapping lines have been shifted up by two subjects.
31
Figure 6: Time course of functional neuroimaging studies Functional neuroimaging studies of mTBI also vary widely in terms of both the number of subjects studied and the time range after injury. Conventions are the same as in Figure 4.
32
focused on acute and semi-acute time ranges. PET, ASL, CE-MRI, and MEG seem to be
rarely used. In addition, there is a lack of PET studies of acute mTBI, and both PET and
MRS publications used groups that were smaller than 40 subjects.
The research areas of neuroimaging studies tend to follow distinct structural and
functional categories, as shown in Figure 7. As mentioned previously, CT was used as
part of the clinical characterization of the subjects in many studies. In papers where CT
was a research focus, it was used to examine structural abnormalities arising from mTBI.
Although the category of “mild” TBI sometimes includes the stipulation of negative
findings, CTs can nonetheless be used to detect abnormalities from a wide range of
trauma severities that appear as hyper- or hypo-intensities observed in gray matter or
white matter regions as well as hematomas, cerebral swelling, and hemorrhages. Gray
matter is composed largely of neuronal cell bodies, glial cells, and the capillaries. White
matter abnormalities are thought to reflect stretched or sheared axon bundles or abnormal
cellular microstructures. Abnormalities not clearly falling into the categories of
hematomas or white/gray matter findings have been designated as “complicated mTBI”
in Figure 7. Complicated mTBI was examined 56 times in our publication sample, mainly
using CT and MRI. Gray matter (GM) abnormalities were only reported in one T1 and
one CT study. Abnormal blood–brain barrier (BBB) permeabilities were reported by
three publications using SPECT and CT (CT studies also detected tau/S100b proteins in
blood). White matter (WM) abnormalities were examined 38 times (primarily with DTI).
Hematomas were only reported in two CT-based studies. Metabolites were analyzed by a
PET, a SPECT, and five MRS publications. Perfusion was examined by one PET, one
33
Figure 7: Themes studied with different imaging modalities Two major goals of neuroimaging studies are to find structural and functional markers of mTBI and to establish links between neuropsychological assessments and neuroimaging. Shown is a breakdown of imaging methods and study focus “themes.” Themes include gray matter (GM) abnormalities; white matter (WM) abnormalities; intracranial hematomas; complicated mTBIs (non-negative imaging results, excluding hematomas and abnormalities localized to GM and WM such as increased blood–brain barrier permeability, contusions, intracranial lesions, and micro bleeds); metabolites (changes in magnetic resonance spectroscopic results); blood–brain barrier (BBB) permeability; perfusion deficits; task-based imaging; connectivity analysis; and neuropsychological assessments (used in conjunction with a neuroimaging modality).
34
ASL, and 12 SPECT publications. Task-based imaging was used in 17 fMRI and 15 EEG
studies, often linking tasks to brain activity. Connectivity analyses were examined 27
times, detecting low-frequency activity, connectivity differences between neural regions,
EEG pathologies, altered resting state networks (using fMRI), and connectivity maps
(using white matter tractography). Neuropsychological assessments were used in 72
publications and were the one theme of investigation (despite not relying on any
particular imaging modality itself) that spanned all structural and functional modalities.
Figures 8–10 display results from our MRI meta-analyses. Figure 8 shows results from a
Ginger ALE analysis of the fMRI publications summarized in Table 1. Table 1 also
examines how many independent publications support each activation likelihood
estimated (ALE) peak. The analysis produced six regions that were consistently more
active in mTBI compared to control groups and seven regions with lower activity for
mTBI. Spatially, these regions suggest an anterior-to-posterior pattern in which activity is
reduced in anterior regions and increased in posterior regions. Of the seven regions with
decreased mTBI activity, six were in the frontal lobe or anterior cingulate, with the
remaining one being relatively posterior (temporal lobe/BA 39). The regions with
increased mTBI activity consisted of two coordinates in the cerebellum, two insula
regions, and two foci in the parietal lobe (BA 40). The mean Talairach anterior-to-
posterior coordinate for regions with reduced activity was Y = 15 mm, compared to Y =
−23 for increased mTBI activity, and a two-sample t-test of Y between the decreased and
increased regions was significant (p = 0.05, two-tailed).
35
Close examination of the number and anatomical locations of publications reporting
white matter abnormalities revealed significant anatomical heterogeneity, as shown in
Figure 9. See Table 2 for white matter regions in the ICBM-81 atlas as well as their
abbreviations and center-of-mass coordinates. It is important to note that several ICBM-
81 regions have not been reported by the publications we examined. These include the
pontine crossing tract, medial lemniscus, inferior cerebellar peduncle, cerebral peduncle,
posterior thalamic radiation, cingulum to hippocampus, superior fronto-occipital
fasciculus, tapetum, and inferior longitudinal fasciculi. These unreported regions were
coded in two ways; both leading to statistically significant correlations between
anatomical location along the anterior-to-posterior gradient and the number of
abnormalities. Specifically, we calculated a Spearman correlation, 1) omitting these
regions from the statistical test (using only regions that had reported abnormal findings in
the literature), resulting in the correlation of ρ2 = 0.26 (p < 0.026); and 2) including the
unreported regions (coding their frequency as zero), resulting in ρ2 = 0.32 (p < 0.0025).
We did not find significant correlations in the inferior-to-superior or left-to-right
directions. Laterality of the regions was also not significant.
36
Figure 8. Activation likelihood estimate (ALE) analysis of fMRI mTBI publications Activation likelihood estimate (ALE) analysis of fMRI mTBI publications shows both increased and decreased BOLD responses for mTBI. As shown, with mTBI have increased response in the cerebellum, insula, and inferior parietal regions (BA 40) compared to controls. Relative to mTBI, control subjects have increased response in several regions in the frontal lobe and in BA 39. Maps have a threshold of p < 0.05 using a false-discovery rate (FDR) correction and a minimal cluster size of 64 μL. Results are displayed on a Talairach brain template.
37
ROI abbrev.
Vol. (μL)
ALE cluster coordinate (X, Y, Z) a
ROI name Supporting publications (ALE coord. within 20 mm)b
mTBI > Control
CBT 304 33 -53 -39 r. cerebellar tonsil Wi2010, Sl2010, Kr2011 BA40 152 58 -32 39 r. inf. par. lobule / BA 40 Wi2010 CBC 144 -33 -34 -26 l. culmen Wi2010, Mc2011 IFG 144 29 9 -12 r. inf. front. gyrus Wi2010, Ma2011 Insula 64 33 17 7 r. insula Wi2010, Ma2011 SMG 64 57 -47 33 r. supramarg. gyrus Wi2010, Ma2011 Control > mTBI
MdFG 1296 32 28 24 r. mid. frontal gyrus Wi2010, Ma2011, Mc2011, Ch2007
ACC 856 3 30 10 r. ant. cingulate Wi2010, Ma2011 MTG 616 51 -59 23 r. mid. temp. gyrus / BA
39 Wi2010, Ma2011, Mc2011
BA9 496 45 17 34 r. precentral gyrus / BA 9 Wi2010, Mc2011, Ch2007 MdFG 304 -50 14 35 l. mid. frontal gyrus Mc2011 BA46 184 46 40 10 r. DLPFC / BA 46 Ma2011, Mc2001, Mc2011 BA46 184 -39 36 19 l. DLPFC / BA 46 Ma2011, Mc2011, Ch2007 Table 1. mTBI vs. control cluster centroids with FDR (p < 0.05) using ALE a The ALE cluster coordinate is the center of mass of the cluster calculated by the ALE algorithm. b Publications codes Wi2010, Sl2010, Kr2011, Mc2001, Mc2011, Ma2011, and Ch2007 represent Witt et al., 2010; Slobounov et al., 2010; Krivitzki et al., 2011; McAllister et al., 2001; McAllister et al., 2011; Mayer et al., 2011; and Chen et al., 2007, respectively. Supporting publications have coordinates at least within 20 mm of the ALE coordinate.
38
Figure 9. Anisotropy abnormalities mapped spatially in mTBI Shown are the ICBM-81 white matter regions, colored to indicate the number of publications reporting white matter abnormalities (regions with no abnormal findings in the literature are not shown). The Montreal Neurological Institute (MNI-152) template is added for anatomical reference. Using the center of mass for each ICBM-81 structure, we determined that a significant anterior-to-posterior relationship exists between frequency in the literature and anatomical location. See Table 2 for full names of the anatomical labels. Note that since more lateral structures are only partially visible, the anatomical labels point to a convenient, visible location and do not necessarily reflect a structure’s center of mass. For example, SLF is mostly covered by more medial structures and is only visible at its most posterior-inferior part. Coordinates are displayed in MNI-152 space.
39
ROI C o o r d i n a t e s Abnorm- ROI Acronym X(L)a Y(P)b Z(I)c alities Name ACR 22 28 10 6 anterior corona radiata alIC 18 8 8 2 anterior limb of internal capsule bCC 1 -5 27 1 body of corpus callosum bFX 1 -6 13 1 body of fornix cFX 28 -24 -6 1 fornix crus CgC 8 -11 30 4 cingulate cortex CgH 21 -31 -13 0 hippocampal part of cingulum CP 12 -18 -12 0 cerebral peduncle CST 6 -25 -33 1 corticospinal tract EC 30 0 1 2 external capsule gCC 0 26 7 6 genu of corpus callosum ICP 8 -44 -37 0 inferior cerebellar peduncle MCP 1 -40 -35 1 middle cerebellar peduncle ML 6 -37 -33 0 medial lemniscus PCR 24 -38 28 2 posterior corona radiata PCT 1 -29 -31 0 pontine crossing tract plIC 20 -13 8 5 posterior limb of internal capsule PTR 34 -55 7 0 posterior thalamic radiation rIC 31 -29 6 1 retrolenticular part of internal capsule sCC 1 -42 18 5 splenium of corpus callosum SCP 6 -42 -25 1 superior cerebellar peduncle SCR 24 -8 30 3 superior corona radiata SFO 20 2 21 0 superior fronto-occipital fasciculus SLF 36 -26 26 2 superior longitudinal fasciculi SS, IFO, ILF
40 -32 -9 2 sagittal stratum, inf. fronto-occipital fasc., inf. longitudinal fasc.
TAP 28 -46 14 0 tapetum UNC 34 0 -16 4 uncinate fasciculi Table 2. Center of mass of white matter structures in the ICBM-81 atlas a X-coordinate assumes symmetry between left and right hemispheres (mean absolute value of centroid reported here). b Y-coordinate increases with anterior direction. c Z-coordinate increases with superiority. All coordinates are in MNI-152 space. Reported anisotropy-abnormalities indicate the number of publications that reported an anisotropy abnormality in ROI.
40
Based on observations by Niogi & Mukherjee (2010) as well as Mayer et al. (2011), we
wanted to examine the hypothesis that anisotropy would be elevated at acute time points
and depressed afterwards. Figure 10 demonstrates the critical importance of accounting
for the time post-injury in diffusion-weighted imaging reports. Shown are the time-
resolved anisotropy results for each study. In Figure 10A, studies compared mTBI groups
to control groups. In Figure 10B, studies correlated mTBI anisotropy values with
neuropsychological performance. Note that some publications report experimental results
at two different chronological points (Grossman et al., 2012; Inglese et al., 2005; Mac
Donald et al., 2011; Mayer et al., 2011; Messe et al., 2011) and thus are represented by
two bars. Another complication with the literature is that some research groups seem to
have used overlapping mTBI cohorts across publications (see potential overlap in boxed
bars in Figure 10). Thus counting each bar in Figure 10 as an independent measure likely
leads to inflated estimates of significance. To account for this, we report the most
conservative statistical estimate by using each boxed group as a single experiment. As
shown, elevated anisotropy values are more frequently reported for studies of acute mTBI
(before 2 weeks), while depressed anisotropy findings are reported more frequently for
post-acute mTBI (after 2 weeks). In Figure 10A, a two-sample t-test confirms that
anisotropy is greater in the acute phase (one-tailed, 14 DOF, p = 0.02). In Figure 10B,
acute-phase anisotropy was also significantly and negatively correlated with
neuropsychological performance compared to a predominant positive correlation in the
post-acute phase (one-tailed, 7 DOF, p < 0.006). Finally, to put these results into the
context of the literature, it is important to note that the participant populations were
41
recruited using different criteria in the acute and chronic periods. Specifically, all of the
acute studies recruited prospective mTBI subjects, the majority of whom would be
expected to recover from their injuries. Conversely, many of the post-acute mTBI studies
of anisotropy selected for symptomatic subjects.
42
Figure 10. Anisotropy compared temporally in relation with mTBI or performance (A) Elevated anisotropy following mTBI is more frequently reported for studies of acute mTBI, while depressed anisotropy is reported more frequently for studies after the acute phase. Each bar represents the ratio of increased (red) to decreased (blue) regions in mTBI vs. a control group for each publication. The gray boxes represent experiments with insignificant FA difference between mTBI and control groups. The lines connecting the bars to the time axis mark the study’s median time post-injury. Also indicated is the number of subjects per experiment and whether the mTBI subjects were prospective (P) or selected (S). Boxed bars indicate potentially overlapping subject cohorts. The colored backgrounds indicate the time axis scales (days, weeks, and years). For statistical tests, we have defined post-injury times less than 14 days as an acute mTBI and post-acute times as 2 weeks and greater. Based on a one-sided, two-sample t-test, the acute anisotropy was significantly greater than the anisotropy after the acute phase (p = 0.02). (B) Shows the analogous information for studies that reported significant relationships between anisotropy measures and neuropsychological performance. Using a t-test analogous to that in 7A, we show that the acute phase anisotropy was significantly anti-correlated with neuropsychological performance compared to a predominant positive correlation after the acute phase (p < 0.006). Statistical significance was designated as * for p < 0.05. Publications are I) Bazarian et al. (2007); II) Chu et al. (2010); III) Wilde et al. (2008); IV) Wu et al. (2010); V) Yallampalli et al. (2010); VI) Inglese et al. (2005); VII) Miles et al. (2008); VIII) Lipton et al. (2009); IX) Mayer et al. (2011); X) Mayer et al. (2010); XI) Messe et al. (2011); XII) Mac Donald et al. (2011); XIII) Holli et al. (2010); XIV) Smits et al. (2011); XV) Grossman et al. (2012); XVI) Lipton et al. (2008); XVII) Maruta et al. (2010); XVIII) Niogi et al. (2008a); XIX) Niogi et al. (2008b); XX) Geary et al. (2010); XXI) Lo et al. (2009).
43
Discussion
An obvious key limitation of this analysis (and of any review that attempts to report a
snapshot of a vibrant field) is that components of it are immediately outdated. Although
we believe that the conclusions we draw here will be qualitatively accurate for the next
few years, the literature we sampled just barely supported any meaningful fMRI meta-
analyses, and the issue of the anisotropy values’ dependence on time post-injury is far
from conclusive. Part of what we have hoped to accomplish here is an assessment of the
modalities, questions, and times post-injury that recent publications have examined. We
hope that research teams can use this assessment to strategically bolster the areas of this
field that would most benefit from additional effort.
Our review of the mTBI literature highlights the fact that currently about one-tenth of
published studies have a human neuroimaging component and that several small and
large studies spanning acute to chronic time points have been conducted. These studies
have examined both structural and functional changes with mTBI, using virtually every
available medical imaging modality. Two key commonalities have been used across the
majority of imaging studies. The first is the comparison between mTBI and control
populations. The second is the attempt to link imaging results with neuropsychological
assessments. Indeed, in addition to highlighting the fact that research topics have largely
(but not exclusively) fallen into either structural or functional categories and providing a
sense of the proportion of work that has been devoted to the two, Figure 7 strikingly
shows that neuropsychology appears to be the predominant commonality across the
imaging literature. This underscores the fact that mTBI neuroimaging efforts have
44
prioritized attempting to link brain measurements with neuropsychological assessments.
Despite the significant efforts to date, neuroimaging methods still lack the individual
patient-level sensitivity and specificity to serve as a diagnostic tool for mTBI. Similar to
other reviews in this area (Belanger et al., 2007; Hunter et al., 2012; Niogi & Mukherjee,
2010; Prabhu, 2011; Pulsipher et al., 2011), we believe that ultimate motivations for
further neuroimaging work are to provide data to predict an individual’s recovery, to
measure her or his transient and persistent cognitive deficits at a level that compliments
or exceeds the sensitivity of neuropsychiatric testing, and to quantify the success of
cognitive and drug-based interventions (McAllister et al., 2011). In the meantime, the
field needs further effort to provide stronger correlative understanding of the
relationships between neuroimaging and neurology, and both types of measures also need
improvement.
Our fMRI and DTI meta-analyses show, however, that there is hope for consistent
neuroimaging markers of structure and function. The fMRI meta-analysis results in
Figure 8 and Table 1 show a frontal vulnerability in mTBI, demonstrated by decreased
signals compared to controls. Studies on fMRI consistently report decreased activity in
mTBI in frontal regions such as the right middle frontal gyrus (MFG), the anterior
cingulate cortex (ACC), and the right precentral gyrus. Our finding of bilateral decreases
in the dorsolateral prefrontal cortex (DLPFC) is consistent with three published studies
(Chen et al., 2007; Mayer et al., 2011; McAllister et al., 2011). The DLPFC may be
involved in working memory. Specific localization of working memory intense regions
was found in nonhuman research (Funahashi et al., 1989) and in human research using
45
fMRI (D’Esposito et al., 1995). Even though DLPFC responds to working memory, it is a
very complex region that serves a multitude of cognitive functions (Kane and Engle,
2002). Interestingly, among the most significant mTBI increases, right cerebellar tonsil
and left culmen were noted in four independent publications (Chen et al., 2007; Mayer et
al., 2011; McAllister et al., 2011; Witt et al., 2010). Krivitzky et al. (2011) noted that the
cerebellum has been implicated for regulating behavior, working memory, and other
aspects of executive control. In addition, it has long been recognized that cerebellar
lesions can lead to postural deficits (Horak and Diener, 1994), which is a prominent issue
in mTBI (Guskiewicz et al., 1996; McCrea et al., 2003). Finally, it is important to note
that the well-known fMRI issue of multiple-comparisons correction may be a limitation
of our Ginger ALE analysis. Most of the publications that we used corrected their
findings for multiple comparisons. However, notable exceptions include McAllister et al.
(2001) and Witt et al. (2010). Another limitation of the fMRI meta-analysis was that,
unlike the diffusion-weighted literature, the number of fMRI studies published did not
have power enough for a statistical analysis using time post-injury. We feel that “time-
resolving” fMRI studies could be an important area for future investigation.
Structurally, diffusion-weighted imaging is currently one of the most promising
techniques for characterizing the subtle and heterogeneous changes that occur with
mTBI. Figure 9 demonstrates that differences in the frequency of reported mTBI white
matter anisotropy values have an anterior-to-posterior gradient. Taken independently,
each DTI study shows white matter differences that seem to be sporadic. When
integrating across all reported findings, these frequencies of abnormalities have a
46
systematic spatial distribution, corroborating previous suggestions that anterior regions of
the brain are more vulnerable to abnormalities (Hashimoto and Abo, 2009; Lipton et al.,
2009; McAllister et al., 1999; Niogi et al., 2008a). That TBI is particularly vulnerable in
frontal and ventral regions is well known (Bigler, 2007). Anterior regions may also be
related to executive/cognitive function disabilities in mTBI (McCrea et al., 2003).
Using the existing literature to time-resolve results, Figure 10 suggests that white matter
anisotropy can serve as an important potential marker of both chronic symptoms as well
as of an acute response to secondary injuries. Taken together, the literature demonstrates
that elevated anisotropy values (and negative correlations with neuropsychological
performance) are more frequently reported for studies of acute mTBI, while depressed
anisotropy findings (and positive correlations with neuropsychological performance) are
reported more frequently for chronic studies. Our analyses show that elevated anisotropy
values are more frequently reported in studies of acute mTBI, while depressed anisotropy
findings are reported more frequently in post-acute studies, corroborating conjectures of
Mayer et al. (2011) and Niogi & Mukherjee (2010). Thus, Figure 10 helps to clarify an
apparent inconsistency in the literature about anisotropy changes after mTBI by “time-
resolving” the results. In addition, Figure 10B raises important new questions about the
interactions of anisotropy, neuropsychological performance, and time post-injury. In
particular, the fact that cognitive performance is negatively correlated with anisotropy
during acute stages but positively correlated with anisotropy chronically is puzzling.
At least two issues remain outstanding with the results in Figure 10. The first is the
inconsistency in Figure 10A of the results for paper XXI (Lo et al., 2009). In discussing
47
their results, Lo et al. (2009) were careful to point out that the small sample size of their
study called for more future work. However, they also suggested potential explanations
for their findings, including the possibility that the observed increases in FA could reflect
recovery from injury; that specific degradation mechanisms could target subsets of fibers,
leading to enhanced FA; or, finally, that the increases in FA are compensatory for other
observed decreases in other brain regions. The second issue relates to the heterogeneity of
the neuropsychological testing in the papers used to generate Figure 10B. Thus a
limitation in the literature (and in our results) is the lack of consistency of
neuropsychological assessment. As previously mentioned, we have “normalized” the
scores such that better performance corresponds to a positive score, regardless of which
test was used. We note, however, in light of the literature’s “uncontrolled assessment
variance,” that our results are conservative as well as robust across cognitive domains and
testing variability. Undoubtedly, more sensitive, standardized assessments that are
specifically adapted to mTBI are needed to further reduce the limitations of these types of
studies.
The results of Figure 10 and the outstanding issues that we have raised suggest that more
studies focused on mTBI and anisotropy are warranted and necessary. At this point we
can only conjecture on possible explanations for our meta-analysis results. One
possibility for the increase in FA, in people with acute mTBI, is that MR diffusion-based
measurements are not purely sensitive to white matter microstructure but are also inflated
by acute-stage secondary injury or compensatory mechanisms (Mayer et al., 2011; Niogi
& Mukherjee, 2010). The idea behind this is that inflammation and secondary injury
48
factors (including ischemia, cerebral hypoxia, and cerebral edema) may increase
anisotropy in the acute phase, but that these factors do not contribute in the chronic
phase. Chronically, though, residual damage of white matter could lead to a decreased
anisotropy signal. A second possible explanation of Figure 10B, however, is that there is
not a simple causal relationship between anisotropy and cognitive performance.
Whatever the mechanism, Figure 10B indicates that poor neuropsychological
performance is associated with high anisotropy scores immediately after injury and with
low anisotropy in the chronic phase. From a statistical point of view, Figure 10A
demonstrates a main effect between anisotropy and time post-injury, but Figure 10B
additionally indicates an interaction with cognitive performance. We have previously
mentioned that there is controversy surrounding anisotropy changes after TBI.
Nonetheless, FitzGerald & Crosson (2011) noted in their review of the literature a general
acceptance that decreases in FA values are expected after TBI-induced axonal injury.
However, they also pointed out that mechanisms for increases in FA are also possible
(e.g., unequal injury in one fiber in voxels where two or more fibers cross). Furthermore,
there have been examples of other disorders (such as autism) for which widespread
differences have been noted in the relationship of FA with cognitive function (Ellmore et
al., 2013). Better understanding of the underlying mechanisms of anisotropy changes
following TBI will likely need to rely on methods and validation studies in animal
models such as those of Budde et al. (2011) and Mac Donald et al. (2011) to relate
pathology to diffusion tensor imaging findings. In addition, a study comparing age and
mTBI effects with DTI (Ilvesmäki et al., 2014) has shown that age effects may be
49
stronger than effects from mTBI. Important future work will be to evaluate if we may
find any age bias in our collected dataset. Finally, the majority of the literature has
focused on FA, but some publications have reported radial diffusion (RD), axial diffusion
(AD), and mean diffusion (MD) diffusion as well. The four measures are related to each
other; usually FA and AD are significant in the same direction, while MD and RD tend to
be significant in the opposite direction.
To summarize, this review and meta-analysis have demonstrated several important points
about the ongoing use of neuroimaging to understand the functional and structural
changes that occur throughout the time course of mTBI recovery. A broad range of
neuroimaging modalities are being applied to this problem across the range of post-injury
time scales, and an important component of many studies is the effort to link imaging
measurements to neuropsychological assessments of cognitive deficits. Upon closer
scrutiny of two MRI-based modalities, first, the importance of interpreting white-matter
measures of anisotropy in the context of time post-injury was demonstrated. Secondly,
consistent findings were available for evaluating anatomical vulnerability to mTBI for
both structure and function. The fact that we have been able to produce statistically
significant results in this area is encouraging. Based on the complexity of the injury,
however, much more work in this area is required. Future publications focused in depth
for each image modality would greatly complement this generic analysis and the field of
mTBI. Here we have been able to preliminarily examine important factors related to
anatomical vulnerability and time post-injury, but these issues need to be refined with
future data, and many other factors remain to be studied (e.g., injury mechanism
50
differences, number and timing of previous concussions, and psychological comorbidities
including major depression and post-traumatic stress; Iverson and Lange, 2003;
Rosenbaum & Lipton, 2012). Ultimately a major driving motivation for continued study
is the further efficacy of neuroimaging for clinical diagnosis at both acute and chronic
stages and the synergistic improvement with neuropsychological assessments and
rehabilitation strategies.
51
Chapter 3: Detecting Motor Rate using Support Vector Machines
Introduction
According to the mTBI-fMRI ALE meta-analysis by Eierud et al. (2014), fMRI studies
may detect mTBI (Figure 8). Even though the results are significant using false discovery
rate (FDR) correction, the mTBI effects are significant only at the group level, and
detection needs to be sensitive at a single-subject level in order to be of use clinically.
Almost all studies included in the meta-analysis were hypothesis driven using the general
linear model (GLM; Friston et al., 1995). In contrast to the hypothesis driven GLM,
SVM, using supervised learning (Cortes & Vapnik, 1995) to make predictions may
complement GLM. In fMRI, the ability of SVM to find neural patterns that can be used
for powerful predictions has been successfully proven (Cox & Savoy, 2003; Fan et al.,
2007; LaConte et al., 2005; LaConte et al., 2007; LaConte et al., 2011; Mitchell et al.,
2004). Furthermore, SVM’s prediction accuracy may increase its sensitivity to track
mTBI at subject level, which may be needed for clinical usage. Bringing neuroimaging
closer to clinical practice by using SVM is the end goal for this chapter. It would be most
beneficial if neuroimaging right after a concussion could predict any chronic mTBI
effects. This could reduce the lengthy and problematic diagnosis of PCS, which is a very
52
costly disability that could benefit from image objectivity and less chance of
misdiagnosis (Evans, 1992).
Even if it does not fully achieve that final goal, this chapter will take a few steps in that
direction. The first step taken in this thesis is to characterize support vector classification
and the newer support vector regression algorithm.
For fMRI and mTBI, a behavioral paradigm is needed that may be either a default mode
network (DMN) or task fMRI paradigm. DMN is the neural signal from a person who is
not performing any behavioral paradigm at all and seems to be at rest (Raichle & Snyder,
2007). The mTBI-fMRI literature is more abundant for fMRI tasks than for DMN in
accordance to the previous literature (Eierud et al., 2014) that was important in our case.
Behavioral Paradigm
In order to characterize SVM properly, the following were required from the behavioral
paradigm:
(1) The maximal neural response signal in order to receive predictions as robust as
possible.
(2) Practicality and accuracy to handle large amounts of labels accurately, preferably by
automation, to save working hours.
(3) A task with abundant previous literature to allow referencing results.
(4) A behavioral paradigm with a wide stimulus spectrum that generates a span of low
response to strong neural response, to maximize the signal from the paradigm.
53
These requirements were fulfilled by a unimanual motor rate task. Motor tasks are known
to yield a strong response. For example, it was seen in Glover (1999) that a motor
response can yield a stronger signal than an auditory stimulus. The motor rate design we
used contained more than 14,000 labels that were automatically and accurately logged.
The abundance of data samples benefitted the SVM train/test. Automation of the labeling
process saved many work hours and accuracy. Many publications in the previous
literature have used similar stimuli for reference. A representative sample of motor rate
publications was listed in the comprehensive meta-analysis of Witt et al. (2008). It was
possible to widen the motor stimuli spectrum, spanning from a very slow to extremely
fast rates that induced neural responses accordingly.
Unimanual Motor Rate Literature
Many unimanual research articles exploring rate effects have found a positive rate effect
between motor rate and neural response in the following regions: contralateral (primary)
sensorimotor cortex (M1), ipsilateral sensorimotor cortex (SMC), supplementary motor
area (SMA), contralateral thalamus, contralateral lentiform nucleus, and cerebellum. The
meta-analysis of Witt et al. (2008) listed 11 publications that studied tapping rate tasks.
However, this chapter is MRI specific, so we eliminated two PET studies (Blinkenberg et
al., 1996; Kawashima et al., 1999) and one TMS study (Rounis et al., 2005). In addition,
two studies were not rate specific in any way comparable to our study (Aramaki et al.,
2006; Lehericy et al., 2006). The remaining six publications matched our study (Jäncke et
al., 1998; *Jäncke et al., 1999; Jäncke et al., 2000; Riecker et al., 2006; *Deiber et al.,
54
1999; Lutz et al., 2004), with two studies that were ROI specific and did not overlap with
all regions our study explored (noted with an asterisk).
Support Vector Machines
An SVM is a supervised learning model. This means that subject data need to be divided
into a training part and a test part. First the training part is fed to the SVM together with
labels describing each data sample, and this generates the SVM brain model. This model
has two features. First, the brain model’s weights may be used to present what regions are
most important for predictions. Secondly, the model enables a second phase of prediction
of the remaining dataset (Figure 11). The model is constructed differently for support
vector classification (SVC) and support vector regression (SVR), which are described in
greater detail below (and see Figure 11). In general, SVC may be used to separate
between different mental states, whereas SVR may be used to decode the degree of a
single mind state. For example, SVC could classify mTBI subjects into two classes, such
as “likely to recover” and “likely transition into PCS,” while mTBI SVR could predict a
probability that the subject will transition into PCS by a percentage, where 0% would
mean recovery and 100% would mean maximal probability to transition into PCS.
Support Vector Classification Background
First the SVC algorithm was trained using at least 4 minutes of images, using stimulus
labels for each image, to generate a specific normal vector (W), also called a weight
55
vector or brain model. Then, in the second phase, the SVC could predict unlabeled data
samples (fMRI volumes) using the brain model.
For explanatory purposes, the most rudimentary SVC algorithm (called a “hard margin”)
is described that only classifies two fully separable classes in the following six steps:
1. Each data sample (brain volume) needs to be transformed into a feature vector (x) in
the feature space. Because our data samples contained at least 30,000 voxels, the feature
space would contain at least 30,000 dimensions, which is impossible to depict on 2D
paper. Therefore, a scatterplot of a hypothetical brain with only two voxels was depicted
to demonstrate the idea (Figure 12).
56
Figure 11. Supervised learning Both the support vector classification (SVC) and the support vector regression (SVR) algorithms are supervised learning algorithms, which means that they have to be trained using labeled data before they can predict unlabeled data. Step 1 (training), shows the data samples (fMRI volumes) being acquired every other second. Simultaneously, the subject’s behavior and stimuli is labeled and tagged to each fMRI volume. The SVM algorithm uses several fMRI volumes to develop an optimized brain model. In Step 2 (testing) the brain model may be used to classify unlabeled fMRI volumes. Figure is adapted from LaConte et al. (2007) with permission from Dr. LaConte and John Wiley & Sons.
57
Figure 12. Hypothetical classification and feature space scatterplots (A) Depicts a hypothetical two-class paradigm. The class y = +1 (star labeled) could be a motor stimulus, and the y = -1 class (diamond labeled) could be a resting state. (B) A feature space scatter plot of a hypothetical two-voxel brain with voxels X1 and X2, within constraints for a hard-margin SVC. (C) is analogous to (B) with the difference that feature vectors now are allowed on the wrong side of the margin with the penalty of the variable ξt.
58
The first step of describing SVC is to explain the rudimentary hard-margin SVC, which
only works in cases where the data are fully separable. The SVC decision function D(x)
is the dot product of the weight vector and the feature vector in accordance with Equation
7.
Equation 7: D ! = ! ⋅ ! +w!
w is the weight vector and x is a feature vector of the data sample. w!/ ! is the
distance between the origin and the hyperplane.
In the hard-margin case, all feature vectors have to be above the margin (Equation 8) for
Task 1 labels and below the margin (Equation 9) for Task 2 labels. For example, the
decision function is usually 1 or -1 (for the two classes). However, in the case a data
sample causes the decision function to result in 0 (D ! = 0) that data sample is placed
right on the hyperplane that is not allowed in the hard margin case.
Equation 8: Case y! = +1 (task 1) ⇒ ! ⋅ ! +w! / ! ≥ +Δ
Equation 9: Case y! = −1 (task 2) ⇒ ! ⋅ ! +w! / ! ≤ −Δ
The Δ is the width of the margin. To normalize the margin with the weight vector such
that ! Δ = 1 both Equation 8 and Equation 9, it may be rewritten as
Equation 10: y! ! ⋅ ! +w! ≥ 1
59
!(!) =!!! !! +
12‖!‖!
!
!!!
The weight vector that is constrained by Equation 10 and minimizes ! ! = ! !, for
t = 1,… ,T separates the two classes with the widest possible margin (optimal weight
vector) and is called the hyperplane. Since η ! = ! ! is minimized at an identical w
to Equation 11, which is easier to calculate, Equation 11 is most often used in the
literature for the SVC minimization.
Equation 11: η ! = !!! !
There are different ways to minimize ! ! in Equation 11, but in literature it is common
to use Lagrangian multipliers.
Since empirical data clusters from different classes often are close enough to overlap, the
hard margin SVM is not very practical. For this reason, the soft-margin SVM was
introduced (Cortes & Vapnik, 1995), making SVMs more usable and popular. The soft-
margin SVM is extended with a term to accept margin errors on top of the equations for
the hard-margin SVM. Equation 10 and Equation 11 are extended into Equation 12 and
Equation 13, respectively, as follows:
Equation 12: y! ! ⋅ ! + !! ≥ 1− !!
Equation 13:
60
Feature vectors overstepping the margin boarders, and the degree to which they do so, are
registered by a slack variable in accordance with Equation 12 (Figure 12C). Then the
slack variables (noise) are summed up for penalty (Equation 13). In the soft-margin case,
Equation 12 accepts overlapping clusters, but instead the width of the margin may be
varied depending on C, which is a constant that modifies the hyperplane margin width. If
C is high, the margin is narrow, and if C is low, the margin becomes wider. This way the
SVM can run for any data, and the C value can be adjusted to optimize the SVM results
for the dataset. Further details on SVC can be found in textbooks such as Cherkassky and
Mulier (2007).
Since an SVC hyperplane only separates two classes, several classes may be predicted by
combining SVC with another algorithm such as the directed acyclic graph (DAG)
algorithm. Combining DAG with SVC a hyperplane (brain volume) requires estimating
each two-class combination, totaling n(n-1)/2 brain volumes for n classes. For example, if
4 classes need to be classified, 6 brain models need to be estimated. An example of a
DAG multiclass classification using several hyperplanes (brain models) is depicted in
Figure 13. A decision tree for a 4-class DAG classification is depicted in Figure 14.
61
Figure 13: Hyperplanes layout for DAG multi-classification (A) Depicts four clustered classes of data separated by borders crafted by six 2-class hyperplanes and direct acyclic graph (DAG). (B) The six 2-class hyperplanes were estimated for the identical clusters as in (A).
62
Figure 14: Direct acyclic graph (DAG) decision tree The unlabeled brain image (Test Input) is tested against six 2-class brain models before final prediction of which of the four classes it belongs to. The six brain models (SVC1 to SVC6) have previously been trained in all possible unique 2-class combinations possible.
63
!(!) =!!!(ξ! + ξ!∗) +
12‖!‖!
!
!!!
Support Vector Regression Background
The linear support vector regression (SVR) may become a great complement to SVC for
detecting mTBI. There are algorithms for nonlinear SVR, which are out of scope of this
research. In the test case, SVR decodes a neural image into a continuous value
representing the level of a mind state or condition. The condition may be a mental
disability such as a pain disorder, but in our case we want to compare how well SVR
decodes fMRI. SVR predicts continuous real numbers instead of integers (representing
classes), which SVC predicts. Very differently from SVC (which uses a hyperplane), the
SVR encapsulates an epsilon tube (ε-tube), preferably outside the feature vectors. The
regression is then decoded by the length of the ε-tube at which the normal of the feature
vector falls. Feature vectors outside the ε-tube are allowed, but they are penalized using a
slack variable, and feature vectors on the edge of the ε-tube are support vectors. A
hypothetical 2D feature space and ε-tube in the SVR case is depicted in Figure 15B. The
loss function for the slack variables is depicted in Figure 15C. The SVR minimizing
function is as
Equation 14 and is constrained by Equation 15 as follows:
Equation 14:
Equation 15: y! − w ⋅ x! −w! ≤ ℇ+ ξ!y! − w ⋅ x! −w! ≤ ℇ+ ξ!∗ ξ!, ξ!∗ ≥ 0,where t = 1,… ,T
Further SVR details may be found in textbooks such as Cherkassky and Mulier (2007).
64
Figure 15: Hypothetical support vector regression scatter plot The SVC uses a hyperplane to separate two classes. Conversely, the SVR models an epsilon tube that ideally should have the feature vectors clustered along its centerline. Any feature vector outside the epsilon tube is a support vector and may even have a slack variable. Since regression models a single variable intensity (perhaps of a mind state), all feature vectors originate from the same stimulus type (B). (A) shows the SVC from Figure 12 for comparison. (C) shows the loss for the feature vector from the center of the epsilon tube. Within the tube there is no loss, but the farther away the feature vector is from the margin, the higher loss it receives.
65
Examples of Differences Between GLM and SVM
GLM is hypothesis driven. Often the result is a single image after using data across the
entire data collection. Conversely, SVM is a supervised learning algorithm. In the test
phase, the SVM generates a prediction from each individual data point. Because each
subject often has several test images (in SVM), there may be a large number of
predictions at every time point in the test phase. Using a group of predictions, inferences
may be drawn in a complementary way to GLM. Also, an SVM prediction, which is
represented by a single number, may be simpler to use in automated systems than GLM.
In the case of GLM, one may need to interpret the GLM contrast image before it can be
used for automation. An automated system could be a real-time feedback system.
In the real-time feedback, a prediction may be made after each image is acquired (such as
a 2-s delay). If the labels are “left” and “right,” the subject could then control a cursor
going left or right on a screen with only thought (LaConte et al., 2007). Real-time
feedback, also called brain computer interface (BCI), is an exciting avenue for future
research. Our lab has developed several BCI models that give real-time feedback to
subjects during imaging (e.g., LaConte et al., 2007; Papageorgiou et al., 2013).
Materials and Methods
All subjects were approved by Baylor College of Medicine’s institutional review board
(IRB).
66
The following protocol was used for the motor rate experiment (N = 15):
• Neural imaging protocols were acquired using a Siemens 3T Trio.
o Structural T1 used MPRAGE with 1mm isotropic voxels.
o fMRIs were acquired using a 3T Trio scanner with the following echo
planar imaging (EPI) parameters:
TE/TR = 30/2000 ms
Flip angle (FA) = 90°
Firld of view (FOV) 220×220×132 mm3
Matrix (X, Y, Z) = 64×64×30
• The subject’s task was to press a button with the right index finger in sync with a
metronome while fMRIs were acquired. Further explanation of the task is
provided below.
• All image processing used the AFNI (Cox, 1996):
o Deobliqued images. Images were oblique (10 to 45 degrees) in order to
maximize cerebellar tissue in field of view.
o Motion-corrected images
o Aligned images to scullstripped T1 image
o Detrended all voxels temporally
o Smoothed all images using Gaussian blur with 6mm full width half
maximum (FWHM)
o Processed images using 3dsvm
• For weight maps, a 3dttest was used at the second level.
67
Unless otherwise stated, all the analyses used a whole-brain SVM. This means that no
part of the brain was masked.
Behavioral Task
Basically, the subjects tapped a button in sync with an audible metronome. The rate of
the metronome was 1, 2, 3, 4, or 5 hertz, which enabled five classes of stimuli. The
metronome tick played at the same rate during each 30-s active block. It was always
followed by a block of rest, which was 10 s. During the active block, the rate was kept
constant, but each block was pseudorandomly set to different rates according to the five
classes. Not very important to the task, we added or subtracted 0.2 Hz for each block to
eliminate any possible habituation effects from any subject. In addition to the auditory
stimuli, the fixation cross was flashing at every metronome tick. During rest, subjects
only gazed at a stable fixation cross and were not allowed to tap the button, as there were
no metronome ticks. Each subject performed 3 runs, which included 16 active blocks.
The 3 runs were almost 33 min long. For analysis, only the last 26 s of each block
yielding 13 TRs (brain volumes) were used for training and testing, unless otherwise
stated. This would eliminate any transitional effects in the beginning of the block. Since
there are 9 to 10 blocks of each rate, 117 to 130 samples per rate exist per subject. Note
that our subjects’ responses were inaccurate at the highest motor rates before we used the
real-time feedback bar that helped subjects to tap more accurately. The feedback bar
indicated the degree of deviation from the perfect tapping rate. The task is depicted in
Figure 16.
68
Figure 16. Stimuli paradigm for real-time motor rate effect The stimulus task consisted of a subject pressing a button in sync with the metronome clicks from the headphones. In addition, the subject was instructed to fixate on the fixation cross that was flashing in sync with the metronome clicks. Since many subjects found it hard to tap the button at the fast pace of 5 Hz, we constructed the user interface above with a real-time feedback bar. If the real-time feedback bar was placed by the grid lines at an inaccurate rate to the left of the fixation cross, the subject was instructed to tap faster. Conversely, the subject was instructed to tap slower if the dynamic feedback bar was shown by the grid lines to be at an inaccurate rate on the right side. Increased behavior accuracy resulted in increased SVM prediction accuracy.
69
Simulation Model
For SVM simulations MATLAB and AFNI (Cox, 1996) with 3dsvm (LaConte et al.,
2005; Joachims, 1999) was used in following 3 steps.
(1) MATLAB generated random 2D images including a signal to be predicted and the
image label. The labels were numbers 1 to 5 to match the rates we measured from
our subject-based data. The signal that MATLAB provided for each image was
corrupted with Gaussian noise to the degree that the signal overlapped across
labels.
(2) AFNI used the built in 3dsvm plugin and half of the MATLAB generated images
with their labels to train an SVM-brain-model.
(3) AFNI used the built in 3dsvm plugin, the SVM-brain-model, and the remaining
images generated from MATLAB to predict the labels.
(4) Even though SVM was not provided the true labels (generated from MATLAB in
Step 1) for the images it predicted those “true” labels existed to validate the SVM
predictions with. Sorting the predicted labels by the true labels
Results
To increase SVM prediction accuracy, the subject had to perform the task accurately.
Since it was important to have a wide spectrum of rates to predict from, the rate extended
up to 5 Hz. With the support from the real-time feedback bar, all rates were very accurate
with the slight exception of a negligible rate discrepancy at 5 Hz (Figure 17). Statistics of
all subjects revealed that SVR prediction is correlated with the true rate (p < 2*10-16),
70
which bolsters that the motor rate predictions were associated with the true motor rates
(Figure 18).
As seen in Figure 19, our motor rate effects were depicted using a GLM contrast map, an
SVC weight, and an SVR weight map. The maps depicted significant extreme points at
analogous regions across the three methods. Note that it was easy to separate individual
ROIs using the standard threshold of p < 0.05 (FDR corr.) for the SVM weight maps.
However, the GLM method was so sensitive that many regions became inseparable at
that threshold. Thus the GLM threshold was changed to p < 0.002 (FDR corr.), which
kept the GLM ROIs separated. It was possible to see that all regions with positive effect
between BOLD response and motor rate in both SVR and SVC were inclusive in the
neural pattern for GLM contrast image. For GLM, we noted an additional rate effect for
SMA (not found using SVM). SMA, which was not as consistently found in the previous
literature (as M1), may be because the SMA rate effect was less significant than for M1.
M1 was consistently found in previous literature and in our data to have stronger
significance. The SVM weight maps were not as sensitive as the GLM maps, which may
explain why the SMA was missing, but the M1 was present in the SVM
71
Figure 17. Group behavior motor rate mean with standard deviation Block trial behavioral motor rate mean from the 15 subjects included in the motor rate experiment and their deviations. Ideally the mean rate (line with darker color) should constantly be the rate at which it is suggested to press the button (color coded according to legend), with minimal deviation (lighter color), throughout the trial. It can be seen that subjects got slightly fatigued with time at 5 Hz, but that was negligible.
72
Figure 18. Support vector regression prediction variance Support vector regression (SVR) predicts the motor rate with the significance of p < 2*10-16 according to linear fit. The X-axis represents the true motor rate. The Y-axis represents the predicted rate. SVR performed 624 predictions for each subject (N=15).
73
Figure 19: Similarities between GLM, SVC, and SVR statistical brain patterns It is noticeable that the weight maps for both SVC and SVR agree with those of GLM, which is the gold standard. As seen in the top row, the GLM contrast image is more salient than both the support vector weight maps. The red color represents regions where the response is higher at faster motor rates. Conversely, blue represents regions with higher response for slower motor rates. Apart from a few blue regions (in E, F, G, and H), it is noticeable that all the regions with higher responses at higher motor rates in both the SVR and SVC cases are inclusive in the GLM pattern. The variable q means a false discover rate (FRD) corrected p-value. The anatomical locations are mostly labeled in the middle row because that contrast map had fewer problems with overlapping ROIs than the top row and at the same time was more significant than the bottom row. All maps are group results from 15 subjects engaged in the motor rate task. The GLM and SVR used all five motor rates, while the SVC had sorted away all data for 2, 3, and 4 Hz for (D) and (F) and 1, 3, and 4 Hz for (E). Anatomical acronyms are: supplemental area (SMA); inferior temporal gyrus (ITG); superior temporal gyrus (STG); sensorimotor cortex (SMC); primary motor cortex (M1); and middle occipital gyrus (MOG).
74
weight maps. Figure 19 depicts our results from the three algorithms. The strongest
weighted region for both SVM and SVR was the M1 rate effect in the positive direction.
Spatially, the SVM results match the GLM results seen in Figure 19 at slice y = -20mm.
The direction and strong intensity were corroborated by all publications that analyzed M1
(Jäncke et al., 1998; Jäncke et al., 2000; Riecker et al., 2006; Lutz et al., 2004). This
suggests that M1 is rate dependent. SMA may also be rate dependent, but further analysis
is needed to clarify the SMA dependence on behavioral rate changes since only Riecker
et al., (2006) found a positive rate effect, using Witt et al.’s (2008) list. Three
publications did not find any significant rate effect in the SMA when the dominant hand
was used (Jäncke et al., 1998; Jäncke et al., 2000; Lutz et al., 2004). The listed
publications were only used as well-representative samples of all unimanual motor rate
studies. Other publications have found both positive motor rate effects with SMA (Rao et
al., 1996; Wurster et al., 2014) and alternate (nonpositive) motor rate effects with SMA
(Sadato et al., 1996) by using analysis of variance (ANOVA) to find the difference. In
addition, previous literature found a positive unimanual rate effect in the ipsilateral dorsal
SMC (Jäncke et al., 1998; Lutz et al., 2004; Rao et al., 1996; Riecker et al., 2006;
Wurster et al., 2014), ipsilateral cerebellum (Figure 19 at Z = -13; Jäncke et al., 1999;
Lutz et al., 2004; Riecker et al., 2006), and contralateral thalamus (Riecker et al., 2006;
Wurster et al., 2014). Interestingly, our positive rate effect in the lentiform nucleus
(putamen), found only in the case of GLM and SVC, was in the negative direction
opposite to the previous literature (Riecker et al., 2006; Wurster et al., 2014). This
discrepancy was not predicted, but may be an effect from our visual interface that did not
75
exist in the other two studies. Further, our GLM analysis found a positive rate effect in
the cingulate motor area in accordance with the previous literature (Deiber et al., 1999;
Lutz et al., 2004), which may be related with projections from the spinal cord (He et al.,
1995). We also found small negative rate effects in the ventral SMC bilaterally in the
SVC case and in the right ventral SMC in the SVR case. Ventral SMC was not
corroborated by previous literature, to our knowledge. Except from the discrepancy of
direction of rate effect in the lentiform nucleus several regions matched well with
previous literature. Further analysis about the directional discrepancy may be needed to
fully elucidate that it is caused by the visual component including the feedback feature of
our task.
Because our task also included stimulus rates from visual and auditory senses, we
received additional responses exclusive to motor-related stimuli. The auditory aspect of
our task may have contributed a positive rate effect, in the cases of SVM and GLM, in
the thalamus, and in the bilateral superior temporal gyri (STG). In the SVM case, a small
negative rate effect was found in the right anterior insula. These three regions agreed with
the previous literature about the auditory rate effect (Ackermann et al., 2001).
Rate effects from the flashing visual stimuli were found in the SVC and GLM cases,
bilaterally in middle occipital gyri (MOG; located in Brodmann area 19, which is part of
visual cortex) and in the right inferior temporal gyrus (ITG). ITG may be related to visual
processes such as face recognition (Gross, 1992). A more elaborate table of regions that
regressed with or against the motor rate stimuli were listed in Table 3.
76
Neural ROIs 1vs 2Hz
1vs 3Hz
1vs 4Hz
1vs 5Hz
2vs 3Hz
2vs 4Hz
2vs 5Hz
3vs 4Hz
3vs 5Hz
4vs 5Hz
SVR
1) L Sensorimotor (dors) 2154 4186 7477 10159 2479 5445 6583 2403 2) L STG 894 1747 5933 7843 1625 3982 4186 1890 3) R STG 1260 4186 4389 1544 1829 935 1593 4) R Culmen (cerebell.) 3779 3617 813 2276 1138 297 5) R Sensorimotor (vent) 2235 1829 975 894 486 6) R Sensorimotor (dors) 1544 1138 7) L Sensorimotor (vent) 1057 8) R MOG (Visual) 1057 1788 9) R Inf. Temp Gyr (ITG) 1219 10) L MOG 1097 11) L Lentiform Nucleus 1016 12) L Thalamus 1016 13) R Insula 975 243 Table 3. Cluster volumes in μ-liters using support vector machine Cluster volumes in μ-liter (μL) using SVC for different rates and SVR. All results were thresholded at p < 0.05 using FDR correction. Columns 2 to 11 are results of SVCs between different pairs of motor rates. The last column shows the regression results.
77
Using the direct acyclic graph algorithm, SVC prediction accuracy was 44%, which was
significantly greater than chance (20% given the five prediction classes). SVC prediction
accuracies were also tabulated by subject in Table 4. In addition, SVR, which is
continuous, could be advantageous for certain measurements such as to predict the level
of pain (DeCharms et al., 2005). Our study used SVR to predict the motor rate of the
subject at the instant the neural image was acquired. SVR prediction from neuroimaging
of motor rate for the best subject CEA013 was depicted over time in Figure 20. The
predicted rate clearly correlated with the true rate, as shown in Figure 21. In Table 5, it
can be seen that all subjects had significant correlation and prediction accuracies. Overall
subjects correlations, and root-mean-square errors (RMSEs) along with their standard
deviations were: ρ = 0.79, σρ = 0.10 and RMSE = 0.84, σRMSE = 0.17.
78
Table 4. Support vector classification accuracies from 1 to 5 Hz per subject Support vector classification prediction accuracy by subject. Predictions in the middle of the rate range (3 Hz) are so poor that they are often below chance level. This is because the predicted rate still has a wide distribution even after our optimizations. When the true rate is at the extremes, the SVC prediction is supported by out-of-range corrections. For example, if the true 5-Hz rate was predicted as 7 Hz, it would be corrected to the closest valid rate, which is 5 Hz and becomes accurate. Since the middle rate gets the least out-of-range correction, it gets the most errors. That the rates are well predictable is shown when looking at results from the combined rates. Future work may correct for the poor middle rate predictions, which may improve the overall SVC performance.
79
Figure 20: Support vector regression of single subject The red line shows the true motor rate performed by the subject. The black line shows the motor rate the SVR predicted the subject would have only after analyzing the neural images. Rest blocks between active blocks have been removed to save graph space.
80
Figure 21. Support vector regression errors per subject graphed Root mean square error (RMSE) of support vector regression (SVR) per subject for the motor rate task.
81
CEA0 13 30 31 32 34 35 37 38 39 40 41 42 43 44 04 ρ 0.91 0.73 0.76 0.78 0.86 0.67 0.65 0.84 0.60 0.90 0.78 0.85 0.83 0.85 0.78
RMSE 0.55 0.97 0.93 0.85 0.70 1.10 1.07 0.76 1.11 0.61 0.87 0.76 0.78 0.75 0.86 Table 5. Support vector regression correlation and root mean square error (RMSE) between predicted and true rate
82
The extremes (1 and 5 Hz) SVC rates had fewer errors than the center classes, as depicted
in Figure 22A. These results were found across all 15 subjects. SVR could greatly
complement SVC by providing a continuous prediction. Conversely to SVC, SVR
yielded larger errors at the extremes (Figure 22B). That the optimal performance for SVC
(at extreme rates) and SVR (at middle rate) were opposite was puzzling so we performed
different simulations to recreate these results. We noted that if we simulated that the
source BOLD data was widely distributed (noisy; Figure 23A) we received an analogous
situation with SVC (Figure 23B) and SVR (Figure 23C) as in our empirical case. To
induce this characteristic effect in both SVR and SVC predictions simultaneously we had
to simulate our signal with such a noise that the standard deviation was so large that it
overlapped across labels. After trying many other simulations that got alternate results we
decided to explore our subject data further. Using the region with most significant effect,
namely the primary motor cortex (M1), we found such a widely distributed noise, which
was systematic and time dependent. Since the motor rate paradigm is a block paradigm
over 30 seconds the BOLD signal should be fully stable after 10 s (Figure 24A)
according to a leading contemporary hypothesis of the hemodynamic response function
(Buxton et al., 1998; Glover, 1999). However, our empirical data from M1 included a
strong time dependent perturbation. To make the exploration as accurate as possible we
used all the user data possible by averaging 9-10 blocks for all 15 subjects per rate. This
resulted in that each of the five mean motor rates were an average of 135 to 150 blocks
depicted in Figure 24B. It is clear from Figure 24B that the signal from M1 has strongly
time dependent overlaps (e.g. at 5 Hz rate at 12 s there is an overlap with the 3 Hz rate at
83
30s) that should not overlap in accordance with contemporary theory (Figure 24A). This
may indicate that the Balloon model (Buxton et al., 1998) may be corrected in our case to
gain higher accuracies. Analogous time dependent perturbations for block paradigms
have been found in previous literature (Fox et al., 2005). If this time dependent
perturbation can be corrected for the predictions may improve, which may be important
to resolve in order to detect mTBI.
84
Figure 22: Comparing support vector classification to regression prediction errors Support vector classification (SVC) has more errors in the middle of the spectrum, while support vector regression (SVR) has more errors at the edges of the distribution. The chance level for percent errors is 80% (for SVC). The chance level of root mean square error (RMSE) is 1.37 in our model.
85
Figure 23: Simulated SVM signal and results
(A) Simulated source signal intensities (inside images) with wide noise profile that intensity wise overlaps across different labels. (B) shows the resulting support vector classification prediction of the simulated images with higher accuracy at extremes of true labels. (C) shows the support vector regression (SVR) had higher prediction accuracy of the simulated images in the middle range of the true labels. The diagonal purple line marks the ideal predictions (and intersects the true and predicted 1-labels as well as the true and predicted 5-labels). As seen for the SVR the true label 5 images are predicted too low, while the true label 3 images are predicted with an accurate mean. The true label 1 images are predicted too high, but may have a smaller mean error than the label 5 images.
86
Figure 24: Theoretical and empirical response time series for signal variations within rates
The theoretical BOLD response in accordance with contemporary literature (Buxton et al., 1998; Glover, 1999) compared to the empirical response may be divergent for block paradigms such as our motor rate task. Both diagrams show the BOLD time series for different rates in accordance to the color label in the legend. (A) shows the theoretical BOLD as a convolution between the hemodynamic response function (Glover, 1999) and our 30 s box-car time course (marked block stimulus) of when task was performed. (B) shows the empirical BOLD from the primary motor cortex (M1) by averaging all 15 subjects and trials per rate. Each motor rate time series was averaged by 135 to 150 blocks (of same motor rate).
87
Discussion
We found that both the wide spectrum of subject motor responses spanning from 1 to 5
Hz and correct behavior were essential for the SVM to be able to predict anything at all.
When we tested experiments with comfortable motor rates (2 to 4 Hz) for subjects, their
behavior was great, but SVM could not predict the results. The interface for our first
experiment of 1 to 5 Hz lacked real-time feedback, so the subjects’ behavioral accuracies
were poor at 4Hz and above, which led to prediction problems. Finally, we rebuilt our
interface to look like Figure 16 and included a real-time feedback bar that gave
quantitative feedback to subjects about their rate discrepancy. This real-time feedback
improved the behavior (Figure 25) and may have improved the SVM predictions as well.
The final results demonstrated that SVR and SVC clearly decodes the neuroimaging data
and that the SVM weight maps detect neural regions that are associated with the
paradigm the subject performs.
We also found that the BOLD response may deviate from the contemporary hypothesis
about the neural response. The contemporary HRF hypothesis (Buxton et al., 1998;
Glover, 1999) may corrupt our SVM predictions and may be corrected for. Our
significant findings in fMRI and SVM research may have brought the fields closer to
mTBI. However, further research is needed to complete the bridge between SVM and
mTBI.
88
Figure 25: Behavioral improvement after real-time feedback Each trace in the quadrants is mean and deviation, from 9 or 10 trials, of the subject’s true motor rate. Ideally the mean (trace in darker color) should match the rate color, with minimal deviation (lighter color), throughout the trial. The left column shows behavior for two subjects before the dynamic real-time feedback bar was implemented into the user interface, and the right column shows the results for the same two subjects after the implementation. Qualitatively, it can be seen that the results are remarkably more stable and accurate after the implementation of the real-time feedback bar.
89
Chapter 4: mTBI Model System Using Executive Function
Introduction
As mentioned before, fMRI may detect mTBI (Figure 8), but these results may be
impeded by heterogeneity (Rosenbaum & Lipton, 2012) and other factors. One such
factor is that the GCS indicator (practical to classify patients clinically) is too brief since
mTBI develops heterogeneously in many ways. A more detailed and standardized
diagnosis could provide valuable information for neuroimaging analysis of mTBI
(Saatman et al., 2008). For example, the same GCS may be used for widely different CT
abnormalities such as epidural hematomas, contusions and parenchymal hematomas,
diffuse axonal injury, subdural hematoma, subarachnoid hemorrhage and intraventricular
hemorrhage, and diffuse brain swelling. However, the different abnormalities, lost if only
looking at GCS, may affect the patient differently at a later time points (Saatman et al.,
2008). The different types of injuries listed in Saatman et al. (2008) are also proof that
TBI (including mTBI) is very heterogeneous. Two other problematic heterogeneities are
individual mTBI uniqueness and diverse outcomes leading patient into different mental
disorders (APA, 2000; Rosenbaum & Lipton, 2012; Saatman et al., 2008). If the
heterogeneous nature of the imaging findings are not taken into account for the
neuroimaging detection method it may corrupt the results. Conversely, taking the
heterogeneity into account may complicate the detection method. That conventional
90
neuroimaging of most mTBI patients do not have any abnormalities may also challenge
detection (Shenton et al., 2012; Rosenbaum & Lipton, 2012). An additional shortcoming
when developing methods to detect mTBI is the loss of a pre-concussion baseline. There
are also problems occurring in mTBI groups with unknown premorbidity, but because
this thesis focuses on longitudinal models (the same participant measured at several time
points) the premorbidity issue may be less problematic than the premorbidity issue in
cross-sectional studies. Although our end goal is to detect mTBI, we present an
intermediate goal, which is to characterize how our GLM, SVM, permutation tests, and
regular t-test methods work before testing them on mTBI directly. A practical solution is
a model system that may include a pre-baseline and decrease executive function similarly
to mTBI (McCrea et al., 2003; Rohling et al., 2011). Note that we are not claiming our
model system will match the actual brain pattern of mTBI, but most other aspects should
be analogous. From the perspective that the patterns are not matching one may call our
model system naïve. We found that sleep deprivation fulfilled our requirements as a naïve
mTBI model. In addition, a control group without treatment is required from the same
population as the model system is provided. According to previous literature, sleep
deprivation increased reaction time (Drummond et al., 2006; Cain et al., 2011) and
induced significant neural patterns in the posterior and anterior cingulate, bilateral insula,
bilateral temporal parietal junction, and other regions (De Havas et al., 2012; Rosales-
Lagarde et al., 2012; Chee & Chuah, 2007).
To measure the severity of mTBI, it is popular to behaviorally measure an executive
function task (Miles et al., 2008; Niogi & Mukherjee, 2010). The multi-source
91
interference task (MSIT) is such a paradigm and is closely related with the Stroop task
(Bush et al., 2003; Sheu et al., 2012). Even though the MSIT has not been used to detect
mTBI, the Stroop task has successfully been used to do so (Binder et al., 1997; Bohnen et
al., 1992; Echemendia et al., 2001; Voller et al., 1999). The aim of this chapter is to
detect reversible effects from reversible executive function treatments using both fMRI
and behavioral measurements and the MSIT. Note that the decreased executive function
found in mTBI is reversed when the injured person recovers (McCrea et al., 2003). The
reaction time interference effect is more than twice as strong in MSIT (312 ms) compared
to the Stroop test (142 ms; Bush et al., 2003; Carter et al., 2000). In addition, the MSIT
neuroimaging signal is so robust in the dorsal anterior cingulate cortex that measurement
time can be decreased compared to other fMRI paradigms, including the Stroop task
(Bush et al., 2003). The dorsal anterior cingulate cortex may be the most important
location known to evaluate executive function using neuroimaging (Bush et al., 2003).
Decreased scanning time may be valuable in clinical settings. The MSIT has easy task
instructions, and both behavior and fMRI patterns are temporally stable (Bush et al.,
2003). Most importantly, Bush et al. (2003) stated that “MSIT yields temporally stable
performance and imaging data, making it useful in treatment and other types of
longitudinal studies” (p. 67), which means MSIT is not supposed to have any training
effects. Subjects in the study of Bush et al. (2003) performed the MSIT during 4 scans,
but in a single visit. Training effects for a group that makes several visits (on different
days) were not tested. The MSIT targets the executive function and induces cognitive
interference by combining three other tasks, which are the Stroop task, the Eriksen
92
flanker task, and the Simon effect, according to Bush et al. There are many variations on
the Stroop task using words, colors, sounds, and other stimuli. In this case, we use the
numbers 1, 2, and 3 and their positions. When 1 is in the first position, it is congruent and
in any other position incongruent. Similarly, the number 2 is congruent in the second
position, but incongruent in position 1 or 3 (Figure 26). The Eriksen flanker paradigm
delays subjects when the flanking positions contain symbols that could be targets in other
trials, but it does not in the current trial. Since the number 0 is never a target in our
version of MSIT, it has no Eriksen effect, which is used in the congruent trials. Finally,
the Simon effect has proven to decrease reaction time if an option is presented on the
screen ipsilateral to the button he or she presses compared with the button being placed
on the contralateral side. If the target information is presented on one side of the screen
and the button the subject is supposed to press is on the opposite side, the reaction time
will be slower. All three separate tasks combined into MSIT widened the reaction time
difference between the incongruent and congruent cases. Figure 26A depicts the MSIT.
Materials and Methods
Both the sleep-deprived (N = 8) and the control (N = 7) longitudinal groups used the
analogous imaging protocol at three time points (TPs) per subject. Subjects were medical
students aged 18 to 55 of both sexes. Both experiments were approved by Virginia
Tech’s institutional review board. The following protocol was identical at each TP:
• Neural imaging protocols were acquired using a Siemens 3T Trio.
o Structural T1 used MPRAGE with 1 mm isotropic voxels.
93
o An fMRI was acquired using a 3T Trio scanner with the following EPI
parameters:
TE/TR = 30/2000 ms
FA = 90°
FOV 220×220×145 mm3
Matrix (X, Y, Z) = 64×64×33
o Diffusion tensor imaging (DTI) was used.
o Susceptibility weighted imaging (SWI) was conducted.
• Two task types were performed while fMRIs were acquired:
o MSIT
o Default mode network (DMN)
• All image processing used the AFNI (Cox, 1996) software. The first level of
image processing used the afni_proc.py command with the following parameters:
o Smoothing with 6 mm kernel
o Contrast was: (+1){incongruent trials} (-1){congruent trials}
• Second-level analysis used either the 3dttest++ included in the AFNI package.
• Specific protocol for the sleep-deprivation group (N = 8):
o (TP1) Subjects had had 3 nights with more than 6h sleep per night.
o (TP2) Acquisition was performed 24 h after TP1, and subjects had not had any
sleep between scans. Sleepiness was measured using the Stanford Sleepiness
Scale (Figure 27).
94
o (TP3) Acquisition was performed between 48 h and 144 h after TP2, and
subjects had had two nights with more than 6 h sleep per night before the
scan.
o Participants’ sleep was assessed using a tablet actigraphy application.
• Control group (N = 7):
o (TP1) Acquisition was performed without treatment.
o (TP2) Acquisition was performed about 30 min after TP1, still without
treatment.
o (TP3) Acquisition was performed 24 h to 120 h after TP1 and TP2.
Statistical Analyses
The MSIT contained 288 congruent and 288 incongruent trials per visit (at TP1, TP2 or
TP3) and subject. From this the averages for congruent and incongruent reaction times
per subject and time point were calculated (tabulated in Table 6, and Table 7). For both
behavior and image data, TP1 and TP3 were treated as baselines. The null hypothesis was
set such that the behavior at TP2 was equal to the mean of the baselines for any specific
subject. Therefore, subtracting the average of the baselines from TP2 would result in zero
if the null hypothesis was true. Even though Bush et al. (2003) stated that MSIT does not
have any practice effect, this baseline subtraction approach works if practice effects are
present or not, assuming a linear approximation. In this way the three time points are
collapsed into a single value, for both behavior data and image data.
95
Figure 26: Multi-source interference task (MSIT) Above is the multisource interference task (MSIT) behavioral paradigm that mTBI and control subjects performed. (A) It consists of a difficult (incongruent) number combination and a simple (congruent) number combination. We see the distribution of reaction time in the congruent case (B) and in the incongruent case (C). (D) is the combination of B and C distributions.
96
Figure 27. Sleepiness according to the Stanford sleepiness scale (SSS) Eight subjects were kept awake between first and second neuroimaging time points (TP). The mean of their Stanford sleepiness scale (SSS) score was plotted.
97
Time point 1 Time point 2 Time point 3
Subject # Cong. [s] Incong. [s] Cong. [s] Incong. [s] Cong. [s] Incong. [s] MAA001 0.60 1.08 0.64 1.09 0.49 0.93 MAA002 0.62 0.97 0.72 0.98 0.59 0.88 MAA003 0.61 0.85 0.63 0.88 0.62 0.84 MAA004 0.54 0.90 0.53 0.83 0.51 0.79 MAA005 0.59 0.94 0.70 0.96 0.57 0.83 MAA006 0.59 0.95 0.66 0.96 0.54 0.82 MAA007 0.54 0.91 1.06 1.32 0.54 0.88 MAA008 0.66 0.90 0.73 0.96 0.60 0.83
Mean 0.59 0.94 0.71 1.00 0.56 0.85 SD 0.04 0.07 0.15 0.15 0.04 0.04
Table 6: Behavioral data from the MSIT sleep-deprived group List of acronyms: standard deviation (SD); congruent (Cong); incongruent (Incong); seconds (s).
98
Time point 1 Time point 2 Time point 3
Subject # Cong. [s] Incong. [s] Cong. [s] Incong. [s] Cong. [s] Incong. [s] SFACON1 0.55 0.87 0.49 0.80 0.49 0.80 SFACON2 0.63 1.01 0.63 1.01 0.59 0.88 SFACON3 0.92 1.05 0.65 0.98 0.55 0.84 SFACON4 0.75 1.09 0.71 1.08 0.63 0.97 SFACON5 0.52 0.78 0.50 0.76 0.48 0.71 SFACON6 0.54 0.87 0.51 0.80 0.54 0.79 SFACON7 0.62 0.85 0.57 0.80 0.57 0.81
Mean 0.65 0.93 0.58 0.89 0.55 0.83 SD 0.14 0.12 0.09 0.13 0.05 0.08
Table 7: Behavioral data from the MSIT pure control group Table acronyms: standard deviation (SD); congruent (Cong); incongruent (Incong); seconds (s).
99
Imaging Analyses
The fMRI software package AFNI (Cox, 1996) was used for all steps. A couple of types
of imaging were used. The simplest was the default MSIT neural pattern, which
generated at the first level contrast images for each control subject using the contrast +1
for incongruent beta values and -1 for congruent beta values. Continuing in a random
effects design, a one-sample t-test was then applied across all available first-level contrast
images, resulting in a second-level contrast image that revealed significant similarities
across the sample representing the default MSIT neural pattern (Bush et al., 2003).
For treatment effects, we used all three time points analogous to the statistical analysis of
the behavioral effects. After the first-level analysis (as described above, using contrast +1
and -1 for incongruent and congruent beta values) were performed for all three time
points. Then we collapsed the three contrast images for each time point into a single
image ( = 2*TP2-TP1-TP3) for each subject, which was a baseline subtraction (analogous
to what was previously done for the behavioral data). The second level was simply a two-
sample t-test between the sleep deprived and control group using 3dttest++. A false
discovery rate (FDR) threshold was set to p < 0.05 (two-tailed).
In order to maximize our effect and to test for an image detecting any practice effects a
paired t-test approach subtracting each subjects contrast image at TP3 from TP1 was
used. Still, the FDR was too conservative so we used the family wise approach and the
command 3dClustSim with an initial uncorrected threshold of p < 0.01. These settings
yielded a minimal cluster size of 70 voxels for a multiple comparisons corrected p < 0.05.
100
Results
After baseline removal, the scatterplot of each subject’s behavioral TP2 reaction time
shows that the sleep-deprived group had considerably longer reaction times than the
control group (Figure 28). All reaction times are also tabulated (Table 6 and Table 7).
A two-sample t-test between the sleep deprived group and the control group confirmed
that the sleep deprived group was significantly slower than control group. Significance
was p < 4.79*10-4 (1 tailed) in the congruent case and in the incongruent case p < 0.028
(1 tailed). Results are also tabulated in Table 8.
We hypothesized not to find any practice effects using the MSIT in accordance with
previous literature (Bush et al., 2003; Bush & Shin, 2006; Sheu et al., 2012). However,
the opposite results were significant using a two-sample t-test between TP1 and TP3 for
both congruent reaction times, p < 0.017, and even more significant practice effects for
incongruent reaction times, p < 0.001 (Figure 29; Table 8).
101
Figure 28: Behavioral response from MSIT cognitive insult treatments The scatter plot shows sleep deprivation (red squares) and control subjects (black crosses). One sleep-deprived outlier could not fit the graph (x=0.51 and y=0.43). The reaction times are baseline corrected by subtracting the mean of the pre- and post-baseline from time point (TP) 2. Pre-baseline is TP1 and post baseline is TP3. In addition a linear fit was applied for both sleep deprived and control subjects (green line).
102
Group RT type Mean (SD.) P-value N Control Congruent -0.018 (0.036) 1 7 Control Incongruent 0.011 (0.041) 1 7 Sleep-deprived Congruent 0.077 (0.045) 4.79*10-4 7 Sleep-deprived Incongruent 0.055 (0.038) 0.028 7 Sleep-deprived (including outlier)
Congruent 0.131 (0.160) 0.015 8
Sleep-deprived (including outlier)
Incongruent 0.101 (0.136) 0.056 8
Practice effects Congruent 0.017 15 Practice effects Incongruent 0.001 15 Table 8. Multi-source interference task behavior statistics Multi-source interference task (MSIT) mean reaction times (RT) and their significances compared to the control groups’ means using two-sample t-tests with exception for the practice effects groups. P-values from practice effects are all 15 subjects from TP1 compared to all subjects from TP3 using a two-sample t-test. All tests were one sided since sleep deprivation is expected to slow down RTs and practice is expected to improve RTs. SD is the standard deviation. N is number of subjects in accordance with the “Group” column.
103
Figure 29: MSIT behavioral practice effects Improving reaction times bolster the existence of practice effects. MSIT was first measured at time point 1 (TP1). Retest at TP3 was measured 2 to 6 days after TP1. The two-sample t-test calculated significant p-values between TP1 and TP3 using 15 subjects at both time points. Since subjects at TP2 were measured under different conditions, unrelated to practice effects, they are excluded from this figure.
104
The first imaging result was used for reference and is the control MSIT contrast image. It
was generated from TP1 using a one-sample t-test FDR-corrected at p < 0.05 and depicts
robust effects with pattern including the DLPFC and the dorsal anterior cingulate cortex
(dACC), which are the center of executive function (Banich, 2009; Bush et al., 2003;
Carter et al., 2000). As seen in Figure 30, the signal is abundant even at a corrected
threshold of p < 0.05 for both increased and decreased effects. These are regions that
have been implicated in various neuropsychiatric disorders (Bush et al., 2003).
Since a lot of behavioral results indicated possible practice effects we wondered if this
task might be exposed to practice effects at least for longitudinal studies despite previous
literature. We decided to test if the supposed practice effects were also visible in a
neuroimaging analysis. It is difficult to know if a decrease or increase of neural response
at the same time the behavior is improved represents practice effects and response
changes may be task dependent. However, since the Stroop task (which is closely related
to MSIT) using a similar model to our MSIT model yields a decrease of signal for
practice effects (Chen et al., 2013), a decrease in our results should represent practice
effects as well. That the Stroop task and MSIT really are similar is demonstrated by the
fact that the neural response patterns (in the control case) are extremely similar (Bush et
al., 2003; Sheu et al., 2012). Unfortunately, our test and retest MSIT contrast image
confirmed practice effects by a spatially abundant decrease for the retest that was
significant (p < 0.05, two-tailed) using a false discovery rate correction (Figure 31).
105
Figure 30: MSIT neural contrast image of control case Robust results are shown for the multisource interference task at time point 1 for 15 subjects. The threshold was a false discovery rate corrected at p < 0.05. Red is increased response and blue is decreased response, while z is the height in mm of the slice in accordance to the LPI convention. First-level analysis was an incongruent–congruent contrast. Second-level analysis was a one-sample t-test comparing the group to zero.
106
Figure 31: Neural contrast image of MSIT practice effects Practice effects induced mostly negative response (blue), but also a small positive response (red) in the ventral anterior cingulate. The response was significant (FDR corrected at p < 0.05). The signal decrease (blue) may represent a lower response at time point (TP) 3 compared to TP1. A total of 15 subjects were used in a paired t-test between T1 and T3.
107
A small region in the ventral anterior cingulate had a positive effect, which may be
ascribed to that the response actually is less negative in TP3 than in TP1. Note that the
ventral anterior cingulate is a region, which MSIT by default has a negative response at
(Figure 30).
Significant sleep deprivation treatment effects appeared as decreased neural responses in
several regions. Interestingly, the regions matched previous sleep deprivation literature
(Chee et al., 2008; De Havas et al., 2012) in terms of both location and direction:
posterior cingulate (increased), dorsomedial prefrontal cortex (increased), bilateral insula
(decreased), and bilateral intraparietal sulcus (decreased; Figure 32).
Since the MSIT generates a stronger signal response compared to the Stroop task (Bush
et al., 2003) it may have supported our results.
108
Figure 32: Neural contrast image of MSIT sleep-deprivation treatment Related response with sleep deprivation was posterior cingulate (positive; red), dorsomedial prefrontal cortex (positive), bilateral insula (negative; blue), and bilateral intraparietal sulcus (negative) in accordance with previous literature (Chee et al., 2008; Chee & Chuah, 2007; De Havas et al., 2012). The signal was significant with multiple comparisons correction (p < 0.05).
109
Discussion
Our longitudinal design allowed us to perform a paired statistical test, which may have a
stronger sensitivity than a regular two-sample model, having only 7 to 8 subjects per
group. The baseline before the sleep deprivation was an important advantage as well.
Using baseline correction, subjects were about 0.09 s (with a standard deviation of 0.06 s)
slower at MSIT when they were sleep deprived compared to when they had had sleep.
However, MAA007 (Table 6) was 0.65 s slower when sleep deprived. MAA007 agreed
with our hypothesis (that sleep deprivation slows down reaction time), but the enormous
slowdown for MAA007 would inflate the standard deviation of the entire group so much
that the t-test would stop working properly. MAA007 was an outlier and was eliminated
from our behavioral data.
Even though the MSIT is simple, it may be confusing (similar to how the Stroop task
may be confusing). A few subjects who reported knowing how to do the task made
incorrect behavior responses anyway. Therefore, Dr. LaConte and I developed an
interface that displays real-time subject responses (Figure 33). The interface notified the
scanner technician immediately if the subject responded in an invalid pattern, which
enabled the technician to immediately stop the data collection. Then the technician could
notify the subject of the correct behavior and restart the task. All this happened with
minimal intervention and went very smoothly. Without this real-time feature, the invalid
data would not have been discovered until the time for analysis and could have corrupted
110
that subject’s data for the study. In retrospect, it is clear the real-time feature saved
several subjects’ data.
Due to a slight time difference for sleep deprived and control subjects after TP1 we
decided to research if this could be an issue. Since both groups measured TP3 about 2 to
6 days after TP2 we thought deviations between groups were negligible. The MSIT is
similar to the Stroop task (Bush et al., 2003; Sheu et al., 2012), and our results are
consistent with other behavior results from the Stroop task (Chen et al., 2013). In
addition, the extremely similar reaction times mean at TP3 for the control, sleep-
deprived, and an additional unpublished group support that the time differences between
TP2 and TP3 are negligible as long they are within 10 days of TP1.
111
Figure 33: Real-time MSIT feedback The interface for the lab technician to monitor the subject behavior in real time is shown. If the technician noted that the subject had misunderstood an important part of the task, then the scan could be interrupted. After a quick reminder, the same subject would respond correctly. This saved valuable subject data from corruption. Both the X-axis and Y-axis represent time in seconds. The vertical lines in the bottom row show the reaction time. If the line goes below the baseline, it means that the subject made a mistake. The top row has red bars showing the mean reaction time of 3 congruent and 3 incongruent blocks. The text “Errors: 0,2,1,2,1,0” means that Block 2 had 2 errors, Block 3 had 1 error, Block 4 had 1 error, and Block 5 also had one error.
112
Conclusions
Sleep deprivation causes a significant behavioral reaction time increase, and neuronal
image effects decrease when using the MSIT task. The sleep deprivation paradigm was
useful for testing our methods in several respects. First of all, undocumented practice
effects for the MSIT paradigm were found. Without this knowledge, any practice effects
in the mTBI study could have led to overinterpretation of the recovering factor in an
mTBI group. Even though the sleep deprivation statistical brain pattern is probably
different from the mTBI statistical pattern, this proved that we could image a reversible
cognitive insult stage using a repeated-measures statistical model. The behavioral
response from the sleep deprivation treatment was analogous to how mTBI responds to
the Stroop task (reaction time changes). From the behavioral perspective, sleep
deprivation seems to behave similarly as mTBI that may be beneficial when used as a
model system for mTBI. Since practice effects were found, it will be much harder to
analyze the mTBI data because the improvements in the mTBI sample may be a
combination of mTBI recovery and practice effects.
113
Chapter 5 Longitudinal mTBI
Introduction
Previous fMRI studies have found significant differences between mTBI and control
groups (Eierud et al., 2014). Several mTBI experts have called for longitudinal mTBI
studies to increase the sensitivity to mTBI for several reasons. One reason of interest is to
image tissue changes at different stages post-injury (Loane and Faden, 2010). In short,
tissues damaged by mTBI may break down in apoptosis, heal so they will function
properly again, or progress by an alternate effect. Images at early and late stages after
injury may then be compared to pinpoint tissue modifications (Mayer et al., 2010; Niogi
et al, 2008a; Niogi et al, 2008b). Another reason is that longitudinal studies may be
required to predict between mTBI recovery and chronic mTBI disabilities (Niogi &
Mukherjee, 2010). In addition, longitudinal studies may have an advantage in using a
statistical model that includes a paired t-test compared to cross-sectional studies that may
not use paired t-tests. Further advantages of longitudinal imaging may be found in the
literature (Lo et al., 2009; Mayer et al., 2009).
Our final goal is to track the mTBI recovery temporally from acute to chronic phase
using both executive function behavior and neuroimaging. Most of our analysis on mTBI
subjects has been at the behavioral level so far, which is presented.
114
Materials and Methods
Twenty mTBI subjects were longitudinally measured at four time points (TPs). The mean
time after injury at which imaging was acquired was: time point (TP) 1 = 4 days (N =
20), TP2 = 3.5 months (N = 13), TP3 = 7.2 months (N = 10), and TP4 = 9.6 months (N =
7). All subjects were male or female 18 to 55 of age. All three experiments were
approved by Virginia Tech’s institutional review board. The following protocol was
identical at each TP:
• Neural imaging protocols were acquired using a Siemens 3T Trio.
o Structural T1 used MPRAGE with 1 mm isotropic voxels.
o An fMRI was acquired using a 3T Trio scanner with the following EPI
parameters:
TE/TR = 30/2000 ms
FA = 90°
FOV 220×220×145 mm3
Matrix (X, Y, Z) = 64×64×33
o Diffusion tensor imaging (DTI)
o Susceptibility weighted imaging (SWI)
• Two task types were performed while fMRIs were acquired:
o MSIT
o Default mode network (DMN)
115
Results
The mTBI results shared in this thesis were that MSIT reaction times significantly
improved with time and number of visits. In the congruent case, significant improvement
(p < 0.05, one sided two-sample t-test) was found between TP1 and TP3. In the
incongruent case, significant improvements (p < 0.05, two-sided two-sample t-tests) were
found between TP1 and TP3 and between TP2 and TP3. The improvements are depicted
in bar diagrams (Figure 34A and Figure 34B). Each subject usually improved reaction
times for each visit, depicted as arrow trajectories for each subject (Figure 34C).
116
Figure 34: mTBI reaction time improvements with time post-concussion
Depiction of mTBI subjects’ reaction time improvements for the MSIT paradigm are shown. (A) Bars depict mean reaction times, in the congruent MSIT condition for time points (TPs) 1 to 4. (B) is analogous to figure A, but shows reaction times for the incongruent condition. One asterisk means a one-sided significant decrease. Two asterisks signify a two-sided significant decrease. (C) Each arrow shows a subject’s reaction time difference between two TPs. Subjects who were measured four times are depicted with a trajectory of three arrows. The color of the arrow represents the days post-injury the measurement was taken in accordance to the legend. Including control subjects (gray dots), a significant positive correlation (ρ = 0.72, p < 2*10-16) is seen for congruent and incongruent reaction time pairs. A linear fit across all the reaction time pairs is drawn in black. Most arrow trajectories trace the subject improvements per visit for both congruent and incongruent reaction times. Further improvements follow the linear fit orderly in most cases. Subjects and time post-injury for each time point (TP) were: TP1, N = 20, 4 days; TP2, N = 13, 3.5 months; TP3, N = 10, 7.2 months; TP4, N = 7; 9.6 months.
117
Discussion
As seen in Figure 34, the mTBI subjects have significantly improved their MSIT reaction
times with time and visits. To some degree, this improvement may reflect mTBI
recovery. Unfortunately, this improvement may not exclusively be from mTBI recovery,
but may also be from practice effects, which was revealed in Chapter 4. To tease out
recovery from practice effects and to use SVM for neuroimaging will be important future
work.
118
Chapter 6: Summary, Significance, and Future Goals
Meta-Analysis Conclusions and Significance
Previous mTBI literature has been limited to studies on mTBI subjects, reviews of these
studies, and a few mTBI meta-analyses. Meta-analyses include one complete evaluation
of the mTBI neuroimaging literature (Borg et al., 2004) and a few CT related mTBI
meta-analyses often from a clinical perspective (Jagoda et al., 2008; Stein et al., 2008).
Since the neuroimaging-mTBI literature recently surpassed 122 publications, we thought
it was time to conduct a new, thorough meta-analysis. The resulting meta-analysis found
several significant results for DTI (21 publications) and fMRI (7 publications) with
significant support for the finding that anterior regions are more abundantly reported to
be abnormal in mTBI for both fMRI and DTI independently. Note that executive function
paradigms that are negatively affected in mTBI (McCrea et al., 2003; Rohling et al.,
2011) may be related with anterior regions as well (Banich, 2009; Bush et al., 2003;
Carter et al., 2000). In addition, our meta-analysis found a significant relationship
between anterior regions using DTI and scores from neuropsychological assessments.
Furthermore, the DTI abnormalities seem to have a temporal dependency. A categorical
test significantly found that acute mTBI groups tend to have higher anisotropy, and
chronic mTBI groups tend to have significantly lower anisotropy. Most remarkable
would be if the results that depend on time post-injury are true even in larger studies.
119
Then it may be a marker of secondary effects in mTBI, scarred tissue, or some other
process that changes over time post-injury and may become a tool that can be used
clinically to track mTBI progression. In addition, we were able to generate an mTBI ALE
map. Because only seven mTBI fMRI publications existed, we were surprised that we
could detect 13 significant clusters from our ALE analysis. This mTBI ALE map may be
of value to match future mTBI fMRI publications. If mTBI fMRI publications that image
time points after 3 months post-injury increase in number, an mTBI-fMRI meta-analysis
across time after injury (similar to Figure 10, but for fMRI) may be generated. A
temporal mTBI-fMRI analysis could give further insight of the time course of mTBI.
Support Vector Machine Conclusions and Significance
According to previous literature, SVM has been proven to predict multidimensional data
accurately and is well applicable for fMRI (Cox & Savoy, 2003; Fan et al., 2007;
LaConte et al., 2005; LaConte et al., 2007; LaConte et al., 2011; Mitchell et al., 2004). If
SVM becomes more specialized to track mTBI, it may be more sensitive than competing
fMRI methods. In this thesis we characterized support vector machines for regression and
classification, and the intermediate results were that SVR can predict rates and that the
imaging of the weights is related with the behavioral paradigm in accordance with
previous literature. We also found that the blood-oxygen-level dependence response from
our subjects deviates from the contemporary hemodynamic response function hypothesis
(Buxton et al., 1998; Glover, 1999) that may corrupt the SVM predictions. Our
significant findings in fMRI and SVM research may have brought the fields closer to
120
mTBI. However, further research is needed to complete the bridge between SVM and
mTBI.
From a more practical view, we found that real-time behavioral feedback improves
subject accuracy on difficult tasks, which was observed from our subjects (Figure 25).
Model System for Mild Traumatic Brain Injury
Working with mTBI may be difficult due to challenges finding subjects who are willing
to attend a study, finding subjects with the same amount of mTBI, and other reasons.
Most importantly, it is almost impossible to get a baseline before mTBI has occurred. A
sleep deprivation model system may solve these problems at the expense that the neural
pattern may not match with mTBI. The model system may still validate most parts of our
mTBI methods.
With help from our model system, we found that the MSIT paradigm was exposed to
training effect across subjects’ multiple visits, which was surprising because previous
literature has reported not having found any such effects (Bush et al., 2003; Bush and
Shin, 2006; Sheu et al., 2012).
Behaviorally, sleep deprivation slows down the executive function similarly to how
mTBI does. In the end, we will use our methods to study mTBI. Still, our model system
detected the reversible treatment of sleep deprivation. Sleep deprivation detection was
121
found in the posterior cingulate (increased), dorsomedial prefrontal cortex (increased),
bilateral insula (decreased), and bilateral intraparietal sulcus (decreased; Figure 32; Chee
et al., 2008; De Havas et al., 2012). Since the reversible sleep deprivation was
successfully validated, our model system has the potential for further research. The next
step could be to implement SVM into the model system before it is tested on mTBI.
Results From mTBI Subjects
In our mTBI subjects, results showed that they improved reaction times with time and
visits. However, the improvements may have been not only because of recovery from
mTBI, but also because of a training effect. Revealing that our mTBI subjects may have
been exposed to training effects complicates the interpretation of our mTBI results. It is
important to perform additional control tests to figure out how to separate the training
effect from the recovery effect within our mTBI subjects.
122
References
Ackermann, H., Riecker, A., Mathiak, K., Erb, M., Grodd, W., et al., 2001. Rate-
dependent activation of a prefrontal-insular-cerebellar network during passive listening to
trains of click stimuli: an fMRI study. Neuroreport 12, 4087-4092.
Adams, J.H., Graham, D.I., Scott, G., Parker, L.S., Doyle, D., 1980. Brain damage in
fatal non-missile head injury. Journal of clinical pathology 33, 1132-1145.
Alexander, M.P., 1995. Mild traumatic brain injury: pathophysiology, natural history, and
clinical management. Neurology 45, 1253-1260.
American Psychiatric Association, 2000. Diagnostic and Statistical Manual of Mental
Disorders. American Psychiatric Association, Fourth Edition, Washington, DC, Text
Revision, 760-762.
Aramaki, Y., Honda, M., Sadato, N., 2006. Suppression of the non-dominant motor
cortex during bimanual symmetric finger movement: a functional magnetic resonance
imaging study. Neuroscience 141, 2147-2153.
123
Arciniegas, D.B., Anderson, C.A., Topkoff, J., McAllister, T.W., 2005. Mild traumatic
brain injury: a neuropsychiatric approach to diagnosis, evaluation, and treatment.
Neuropsychiatric disease and treatment 1, 311-327.
Banich, M. T., 2009. Executive function the search for an integrated account. Current
Directions in Psychological Science, 18(2), 89-94.
Bazarian, J.J., Zhong, J., Blyth, B., Zhu, T., Kavcic, V., et al., 2007. Diffusion tensor
imaging detects clinically important axonal damage after mild traumatic brain injury: a
pilot study. Journal of neurotrauma 24, 1447-1459.
Beaumont, A., Gennarelli, T., 2006. CT prediction of contusion evolution after closed
head injury: the role of pericontusional edema. Acta neurochirurgica. Supplement 96, 30-
32.
Belanger, H.G., Vanderploeg, R.D., Curtiss, G., Warden, D.L., 2007. Recent
neuroimaging techniques in mild traumatic brain injury. The Journal of neuropsychiatry
and clinical neurosciences 19, 5-20.
Bernstein, M. A., King, K. F., & Zhou, X. J. (2004). Handbook of MRI pulse sequences.
Elsevier.
124
Bigler, E.D., 2007. Anterior and middle cranial fossa in traumatic brain injury: relevant
neuroanatomy and neuropathology in the study of neuropsychological outcome.
Neuropsychology 21, 515-531.
Binder, L.M., Rohling, M.L., Larrabee, G.J., 1997. A review of mild head trauma. Part I:
Meta-analytic review of neuropsychological studies. Journal of clinical and experimental
neuropsychology 19, 421-431.
Blinkenberg, M., Bonde, C., Holm, S., Svarer, C., Andersen, J., et al., 1996. Rate
dependence of regional cerebral activation during performance of a repetitive motor task:
a PET study. Journal of cerebral blood flow and metabolism : official journal of the
International Society of Cerebral Blood Flow and Metabolism 16, 794-803.
Bohnen, N., Jolles, J., 1992. Neurobehavioral aspects of postconcussive symptoms after
mild head injury. The Journal of nervous and mental disease 180, 683-692.
Borg, J., Holm, L., Cassidy, J.D., Peloso, P.M., Carroll, L.J., et al., 2004. Diagnostic
procedures in mild traumatic brain injury: results of the WHO Collaborating Centre Task
Force on Mild Traumatic Brain Injury. Journal of rehabilitation medicine : official
journal of the UEMS European Board of Physical and Rehabilitation Medicine, 61-75.
125
Brain Injury Association of America. (2014, August 15). Home page. Retrieved from
http://www.biausa.org/
Brandstack, N., Kurki, T., Tenovuo, O., Isoniemi, H., 2006. MR imaging of head trauma:
visibility of contusions and other intraparenchymal injuries in early and late stage. Brain
injury : [BI] 20, 409-416.
Budde, M.D., Janes, L., Gold, E., Turtzo, L.C., Frank, J.A., 2011. The contribution of
gliosis to diffusion tensor anisotropy and tractography following traumatic brain injury:
validation in the rat using Fourier analysis of stained tissue sections. Brain : a journal of
neurology 134, 2248-2260.
Buki, A., Povlishock, J.T., 2006. All roads lead to disconnection?--Traumatic axonal
injury revisited. Acta neurochirurgica 148, 181-193; discussion 193-184.
Bush, G., Shin, L.M., 2006. The Multi-Source Interference Task: an fMRI task that
reliably activates the cingulo-frontal-parietal cognitive/attention network. Nature
protocols 1, 308-313.
Bush, G., Shin, L.M., Holmes, J., Rosen, B.R., Vogt, B.A., 2003. The Multi-Source
Interference Task: validation study with fMRI in individual subjects. Molecular
psychiatry 8, 60-70.
126
Buxton, R.B., Wong, E.C., Frank, L.R., 1998. Dynamics of blood flow and oxygenation
changes during brain activation: the balloon model. Magnetic resonance in medicine :
official journal of the Society of Magnetic Resonance in Medicine / Society of Magnetic
Resonance in Medicine 39, 855-864.
Cain, S.W., Silva, E.J., Chang, A.M., Ronda, J.M., Duffy, J.F., 2011. One night of sleep
deprivation affects reaction time, but not interference or facilitation in a Stroop task.
Brain and cognition 76, 37-42.
Carter, C.S., Macdonald, A.M., Botvinick, M., Ross, L.L., Stenger, V.A., et al., 2000.
Parsing executive processes: strategic vs. evaluative functions of the anterior cingulate
cortex. Proceedings of the National Academy of Sciences of the United States of America
97, 1944-1948.
Cassidy, J.D., Carroll, L.J., Peloso, P.M., Borg, J., von Holst, H., et al., 2004. Incidence,
risk factors and prevention of mild traumatic brain injury: results of the WHO
Collaborating Centre Task Force on Mild Traumatic Brain Injury. Journal of
rehabilitation medicine : official journal of the UEMS European Board of Physical and
Rehabilitation Medicine, 28-60.
127
Chee, M.W., Chuah, Y.M., 2007. Functional neuroimaging and behavioral correlates of
capacity decline in visual short-term memory after sleep deprivation. Proceedings of the
National Academy of Sciences of the United States of America 104, 9487-9492.
Chee, M.W., Tan, J.C., Zheng, H., Parimal, S., Weissman, D.H., et al., 2008. Lapsing
during sleep deprivation is associated with distributed changes in brain activation. The
Journal of neuroscience : the official journal of the Society for Neuroscience 28, 5519-
5528.
Chen, J.K., Johnston, K.M., Collie, A., McCrory, P., Ptito, A., 2007. A validation of the
post concussion symptom scale in the assessment of complex concussion using cognitive
testing and functional MRI. Journal of neurology, neurosurgery, and psychiatry 78, 1231-
1238.
Chen, Z., Lei, X., Ding, C., Li, H., & Chen, A., 2013. The neural mechanisms of
semantic and response conflicts: An fMRI study of practice-related effects in the Stroop
task. NeuroImage, 66, 577-584.
Cherkassky, V., & Mulier, F. M., 2007. Learning from data: concepts, theory, and
methods. John Wiley & Sons.
128
Chu, Z., Wilde, E.A., Hunter, J.V., McCauley, S.R., Bigler, E.D., et al., 2010. Voxel-
based analysis of diffusion tensor imaging in mild traumatic brain injury in adolescents.
AJNR. American journal of neuroradiology 31, 340-346.
Cortes, C., & Vapnik, V., 1995. Support-vector networks. Machine learning, 20(3), 273-
297.
Cox, D.D., Savoy, R.L., 2003. Functional magnetic resonance imaging (fMRI) "brain
reading": detecting and classifying distributed patterns of fMRI activity in human visual
cortex. NeuroImage 19, 261-270.
Cox, R.W., 1996. AFNI: software for analysis and visualization of functional magnetic
resonance neuroimages. Computers and biomedical research, an international journal 29,
162-173.
D'Esposito, M., Detre, J.A., Alsop, D.C., Shin, R.K., Atlas, S., et al., 1995. The neural
basis of the central executive system of working memory. Nature 378, 279-281.
Damadian, R., 1971. Tumor detection by nuclear magnetic resonance. Science 171, 1151-
1153.
129
Damadian, R., Zaner, K., Hor, D., DiMaio, T., 1974. Human tumors detected by nuclear
magnetic resonance. Proceedings of the National Academy of Sciences of the United
States of America 71, 1471-1473.
De Havas, J.A., Parimal, S., Soon, C.S., Chee, M.W., 2012. Sleep deprivation reduces
default mode network connectivity and anti-correlation during rest and task performance.
NeuroImage 59, 1745-1751.
deCharms, R.C., Maeda, F., Glover, G.H., Ludlow, D., Pauly, J.M., et al., 2005. Control
over brain activation and pain learned by using real-time functional MRI. Proceedings of
the National Academy of Sciences of the United States of America 102, 18626-18631.
Deiber, M.P., Honda, M., Ibanez, V., Sadato, N., Hallett, M., 1999. Mesial motor areas in
self-initiated versus externally triggered movements examined with fMRI: effect of
movement type and rate. Journal of neurophysiology 81, 3065-3077.
Dikmen, S., Reitan, R.M., Temkin, N.R., Machamer, J.E., 1992. Minor and severe head
injury emotional sequelae. Brain injury : [BI] 6, 477-478.
Drummond, S.P., Paulus, M.P., Tapert, S.F., 2006. Effects of two nights sleep
deprivation and two nights recovery sleep on response inhibition. Journal of sleep
research 15, 261-265.
130
Echemendia, R.J., Putukian, M., Mackin, R.S., Julian, L., Shoss, N., 2001.
Neuropsychological test performance prior to and following sports-related mild traumatic
brain injury. Clinical journal of sport medicine : official journal of the Canadian
Academy of Sport Medicine 11, 23-31.
Eickhoff, S.B., Bzdok, D., Laird, A.R., Kurth, F., Fox, P.T., 2012. Activation likelihood
estimation meta-analysis revisited. NeuroImage 59, 2349-2361.
Eickhoff, S.B., Laird, A.R., Grefkes, C., Wang, L.E., Zilles, K., et al., 2009. Coordinate-
based activation likelihood estimation meta-analysis of neuroimaging data: a random-
effects approach based on empirical estimates of spatial uncertainty. Human brain
mapping 30, 2907-2926.
Eierud, C., Craddock, R. C., Fletcher, S., Aulakh, M., King-Casas, B., Kuehl, D., &
LaConte, S. M., 2014. Neuroimaging after mild traumatic brain injury: Review and meta-
analysis. NeuroImage: Clinical, 4, 283-294.
Ellmore, T.M., Li, H., Xue, Z., Wong, S.T., Frye, R.E., 2013. Tract-based spatial
statistics reveal altered relationship between non-verbal reasoning abilities and white
matter integrity in autism spectrum disorder. Journal of the International
Neuropsychological Society : JINS 19, 723-728.
131
Evans, R.W., 1992. The postconcussion syndrome and the sequelae of mild head injury.
Neurologic clinics 10, 815-847.
Fan, Y., Rao, H., Hurt, H., Giannetta, J., Korczykowski, M., et al., 2007. Multivariate
examination of brain abnormality using both structural and functional MRI. NeuroImage
36, 1189-1199.
Fisher, R. A., 1936. The use of multiple measurements in taxonomic problems. Annals of
eugenics, 7(2), 179-188.
FitzGerald, D.B., Crosson, B.A., 2011. Diffusion weighted imaging and
neuropsychological correlates in adults with mild traumatic brain injury. International
journal of psychophysiology : official journal of the International Organization of
Psychophysiology 82, 79-85.
Freitas, R. A., 1998. Nanomedicine Foresight Institute, 749-752, 817.
Friston, K. J., Holmes, A. P., Worsley, K. J., Poline, J. P., Frith, C. D., & Frackowiak, R.
S., 1994. Statistical parametric maps in functional imaging: a general linear approach.
Human brain mapping, 2(4), 189-210.
132
Funahashi, S., Bruce, C.J., Goldman-Rakic, P.S., 1989. Mnemonic coding of visual space
in the monkey's dorsolateral prefrontal cortex. Journal of neurophysiology 61, 331-349.
Gasparovic, C., Yeo, R., Mannell, M., Ling, J., Elgie, R., et al., 2009. Neurometabolite
concentrations in gray and white matter in mild traumatic brain injury: an 1H-magnetic
resonance spectroscopy study. Journal of neurotrauma 26, 1635-1643.
Geary, E.K., Kraus, M.F., Pliskin, N.H., Little, D.M., 2010. Verbal learning differences
in chronic mild traumatic brain injury. Journal of the International Neuropsychological
Society : JINS 16, 506-516.
Gennarelli, T.A., Thibault, L.E., Adams, J.H., Graham, D.I., Thompson, C.J., et al., 1982.
Diffuse axonal injury and traumatic coma in the primate. Annals of neurology 12, 564-
574.
Giza, C.C., Hovda, D.A., 2001. The Neurometabolic Cascade of Concussion. Journal of
athletic training 36, 228-235.
Glover, G.H., 1999. Deconvolution of impulse response in event-related BOLD fMRI.
NeuroImage 9, 416-429.
133
Gross, C.G., 1992. Representation of visual stimuli in inferior temporal cortex.
Philosophical transactions of the Royal Society of London. Series B, Biological sciences
335, 3-10.
Grossman, E.J., Ge, Y., Jensen, J.H., Babb, J.S., Miles, L., et al., 2012. Thalamus and
cognitive impairment in mild traumatic brain injury: a diffusional kurtosis imaging study.
Journal of neurotrauma 29, 2318-2327.
Gurdjian, E.S., 1975. Re-evaluation of the biomechanics of blunt impact injury of the
head. Surgery, gynecology & obstetrics 140, 845-850.
Guskiewicz, K.M., McCrea, M., Marshall, S.W., Cantu, R.C., Randolph, C., et al., 2003.
Cumulative effects associated with recurrent concussion in collegiate football players: the
NCAA Concussion Study. JAMA : the journal of the American Medical Association 290,
2549-2555.
Guskiewicz, K.M., Mihalik, J.P., Shankar, V., Marshall, S.W., Crowell, D.H., et al.,
2007. Measurement of head impacts in collegiate football players: relationship between
head impact biomechanics and acute clinical outcome after concussion. Neurosurgery 61,
1244-1252; discussion 1252-1243.
134
Guskiewicz, K.M., Perrin, D.H., Gansneder, B.M., 1996. Effect of mild head injury on
postural stability in athletes. Journal of athletic training 31, 300-306.
Hanlon, R.E., Demery, J.A., Martinovich, Z., Kelly, J.P., 1999. Effects of acute injury
characteristics on neuropsychological status and vocational outcome following mild
traumatic brain injury. Brain injury : [BI] 13, 873-887.
Hashimoto, K., Abo, M., 2009. Abnormal regional benzodiazepine receptor uptake in the
prefrontal cortex in patients with mild traumatic brain injury. Journal of rehabilitation
medicine : official journal of the UEMS European Board of Physical and Rehabilitation
Medicine 41, 661-665.
He, S.Q., Dum, R.P., Strick, P.L., 1995. Topographic organization of corticospinal
projections from the frontal lobe: motor areas on the medial surface of the hemisphere.
The Journal of neuroscience : the official journal of the Society for Neuroscience 15,
3284-3306.
Holli, K.K., Waljas, M., Harrison, L., Liimatainen, S., Luukkaala, T., et al., 2010. Mild
traumatic brain injury: tissue texture analysis correlated to neuropsychological and DTI
findings. Academic radiology 17, 1096-1102.
135
Holm, L., Cassidy, J.D., Carroll, L.J., Borg, J., 2005. Summary of the WHO
Collaborating Centre for Neurotrauma Task Force on Mild Traumatic Brain Injury.
Journal of rehabilitation medicine 37, 137-141.
Horak, F.B., Diener, H.C., 1994. Cerebellar control of postural scaling and central set in
stance. Journal of neurophysiology 72, 479-493.
Hunter, J.V., Wilde, E.A., Tong, K.A., Holshouser, B.A., 2012. Emerging imaging tools
for use with traumatic brain injury research. Journal of neurotrauma 29, 654-671.
Huynh, T., Jacobs, D.G., Dix, S., Sing, R.F., Miles, W.S., et al., 2006. Utility of
neurosurgical consultation for mild traumatic brain injury. The American surgeon 72,
1162-1165; discussion1166-1167.
Ilvesmäki, T., Luoto, T. M., Hakulinen, U., Brander, A., Ryymin, P., et al., 2014. Acute
mild traumatic brain injury is not associated with white matter change on diffusion tensor
imaging. Brain, awu095.
Inglese, M., Bomsztyk, E., Gonen, O., Mannon, L.J., Grossman, R.I., et al., 2005. Dilated
perivascular spaces: hallmarks of mild traumatic brain injury. AJNR. American journal of
neuroradiology 26, 719-724.
136
Iverson, G.L., 2005. Outcome from mild traumatic brain injury. Current opinion in
psychiatry 18, 301-317.
Iverson, G.L., 2006. Complicated vs uncomplicated mild traumatic brain injury: acute
neuropsychological outcome. Brain injury : [BI] 20, 1335-1344.
Iverson, G.L., Lange, R.T., 2003. Examination of "postconcussion-like" symptoms in a
healthy sample. Applied neuropsychology 10, 137-144.
Jagoda, A.S., Bazarian, J.J., Bruns, J.J., Jr., Cantrill, S.V., Gean, A.D., et al., 2008.
Clinical policy: neuroimaging and decisionmaking in adult mild traumatic brain injury in
the acute setting. Annals of emergency medicine 52, 714-748.
Jancke, L., Peters, M., Himmelbach, M., Nosselt, T., Shah, J., et al., 2000. fMRI study of
bimanual coordination. Neuropsychologia 38, 164-174.
Jancke, L., Peters, M., Schlaug, G., Posse, S., Steinmetz, H., et al., 1998. Differential
magnetic resonance signal change in human sensorimotor cortex to finger movements of
different rate of the dominant and subdominant hand. Brain research. Cognitive brain
research 6, 279-284.
137
Jancke, L., Specht, K., Mirzazade, S., Peters, M., 1999. The effect of finger-movement
speed of the dominant and the subdominant hand on cerebellar activation: A functional
magnetic resonance imaging study. NeuroImage 9, 497-507.
Jennett, B., 1998. Epidemiology of head injury. Archives of disease in childhood 78, 403-
406.
Joachims, T. (1999). Making large-Scale SVM Learning Practical. Advances in Kernel
Methods-Support Vector Learning. Schölkopf B. and Burges C. and Smola A. MIT
Press, 1999.
Kane, M.J., Engle, R.W., 2002. The role of prefrontal cortex in working-memory
capacity, executive attention, and general fluid intelligence: an individual-differences
perspective. Psychonomic bulletin & review 9, 637-671.
Kawashima, R., Inoue, K., Sugiura, M., Okada, K., Ogawa, A., et al., 1999. A positron
emission tomography study of self-paced finger movements at different frequencies.
Neuroscience 92, 107-112.
Kraus, M.F., Susmaras, T., Caughlin, B.P., Walker, C.J., Sweeney, J.A., et al., 2007.
White matter integrity and cognition in chronic traumatic brain injury: a diffusion tensor
imaging study. Brain : a journal of neurology 130, 2508-2519.
138
Krivitzky, L.S., Roebuck-Spencer, T.M., Roth, R.M., Blackstone, K., Johnson, C.P., et
al., 2011. Functional magnetic resonance imaging of working memory and response
inhibition in children with mild traumatic brain injury. Journal of the International
Neuropsychological Society : JINS 17, 1143-1152.
Kurca, E., Sivak, S., Kucera, P., 2006. Impaired cognitive functions in mild traumatic
brain injury patients with normal and pathologic magnetic resonance imaging.
Neuroradiology 48, 661-669.
Kushner, D., 1998. Mild traumatic brain injury: toward understanding manifestations and
treatment. Archives of internal medicine 158, 1617-1624.
LaConte, S., Strother, S., Cherkassky, V., Anderson, J., Hu, X., 2005. Support vector
machines for temporal classification of block design fMRI data. NeuroImage 26, 317-
329.
LaConte, S.M., 2011. Decoding fMRI brain states in real-time. NeuroImage 56, 440-454.
LaConte, S.M., Peltier, S.J., Hu, X.P., 2007. Real-time fMRI using brain-state
classification. Human brain mapping 28, 1033-1044.
139
Landre, N., Poppe, C.J., Davis, N., Schmaus, B., Hobbs, S.E., 2006. Cognitive
functioning and postconcussive symptoms in trauma patients with and without mild TBI.
Archives of clinical neuropsychology : the official journal of the National Academy of
Neuropsychologists 21, 255-273.
Lange, R.T., Iverson, G.L., Brubacher, J.R., Madler, B., Heran, M.K., 2012. Diffusion
tensor imaging findings are not strongly associated with postconcussional disorder 2
months following mild traumatic brain injury. The Journal of head trauma rehabilitation
27, 188-198.
Langlois, J.A., Rutland-Brown, W., Thomas, K.E., 2004. Traumatic brain injury in the
United States: emergency department visits, hospitalizations, and deaths. Department of
Health and Human Services, Centers for Disease Control and Prevention, Division of
Acute Care, Rehabilitation Research and Disability Prevention, National Center for
Injury Prevention and Control.
Lauterburg, P.C., 1973. Image Formation by Induced LocalInteractions: Expamples
Employing Nuclear Magnetic Resonance. Nature 242, 190-191.
Lee, H., Wintermark, M., Gean, A.D., Ghajar, J., Manley, G.T., et al., 2008. Focal lesions
in acute mild traumatic brain injury and neurocognitive outcome: CT versus 3T MRI.
Journal of neurotrauma 25, 1049-1056.
140
Lehericy, S., Bardinet, E., Tremblay, L., Van de Moortele, P.F., Pochon, J.B., et al.,
2006. Motor control in basal ganglia circuits using fMRI and brain atlas approaches.
Cerebral cortex 16, 149-161.
Levin, H.S., Williams, D.H., Eisenberg, H.M., High, W.M., Jr., Guinto, F.C., Jr., 1992.
Serial MRI and neurobehavioural findings after mild to moderate closed head injury.
Journal of neurology, neurosurgery, and psychiatry 55, 255-262.
Likes, R. S., 1981. U.S. Patent No. 4,307,343. Washington, DC: U.S. Patent and
Trademark Office.
Lipton, M.L., Gellella, E., Lo, C., Gold, T., Ardekani, B.A., et al., 2008. Multifocal white
matter ultrastructural abnormalities in mild traumatic brain injury with cognitive
disability: a voxel-wise analysis of diffusion tensor imaging. Journal of neurotrauma 25,
1335-1342.
Lipton, M.L., Gulko, E., Zimmerman, M.E., Friedman, B.W., Kim, M., et al., 2009.
Diffusion-tensor imaging implicates prefrontal axonal injury in executive function
impairment following very mild traumatic brain injury. Radiology 252, 816-824.
141
Lipton, M.L., Kim, N., Park, Y.K., Hulkower, M.B., Gardin, T.M., et al., 2012. Robust
detection of traumatic axonal injury in individual mild traumatic brain injury patients:
intersubject variation, change over time and bidirectional changes in anisotropy. Brain
imaging and behavior 6, 329-342.
Lo, C., Shifteh, K., Gold, T., Bello, J.A., Lipton, M.L., 2009. Diffusion tensor imaging
abnormalities in patients with mild traumatic brain injury and neurocognitive impairment.
Journal of computer assisted tomography 33, 293-297.
Loane, D.J., Faden, A.I., 2010. Neuroprotection for traumatic brain injury: translational
challenges and emerging therapeutic strategies. Trends in pharmacological sciences 31,
596-604.
Lutz, A., Greischar, L.L., Rawlings, N.B., Ricard, M., Davidson, R.J., 2004. Long-term
meditators self-induce high-amplitude gamma synchrony during mental practice.
Proceedings of the National Academy of Sciences of the United States of America 101,
16369-16373.
Mac Donald, C.L., Dikranian, K., Bayly, P., Holtzman, D., Brody, D., 2007. Diffusion
tensor imaging reliably detects experimental traumatic axonal injury and indicates
approximate time of injury. The Journal of neuroscience : the official journal of the
Society for Neuroscience 27, 11869-11876.
142
Mac Donald, C.L., Johnson, A.M., Cooper, D., Nelson, E.C., Werner, N.J., et al., 2011.
Detection of blast-related traumatic brain injury in U.S. military personnel. The New
England journal of medicine 364, 2091-2100.
Maruta, J., Suh, M., Niogi, S.N., Mukherjee, P., Ghajar, J., 2010. Visual tracking
synchronization as a metric for concussion screening. The Journal of head trauma
rehabilitation 25, 293-305.
Mayer, A.R., Ling, J., Mannell, M.V., Gasparovic, C., Phillips, J.P., et al., 2010. A
prospective diffusion tensor imaging study in mild traumatic brain injury. Neurology 74,
643-650.
Mayer, A.R., Mannell, M.V., Ling, J., Elgie, R., Gasparovic, C., et al., 2009. Auditory
orienting and inhibition of return in mild traumatic brain injury: a FMRI study. Human
brain mapping 30, 4152-4166.
Mayer, A.R., Mannell, M.V., Ling, J., Gasparovic, C., Yeo, R.A., 2011. Functional
connectivity in mild traumatic brain injury. Human brain mapping 32, 1825-1835.
McAllister, T.W., Flashman, L.A., McDonald, B.C., Ferrell, R.B., Tosteson, T.D., et al.,
2011. Dopaminergic challenge with bromocriptine one month after mild traumatic brain
143
injury: altered working memory and BOLD response. The Journal of neuropsychiatry and
clinical neurosciences 23, 277-286.
McAllister, T.W., Flashman, L.A., McDonald, B.C., Saykin, A.J., 2006. Mechanisms of
working memory dysfunction after mild and moderate TBI: evidence from functional
MRI and neurogenetics. Journal of neurotrauma 23, 1450-1467.
McAllister, T.W., Saykin, A.J., Flashman, L.A., Sparling, M.B., Johnson, S.C., et al.,
1999. Brain activation during working memory 1 month after mild traumatic brain injury:
a functional MRI study. Neurology 53, 1300-1308.
McAllister, T.W., Sparling, M.B., Flashman, L.A., Guerin, S.J., Mamourian, A.C., et al.,
2001. Differential working memory load effects after mild traumatic brain injury.
NeuroImage 14, 1004-1012.
McCrea, M., American Academy of Clinical Neuropsychology., 2008. Mild traumatic
brain injury and postconcussion syndrome : the new evidence base for diagnosis and
treatment. Oxford University Press, Oxford ; New York.
McCrea, M., Guskiewicz, K.M., Marshall, S.W., Barr, W., Randolph, C., et al., 2003.
Acute effects and recovery time following concussion in collegiate football players: the
144
NCAA Concussion Study. JAMA : the journal of the American Medical Association 290,
2556-2563.
McCrea, M., Kelly, J.P., Randolph, C., Cisler, R., Berger, L., 2002. Immediate
neurocognitive effects of concussion. Neurosurgery 50, 1032-1040; discussion 1040-
1032.
McCullagh, S., Oucherlony, D., Protzner, A., Blair, N., Feinstein, A., 2001. Prediction of
neuropsychiatric outcome following mild trauma brain injury: an examination of the
Glasgow Coma Scale. Brain injury : [BI] 15, 489-497.
Messe, A., Caplain, S., Paradot, G., Garrigue, D., Mineo, J.F., et al., 2011. Diffusion
tensor imaging and white matter lesions at the subacute stage in mild traumatic brain
injury with persistent neurobehavioral impairment. Human brain mapping 32, 999-1011.
Meythaler, J.M., Peduzzi, J.D., Eleftheriou, E., Novack, T.A., 2001. Current concepts:
diffuse axonal injury-associated traumatic brain injury. Archives of physical medicine
and rehabilitation 82, 1461-1471.
Miles, L., Grossman, R.I., Johnson, G., Babb, J.S., Diller, L., et al., 2008. Short-term DTI
predictors of cognitive dysfunction in mild traumatic brain injury. Brain injury : [BI] 22,
115-122.
145
Mitchell, T. M., Hutchinson, R., Niculescu, R. S., Pereira, F., Wang, X., Just, M., &
Newman, S., 2004. Learning to decode cognitive states from brain images. Machine
Learning, 57(1-2), 145-175.
Mittl, R.L., Grossman, R.I., Hiehle, J.F., Hurst, R.W., Kauder, D.R., et al., 1994.
Prevalence of MR evidence of diffuse axonal injury in patients with mild head injury and
normal head CT findings. AJNR. American journal of neuroradiology 15, 1583-1589.
Mori, S., Oishi, K., Jiang, H., Jiang, L., Li, X., et al., 2008. Stereotaxic white matter atlas
based on diffusion tensor imaging in an ICBM template. NeuroImage 40, 570-582.
Niogi, S.N., Mukherjee, P., 2010. Diffusion tensor imaging of mild traumatic brain
injury. The Journal of head trauma rehabilitation 25, 241-255.
Niogi, S.N., Mukherjee, P., Ghajar, J., Johnson, C., Kolster, R.A., et al., 2008a. Extent of
microstructural white matter injury in postconcussive syndrome correlates with impaired
cognitive reaction time: a 3T diffusion tensor imaging study of mild traumatic brain
injury. AJNR. American journal of neuroradiology 29, 967-973.
146
Niogi, S.N., Mukherjee, P., Ghajar, J., Johnson, C.E., Kolster, R., et al., 2008b. Structural
dissociation of attentional control and memory in adults with and without mild traumatic
brain injury. Brain : a journal of neurology 131, 3209-3221.
Paniak, C., Reynolds, S., Toller-Lobe, G., Melnyk, A., Nagy, J., et al., 2002. A
longitudinal study of the relationship between financial compensation and symptoms
after treated mild traumatic brain injury. Journal of clinical and experimental
neuropsychology 24, 187-193.
Papageorgiou, T.D., Lisinski, J.M., McHenry, M.A., White, J.P., LaConte, S.M., 2013.
Brain-computer interfaces increase whole-brain signal to noise. Proceedings of the
National Academy of Sciences of the United States of America 110, 13630-13635.
Parikh, S., Koch, M., Narayan, R.K., 2007. Traumatic brain injury. International
anesthesiology clinics 45, 119-135.
Povlishock, J.T., Erb, D.E., Astruc, J., 1992. Axonal response to traumatic brain injury:
reactive axonal change, deafferentation, and neuroplasticity. Journal of neurotrauma 9
Suppl 1, S189-200.
Prabhu, S.P., 2011. The role of neuroimaging in sport-related concussion. Clinics in
sports medicine 30, 103-114, ix.
147
Ptito, A., Chen, J.K., Johnston, K.M., 2007. Contributions of functional magnetic
resonance imaging (fMRI) to sport concussion evaluation. NeuroRehabilitation 22, 217-
227.
Pulsipher, D.T., Campbell, R.A., Thoma, R., King, J.H., 2011. A critical review of
neuroimaging applications in sports concussion. Current sports medicine reports 10, 14-
20.
Raichle, M.E., Snyder, A.Z., 2007. A default mode of brain function: a brief history of an
evolving idea. NeuroImage 37, 1083-1090; discussion 1097-1089.
Rao, S.M., Bandettini, P.A., Binder, J.R., Bobholz, J.A., Hammeke, T.A., et al., 1996.
Relationship between finger movement rate and functional magnetic resonance signal
change in human primary motor cortex. Journal of cerebral blood flow and metabolism :
official journal of the International Society of Cerebral Blood Flow and Metabolism 16,
1250-1254.
Riecker, A., Groschel, K., Ackermann, H., Steinbrink, C., Witte, O., et al., 2006.
Functional significance of age-related differences in motor activation patterns.
NeuroImage 32, 1345-1354.
148
Rohling, M.L., Binder, L.M., Demakis, G.J., Larrabee, G.J., Ploetz, D.M., et al., 2011. A
meta-analysis of neuropsychological outcome after mild traumatic brain injury: re-
analyses and reconsiderations of Binder et al. (1997), Frencham et al. (2005), and Pertab
et al. (2009). The Clinical neuropsychologist 25, 608-623.
Rosales-Lagarde, A., Armony, J.L., Del Rio-Portilla, Y., Trejo-Martinez, D., Conde, R.,
et al., 2012. Enhanced emotional reactivity after selective REM sleep deprivation in
humans: an fMRI study. Frontiers in behavioral neuroscience 6, 25.
Rosenbaum, S.B., Lipton, M.L., 2012. Embracing chaos: the scope and importance of
clinical and pathological heterogeneity in mTBI. Brain imaging and behavior 6, 255-282.
Rounis, E., Lee, L., Siebner, H.R., Rowe, J.B., Friston, K.J., et al., 2005. Frequency
specific changes in regional cerebral blood flow and motor system connectivity following
rTMS to the primary motor cortex. NeuroImage 26, 164-176.
Ruff, R.M., 2011. Mild traumatic brain injury and neural recovery: rethinking the debate.
NeuroRehabilitation 28, 167-180.
Saatman, K.E., Duhaime, A.C., Bullock, R., Maas, A.I., Valadka, A., et al., 2008.
Classification of traumatic brain injury for targeted therapies. Journal of neurotrauma 25,
719-738.
149
Sadato, N., Ibanez, V., Deiber, M.P., Campbell, G., Leonardo, M., et al., 1996.
Frequency-dependent changes of regional cerebral blood flow during finger movements.
Journal of cerebral blood flow and metabolism : official journal of the International
Society of Cerebral Blood Flow and Metabolism 16, 23-33.
Satz, P.S., Alfano, M.S., Light, R.F., Morgenstern, H.F., Zaucha, K.F., et al., 1999.
Persistent Post-Concussive Syndrome: A proposed methodology and literature review to
determine the effects, if any, of mild head and other bodily injury. Journal of clinical and
experimental neuropsychology 21, 620-628.
Shenton, M.E., Hamoda, H.M., Schneiderman, J.S., Bouix, S., Pasternak, O., et al., 2012.
A review of magnetic resonance imaging and diffusion tensor imaging findings in mild
traumatic brain injury. Brain imaging and behavior 6, 137-192.
Sheu, L.K., Jennings, J.R., Gianaros, P.J., 2012. Test-retest reliability of an fMRI
paradigm for studies of cardiovascular reactivity. Psychophysiology 49, 873-884.
Silver, C.H., 2000. Ecological validity of neuropsychological assessment in childhood
traumatic brain injury. The Journal of head trauma rehabilitation 15, 973-988.
150
Slobounov, S.M., Zhang, K., Pennell, D., Ray, W., Johnson, B., et al., 2010. Functional
abnormalities in normally appearing athletes following mild traumatic brain injury: a
functional MRI study. Experimental brain research. Experimentelle Hirnforschung.
Experimentation cerebrale 202, 341-354.
Smith, S.M., Jenkinson, M., Woolrich, M.W., Beckmann, C.F., Behrens, T.E., et al.,
2004. Advances in functional and structural MR image analysis and implementation as
FSL. NeuroImage 23 Suppl 1, S208-219.
Smith-Seemiller, L., Fow, N.R., Kant, R., Franzen, M.D., 2003. Presence of post-
concussion syndrome symptoms in patients with chronic pain vs mild traumatic brain
injury. Brain injury : [BI] 17, 199-206.
Smits, M., Houston, G.C., Dippel, D.W., Wielopolski, P.A., Vernooij, M.W., et al., 2011.
Microstructural brain injury in post-concussion syndrome after minor head injury.
Neuroradiology 53, 553-563.
Stein, S.C., Fabbri, A., Servadei, F., 2008. Routine serial computed tomographic scans in
mild traumatic brain injury: when are they cost-effective? The Journal of trauma 65, 66-
72.
151
Talairach, J., & Szikla, G., 1967. Atlas of stereotaxic anatomy of the telencephalon.
Masson.
Teasdale, G., Jennett, B., 1974. Assessment of coma and impaired consciousness. A
practical scale. Lancet 2, 81-84.
Tellier, A., Marshall, S.C., Wilson, K.G., Smith, A., Perugini, M., et al., 2009. The
heterogeneity of mild traumatic brain injury: Where do we stand? Brain injury : [BI] 23,
879-887.
Turkeltaub, P.E., Eickhoff, S.B., Laird, A.R., Fox, M., Wiener, M., et al., 2012.
Minimizing within-experiment and within-group effects in Activation Likelihood
Estimation meta-analyses. Human brain mapping 33, 1-13.
Vapnik, V., & Lerner, A. J., 1963. Generalized portrait method for pattern recognition.
Automation and Remote Control, 24(6), 774-780.
Voller, B., Benke, T., Benedetto, K., Schnider, P., Auff, E., et al., 1999.
Neuropsychological, MRI and EEG findings after very mild traumatic brain injury. Brain
injury : [BI] 13, 821-827.
152
Wilde, E.A., McCauley, S.R., Hunter, J.V., Bigler, E.D., Chu, Z., et al., 2008. Diffusion
tensor imaging of acute mild traumatic brain injury in adolescents. Neurology 70, 948-
955.
Williams, D.H., Levin, H.S., Eisenberg, H.M., 1990. Mild head injury classification.
Neurosurgery 27, 422-428.
Witt, S.T., Laird, A.R., Meyerand, M.E., 2008. Functional neuroimaging correlates of
finger-tapping task variations: an ALE meta-analysis. NeuroImage 42, 343-356.
Witt, S.T., Lovejoy, D.W., Pearlson, G.D., Stevens, M.C., 2010. Decreased prefrontal
cortex activity in mild traumatic brain injury during performance of an auditory oddball
task. Brain imaging and behavior 4, 232-247.
Wu, T.C., Wilde, E.A., Bigler, E.D., Yallampalli, R., McCauley, S.R., et al., 2010.
Evaluating the relationship between memory functioning and cingulum bundles in acute
mild traumatic brain injury using diffusion tensor imaging. Journal of neurotrauma 27,
303-307.
Xiong, Y., Gu, Q., Peterson, P.L., Muizelaar, J.P., Lee, C.P., 1997. Mitochondrial
dysfunction and calcium perturbation induced by traumatic brain injury. Journal of
neurotrauma 14, 23-34.
153
Yallampalli, R., Wilde, E.A., Bigler, E.D., McCauley, S.R., Hanten, G., et al., 2010.
Acute White Matter Differences in the Fornix Following Mild Traumatic Brain Injury
Using Diffusion Tensor Imaging. Journal of neuroimaging : official journal of the
American Society of Neuroimaging.
Westbrook, C., & Roth, C. K., 2013. MRI in Practice. John Wiley & Sons.
Wurster, C.D., Graf, H., Ackermann, H., Groth, K., Kassubek, J., et al., 2014. Neural
correlates of rate-dependent finger-tapping in Parkinson's disease. Brain structure &
function, 1-12.