University of Southern California Dissertations and Theses

Diffusion tensor tractography: visualization and quantitation with applications to Alzheimer disease and traumatic brain injury

(USC Thesis Other)

Diffusion tensor tractography: visualization and quantitation with applications to Alzheimer disease and traumatic brain injury

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content DIFFUSION TENSOR TRACTOGRAPHY:
VISUALIZATION AND QUANTITATION WITH APPLICATIONS TO
ALZHEIMER DISEASE AND TRAUMATIC BRAIN INJURY
by
Darryl Hwa Hwang
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BIOMEDICAL ENGINEERING)
December 2012
Copyright 2012 Darryl Hwa Hwang
To my mentor, Manbir Singh, Ph.D.
ii
Acknowledgments
I owe the completion of this thesis to the guidance of my committee members, help form
my lab mates, and support from my family and friends.
I would like to give my deepest thanks to my late mentor, advisor, and committee
chair, Manbir Singh, whose radiant intellect and endless patience changed my life. I
would like to thank David D’Argenio, who took on the mantle of chair of my defense,
for providing further guidance towards the completion of my degree. Many thanks to
Walter Wolf and Jesse Yen for their continued support and advisement as members of
my defense committee. I would like to express my gratitude to Helena Chui and Meng
Law, my clinical advisors, who helped direct my work towards meaningful applications.
Special thanks goes to my technical advisor Natasha Lepore who was instrumental in
fine tuning this dissertation and providing much needed counsel on the future.
Countless thanks to my lab mates with which I spent so many of my days, especially
Sungheon Kim, Witaya Sungkarat, Amrita Rajagopalan, Aarti Shetty, Jeongwon Kim,
Chi-Wah “Alec” Wong, Bryce Wilkins, Sinchai Tsao, and Niharika Gajawalli. Without
your genius this dissertation could not exist.
To my family and friends, thank you for your unwavering support as I pursued my
doctorate. Thank you from the bottom of my heart for enduring this long process with
me. Your support and motivation has made it possible for me to reach this milestone.
iii
Table of Contents
Dedication ii
Acknowledgments iii
List of Figures viii
List of Tables xiii
Abstract xiv
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Specific Aims of This Dissertation 6
2.1 Visualization of Tractography . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Quantitation of Tractography . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Non-linear Co-registration of Data . . . . . . . . . . . . . . . . 7
2.2.2 Seed Distribution . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.3 Tract Normalization . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.4 Tractography Tools . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Human Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Background 9
3.1 Diffusion MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 Calculating Tensors . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 DTI Value Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.1 Fractional Anisotropy . . . . . . . . . . . . . . . . . . . . . . 13
3.2.2 Mean Diffusivity . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Eddy Current Correction . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Streamline Tractography . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.1 FACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4.2 Tensor Interpolated Tractography . . . . . . . . . . . . . . . . 15
iv
3.4.3 ICA Tractography . . . . . . . . . . . . . . . . . . . . . . . . 16
4 Visualization of Tractography 18
4.1 Software Feature Comparison . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Tractography Visualization Software . . . . . . . . . . . . . . . . . . . 18
4.3 Tract Render . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Anatomical Shell Rendering . . . . . . . . . . . . . . . . . . . . . . . 25
4.4.1 Anatomy Preprocessing . . . . . . . . . . . . . . . . . . . . . 25
4.4.2 TractRender Rendering . . . . . . . . . . . . . . . . . . . . . . 26
4.5 Tractography Display . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.5.1 NeuroTract . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.5.2 TractRender Rendering . . . . . . . . . . . . . . . . . . . . . . 27
4.6 ROI Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.6.1 ROI Selector . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.6.2 TractRender Rendering . . . . . . . . . . . . . . . . . . . . . . 29
4.7 Anatomical Background Rendering . . . . . . . . . . . . . . . . . . . 29
4.8 Variable Resolution Output . . . . . . . . . . . . . . . . . . . . . . . . 30
4.9 Video Scripting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.10 Software Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Non-linear Registration of MR Images 34
5.1 Statistical Parametric Mapping . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Normalization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2.1 FA Template Normalization . . . . . . . . . . . . . . . . . . . 37
5.2.2 DARTEL Template Normalization . . . . . . . . . . . . . . . . 38
6 Seed Distribution 41
6.1 Template Based Seed Distribution . . . . . . . . . . . . . . . . . . . . 41
6.2 ICV Based Seed Distribution . . . . . . . . . . . . . . . . . . . . . . . 42
7 Tractography Normalization 45
7.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8 Tractography Metrics and Software Toolkit 50
8.1 Tract Count Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.2 Tract-length Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.3 Tract Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.4 Tract Editing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
9 Equal-time Protocol Evaluation 55
9.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
9.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
9.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
v
9.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 63
9.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
9.6 Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
10 Application 1: Traumatic Brain Injury 69
10.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
10.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
10.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
10.3.1 Patients and Control Subjects . . . . . . . . . . . . . . . . . . 77
10.3.2 MRI Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
10.3.3 DTI Data Processing . . . . . . . . . . . . . . . . . . . . . . . 81
10.3.4 Detection and Quantification of Anisotropy Changes . . . . . . 82
10.3.5 Seed Placement and Tractography . . . . . . . . . . . . . . . . 84
10.3.6 Detection and Quantification of Affected Tracts . . . . . . . . . 85
10.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
10.4.1 Individual TBI Subjects . . . . . . . . . . . . . . . . . . . . . 94
10.4.2 Group Study: Control vs. TBI Group . . . . . . . . . . . . . . 98
10.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
10.6 Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
11 Application 2: Alzheimer Disease 106
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
11.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
11.2.1 Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
11.2.2 MR Scanners . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
11.2.3 MR Image Data . . . . . . . . . . . . . . . . . . . . . . . . . . 108
11.3 DTT Difference in AD Populations . . . . . . . . . . . . . . . . . . . . 109
11.3.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 109
11.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
11.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
11.4 Pathway Isolation Using Objective FA Difference ROIs . . . . . . . . . 112
11.4.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 112
11.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
11.5 Fornix Tract Count Using Objective Hippocampus ROI . . . . . . . . . 116
11.5.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 116
11.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
11.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
11.6 Identification of Damaged Tracts Using Improved Normalization . . . . 118
11.6.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 119
11.6.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
11.6.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
vi
11.7 Isolation of Damage Along the Fornix and Cingulum Tracts . . . . . . . 120
11.7.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11.7.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11.7.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
12 Conclusion and Future Work 124
12.1 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 124
12.1.1 USC Biomedical Imaging Lab DTI Software Suite . . . . . . . 124
12.1.2 DTT Normalization . . . . . . . . . . . . . . . . . . . . . . . . 125
12.1.3 DTI Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
12.1.4 Human Studies . . . . . . . . . . . . . . . . . . . . . . . . . . 125
12.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Reference List 128
vii
List of Figures
3.1 Graphical display of diffusion gradient directions distributed on a sphere. 11
3.2 3D rendering of ideal diffusion tensors where
1
,
2
, and
3
are the
eigenvalues and
1
,
1
, and
3
are the eigenvectors corresponding to the
eigenvalues. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.1 Screenshots of various tractography visualization software. DTI-Query
[Ope12] (top left); DTIStudio [RSD12] (top right); medInria [Asc12]
(middle left); Track Vis [WW12] (middle right); 3D Slicer [3DS12]
(bottom left); Fiber Navigator [LNN12] (bottom right). . . . . . . . . . 20
4.2 TractRender: Matlab-based GUI driven tractography render program. . 23
4.3 Examples of rendered brain volumes from TractRender with and with-
out lighting options. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.4 Examples of seed and tractography rendering from TractRender. . . . . 28
4.5 Example of brain rendered with anatomical slices. . . . . . . . . . . . . 30
4.6 Example of high resolution image rendered for the cover for Keck Medicine
Magazine (Winter 2010) . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1 Workflow for DARTEL template normalization. Dark gray sections
completed in FreeSurfer [SZE98]. Light gray sections completed in
SPM. Sections outlined in dotted lines generated by custom Matlab code. 39
6.1 Seed distribution of slice 34 of 69. . . . . . . . . . . . . . . . . . . . . 42
7.1 Tractography of slice 34 of 69. . . . . . . . . . . . . . . . . . . . . . . 47
7.2 Novel tract warping method. The portion enclosed in the dotted line was
written in C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.1 ROI-based tract editing algorithm. The portion enclosed in the dotted
line written in C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
9.1 Fractional Anisotropy map of a 28-slice brain volume . . . . . . . . . . 62
9.2 Tractography from data obtained using 25 gradient directions. . . . . . 64
9.3 Tractography from data obtained using 6 gradient directions. . . . . . . 65
viii
9.4 Track-length histograms for 25 gradient direction and 6 gradient direc-
tion data sets (left) and difference between the two histograms (25 - 6)
(right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9.5 Tract-length histograms (left) and difference between histogram (25 gra-
dients - 6 gradients) (right) for second subject. . . . . . . . . . . . . . . 66
9.6 Tract-length histograms (left) and difference between histogram (25 gra-
dients - 6 gradients) (right) for third subject. . . . . . . . . . . . . . . . 67
9.7 Tract-length histogram (left) and difference (right) between histograms
for data obtained with an 8-element coil and a quadrature coil. . . . . . 67
10.1 Block diagram of the procedure to detect FA changes between an indi-
vidual TBI subject and a group of controls following normalization of
all FA images to an FA template in MNI space. Thet map of the FA dif-
ferences in MNI space was inverse mapped to the TBI subject’s space
and thresholded to generate regions of interest (ROIs) showing signifi-
cant FA changes due to injury. . . . . . . . . . . . . . . . . . . . . . . 78
10.2 Superposition of fronto-occipital tracts from 10 control subjects, obtained
after mapping whole-brain tracts from all control subjects onto the head
of a TBI subject and sorting in the TBI subject’s head space using a
common set of frontal and occipital regions as filters. (Left) Axial view;
(right) sagittal view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
10.3 The procedure used to normalize and sort tracts in each TBI subject’s
space. (Top row) Seeds from the standard MNI space were first dis-
tributed within each control and TBI subject’s head using inverse nor-
malization to maintain the same number of seeds and anatomical equiv-
alency of seed distribution in each subject. (Left to right blocks) Whole-
brain tractography was conducted in each subject and all tracts from all
control subjects were first mapped to MNI space and then mapped onto
the head space of individual TBI subjects. ROIs were identified in the
TBI subject’s head using the procedure outlined in Fig. 10.1. Individ-
ual ROIs were then extracted one by one from the FA-difference map
between the TBI subject and controls, and used to sort tracts. An exam-
ple of a particular ROI (green) is shown at the bottom of the second col-
umn. Tracts were sorted in the TBI subject’s space using this ROI from
all control subjects and the TBI subject, and the mean number of tracts
from the control subjects was compared to that from the TBI subject to
generate the TC metric for this ROI. The sorted and superimposed tracts
for this ROI from two normal subjects (in red and purple, respectively)
and the TBI subject (blue) are shown at the bottom of the third column. 80
ix
10.4 Comparison of previous tensor reorientation-based normalized tractog-
raphy (left, in blue) to our normalized approach where all subject space
tracts are individually mapped to normalized (MNI space) using point-
to-point transformation (right, in red). Sorting of both sets of normalized
tractography was conducted with identical ROIs, located in the frontal
and occipital areas (shown in green) in template space. The tracts gener-
ated with the previous voxel-based normalization and reoriented tensor
method tend to suffer from discontinuity toward the ends. Compared
to our approach, the previous method also appears to increase the con-
founds of partial volume effects (arrows indicate areas of tract disconti-
nuity and redirection around the posterior right corner of the ventricle)
due to the interpolation inherent to voxel-based normalization. . . . . . 81
10.5 FA-reduced regions corresponding to t3.0 (see color bar) superim-
posed on the FLAIR images of TBI Subject 1. An extent thresholdk12
was also used to identify clusters. Some FA-reduced regions overlap
completely, others partially and some do not overlap with the FLAIR
spots. (In this montage, the left hemisphere L appears at the right in
each image.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
10.6 Three pathways, hippocampal/fornix (HC/FX), inferior fronto-occipital
(IFO) and inferior longitudinal fasciculus (ILF), identified as crossing
the voxels of the highest t-score ROI in a TBI subject (Subject 1). (Top
panel) Coronal, sagittal and axial views of the ROI with color coding
of the t-score as indicated in the color bar. (Middle row) The three
pathways (HC/FX—red, IFO—magenta, ILF—black) in a normal sub-
ject shown in axial and sagittal views. (Bottom row) The same three
pathways (HC/FX—blue, IFO—magenta, ILF—black) in a TBI sub-
ject. (The yellow and green colors in the middle and bottom row pic-
tures were used to identify sub-ROIs within the ROI as described in the
text.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
10.7 Tract count-based sensitivity (effect size) to differentiate TBI subjects
from controls as a function of FA. . . . . . . . . . . . . . . . . . . . . 89
10.8 Similar to Fig. 10.6, pathways associated with two other ROIs within
TBI Subject 1 are shown in (A) and (B), respectively. Coronal, sagittal
and axial views of the two ROIs with color coding of the t-score as
indicated in the color bar are shown in the top row. The ROI in (A)
identified tracts through the posterior portion of the corpus callosum,
whereas (B) identified the right HC/FX tracts. Like Fig. 10.6, tracts
for a normal subject are shown in red and those in the TBI subject are
shown in blue. The reduction of tract counts in the TBI subject with
respect to the control subject is obvious for these ROIs. . . . . . . . . . 92
x
10.9 Similar to Fig. 10.5, the FA-reduced regions for TBI Subject 2 super-
imposed on the FLAIR images of the subject. (In this montage, the left
hemisphere L appears at the right in each image.) . . . . . . . . . . . . 94
10.10Similar to Fig. 10.6, pathways associated with two ROIs in TBI Subject
2 are shown in (A) and (B), respectively. The ROI in (A) identified the
left HC/FX tracts, whereas (B) identified the right HC/FX tracts. Like
Fig. 10.6, tracts for a normal subject are shown in red and those in the
TBI subject are shown in blue. The reduction of tract counts in the TBI
subject with respect to the control subject is obvious in these ROIs. . . . 96
10.11A superposition of the FA-reduced regions (in color) detected by an
SPM-based statistical comparison of the TBI group (n=12) to the nor-
mal group (n=10) at P
FDR
.05, k12 on the FA template in MNI
space. Arrows point to the three ROIs (genu of the corpus callosum,
hippocampal region, splenium of the corpus callosum) used in Fig. 10.12. 99
10.12Similar to Fig. 10.6, pathways associated with three ROIs identified as
FA-reduced regions in the group study between 10 controls and 12 TBI
subjects are shown in (A), (B), and (C), respectively. The ROI in the
middle row of (A) identified the left IFO, HC/FX and a portion of the
ILF for a normal control subject (IFO—purple; ILF—black, HC/FX—
red); (B) identified the anterior portion of the corpus callosum (red) and
(C) identified the posterior portion of the corpus callosum (red). The
bottom row shows corresponding tracts identified by the same ROIs in
a TBI subject. (Bottom row, A) IFO—purple; ILF—black; HC/FX—
blue. (Bottom row, B) Anterior corpus callosum—blue. (Bottom row,
C) Posterior corpus callosum—blue. The reduction of tract counts in the
TBI subject with respect to the control subject is obvious in these ROIs. 101
11.1 Fronto-Occipital ROIs. The ROIs are used pair-wise (shown by use of
different colors) to isolate tracts first in the left hemisphere, then in the
right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
11.2 Fronto-Occipital Tracts. The images to the left are from a normal control
and the right are from a probable AD. . . . . . . . . . . . . . . . . . . 111
11.3 Two value t-test results for FA where subject groups had lower FA than
the normal control group. . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.4 Intersection between normal/probable AD and normal/MCI populations. 114
11.5 Tracts isolated from ROIs . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.6 Study trend charts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
11.7 Progressive degeneration of the fornix tract. . . . . . . . . . . . . . . . 118
11.8 Two value t-test results for FA where subject groups had lower FA than
the normal control group. . . . . . . . . . . . . . . . . . . . . . . . . . 119
xi
11.9 N>MCI: Fornix (top row) and cingulum (bottom row) FA t-Score Com-
parison. Colors denote the t-score difference between the mean FA of
the various populations (blue=0, red=3", line marks p<0.005 on the
color scale). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
11.10MCI>AD: Fornix (top row) and cingulum (bottom row) FA t-Score
Comparison. Colors denote the t-score difference between the mean
FA of the various populations (blue=0, red=3", line marksp<0.005 on
the color scale). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
xii
List of Tables
4.1 Feature matrix of various tractography visualization programs. Actively
Updated was judged via the last update posted for the software; any
software with consistent version updates within the last year is consid-
ered Actively Updated. Brain Shell refers to the cortical surface render-
ing (see Chapter 4.4); all programs can render tracts as tubes. Lighting
Control refers to the ability of a program to change the position and
intensity of lighting sources during rendering. Opacity Control refers
to the ability to adjust transparency of various objects during rendering.
Video Scripting is the creation of an automated camera path for video
generation. Multi-Resolution refers to the ability for the program to dis-
play images of differing resolutions all registered to the same real world
space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
9.1 Tract-lengths for 25 vs. 6 gradients. . . . . . . . . . . . . . . . . . . . 66
9.2 Tract-lengths for 8-element vs. quadrature coil. . . . . . . . . . . . . . 67
10.1 List of affected pathways and diffusion metrics for 12 ROIs identified
by FA reduction in a TBI subject (Subject 1) compared to a group of 10
normal controls (t3.00,k12) . . . . . . . . . . . . . . . . . . . . . 91
10.2 List of affected pathways and diffusion metrics for 11 ROIs identified
by FA reduction in a TBI subject (Subject 2) compared to a group of 10
normal controls (t3.0,k12) . . . . . . . . . . . . . . . . . . . . . . 98
10.3 List of affected pathways and diffusion metrics for 12 ROIs identified
by SPM group analysis as FA-reduced regions in the TBI group (n=12)
compared to the normal control group (n=10) (P
FDR
.05,k12) . . . 102
11.1 Fronto-Occipital tract count for 8 normal control and 7 AD subjects . . 110
xiii
Abstract
With the advent of diffusion tensor imaging (DTI) came insight into the organization of
the most complex organic computer in existence—the human brain. Diffusion Tensor
Tractography (DTT) introduced the ability to visualize, in vivo, axonal fiber bundles, the
brain’s internal wiring structures.
Rendering tractography in three dimensions aids in the understanding of how the
axonal connections of the brain are organized, and is an important tool in illustrating the
complex geometry of fiber bundles. To better facilitate the use of 3D visualization for
tractography, we wrote flexible custom software targeted at researchers.
The use of tractography need not be limited to visualization; quantitation allows for
tractography to be used for clinical applications. In order to create objective metrics of
tractography for group analysis, we have created new algorithms to register diffusion
data to a single space for comparison, introduced new metrics such as tract count and
tract-length histograms to quantify tractography, and developed methods to properly
distribute seed points for tractography, a necessary step for quantitation. Finally, we
have provided the ability to conduct tractography in the original, and most accurate,
acquisition space, and then bring the tracts into a common space for comparison.
Validation of theses metrics and techniques required application to conditions that
affect axonal integrity. We examined data from subjects with Alzheimer Disease (AD)
and traumatic brain injury (TBI), two afflictions believed to compromise axons of the
xiv
brain. These studies and others indicate that our understanding of both conditions can
be greatly enhanced by the application of DTI and DTT.
xv
Chapter 1
Introduction
Diffusion Tensor Imaging (DTI) is currently being used by researchers and clinicians
to help gain insight into many neurological diseases. DTI has the singular ability to
image the physical structure and organization of the brain’s white matter in vivo. As
an extension of DTI, Diffusion Tensor Tractography (DTT) allows researchers to infer
three dimensional connections; this leads to the need for visualization.
Several programs have been implemented for DTT. Traditionally, clinicians are
trained extensively in interpreting two dimensional images; the relatively new addition
of three dimensional data provided by these programs requires new methods of visual-
ization.
DTI has many derived quantitative metrics, but there is a distinct lack of quan-
titation with DTT. For tractography to become more clinically relevant and to progress
from being used solely for visualization, objective metrics must be derived which take
into account connectivity, a feature inherent in DTT but lacking in DTI. We have intro-
duced several metrics for DTT as will be described here.
Two neurological conditions currently under investigation in the USC Biomed-
ical Imaging Lab are Alzheimer Disease (AD) and traumatic brain injury (TBI). Both
conditions provide unique challenges which complicate objective analysis of DTI.
This dissertation proposes a new visualization tool and novel methods for quan-
titation of DTT studies. Included are discussions about the application of quantitative
measurements to study the effect of Alzheimer Disease and Traumatic Brain Injury.
1
1.1 Motivation
DTT is currently an underutilized method in clinical medicine. Lack of standardized
display and relegation to subjective visual analysis has prevented DTT from becoming
standard protocol in clinical applications.
We seek to provide the ability to visualize tractography data with complete visual
control for the user, created on a platform which allows quick modifications to accom-
modate new visualization needs.
In addition to providing ease of display, we seek to quantize DTT by provid-
ing new metrics and methods of comparison. The ultimate goal is the use of DTT for
objective quantifiable measures of diffusion MRI scans.
According to the Alzheimer’s Association, an estimated 5.4 million Americans
will be afflicted with Alzheimer Disease in 2012 [Ass12]. AD is a fatal degenerative
neurological condition characterized by neuron death, brain atrophy, and formation of
beta-amyloid plaques and tau protein tangles. Many of the physical changes in the brain
occur in the white matter and involve the targeted destruction of specific pathways.
Traumatic brain injuries have recently been the subject of renewed scientific
interest. The United States of America and its Allied Forces has seen the return of
many veterans, from conflict areas such as Afghanistan and Iraq, who have suffered
injuries by improvised explosive devices (IED) [HMT
+
08]. Due to advances in pro-
tective equipment, many soldiers survive the initial explosion but suffer from TBI. In
athletes, TBI has also been a topic of debate. Concussive injuries suffered by athletes
in contact sports, such as football, have been linked with long term neurological effects
[GMB
+
05]. Current methods of prognosis for recovery often yield expectations that do
not match final outcome.
DTT offers the ability to visualize damage to the axonal fibers of the brain and
provides insight into which neuronal pathways are affected, the degree of damage, and
2
the prognosis for recovery. DTT can be an invaluable tool that has the potential to aid
diagnosis, track disease progression, and detect the changes influenced by treatment.
1.2 Outline
Chapter 2 is a summary of the specific aims for this dissertation. An overview of DTI
technology and standard data processing methods starts in Chapter 3. This chapter will
address the theoretical basis for DTI, derivation of standard DTI metrics, clinical con-
siderations for acquisition, and confounds found during image acquisitions and recon-
struction. The chapter then continues with an explanation of streamline tractography
methods.
Chapter 4 addresses the need for and creation of custom 3D rendering software
to efficiently display tractography. Initially there was a lack of software on the market
to display these images. This led to our development of the 3D “blue brain” rendered
images, which have now become our standard presentation method. This chapter also
discusses the need for and uses of rendering and video generation software.
Chapter 5 discusses the need for non-linear co-registration (normalization) of
images for group comparisons. As different populations need different normalization
techniques, the challenges faced when attempting to normalize images from aged sub-
jects with brain atrophy are different from those where the subjects have youthful brains.
Two normalization methods are presented for discussion.
Chapter 6 focuses on the need to control seed distributions for streamline trac-
tography in order to obtain quantitative results. Two methods are discussed, template
and intracranial volume normalized seed distributions.
Chapter 7 examines our novel tract normalization technique. Tractography nor-
malization is typically accomplished by warping the individual tensors, including tensor
3
reorientation after spatial shifting, followed by streamline tractography using the new
diffusion tensor field. Our method streamlines with process; adjusting the tensors into
a new space is no longer necessary. This increases accuracy because tensor recalcula-
tion is not needed, and allows our tractography to be calculated using non-interpolated
acquisition space data.
Chapter 8 details the creation of several important algorithms central to DTI and
DTT processing. This includes discussion on the utilization of tract count and tract-
length histograms metrics to measure the relative quality of different acquisitions, tract
filtering, and tract editing methods.
Chapter 9 examines the use of tract count and tract-length histograms for the
objective ranking of equal time-length protocols having different parameters. As clin-
ical MR research is often limited by the total amount of time a subject can remain in
a scanner, this method objectively determine optimal scanning parameters given time
constrains.
Chapter 10 examines the application of our methods to a TBI data set. Tech-
niques presented in previous chapters are used to illuminate changes found in TBI
patients.
Chapter 11 chronicles our original research in AD, while taking part in an on-
going collaborative research project. Data for this study was acquired throughout the
course of the project. Many techniques have been developed and refined to examine
changes observed in the brain from the progression of AD.
Chapter 12 overviews future work, such as the continuation of development in
multiple tensors DTI. An inherent limitation of the traditional DTI model is the inability
to adapt to multiple fibers in a single voxel, known as the partial volume effect. This
incorrect modeling of diffusion leads to erroneous calculation of the diffusion tensor,
and subsequent miscalculations of DTI metrics. In contrast, we use ICA tractography to
4
determine multiple fiber orientations, which is our starting point to fit multiple tensors
to the diffusion data.
5
Chapter 2
Specific Aims of This Dissertation
The purpose of this dissertation is to improve quantitation and visualization of diffusion
tensor tractography. Included is a summary of novel methods developed for tractogra-
phy. These methods are then applied to the study of two specific neurological conditions.
2.1 Visualization of Tractography
Visualization of tractography data is a non-standardized process with no consensus on
file format for tractography data. There exist many different programs which display
DTI tractography, but all were created with their own custom formats in mind. We cre-
ated TractRender in Matlab
r
as a companion program to our own tractography pipeline
and format, which allows for the rendering of a transparent brain volume and the display
of tracts in three-dimensional space. A GUI for control over surface opacity, lighting,
color, and camera rotation allows custom displays and animations to be made quickly
and efficiently.
2.2 Quantitation of Tractography
Tractography has generally been used to qualitatively evaluate diffusion tensor imaging
by showing possible connections between different parts of the brain. Our goal was to
use tractography to derive several objective measures for the integrity of the white matter
6
tracts in the brain. To achieve the goal of quantitation, we developed the techniques
outlined in this subsection.
2.2.1 Non-linear Co-registration of Data
Non-linear co-registration is the warping of data to create a common space where com-
parisons can be done. The need for normalization is further exacerbated by age-related
changes seen in many subjects. We created new workflows to register different popu-
lations into a standard space by creating custom templates and multi-stage image co-
registration pipelines. These processes are applied to the study of Alzheimer Disease
and Traumatic Brain Injury.
2.2.2 Seed Distribution
For any quantitation of tractography to be meaningful, the starting point must be the
same for all subjects in a study. Tractography begins with seed point distribution. Tradi-
tionally, seeds are distributed evenly in the subject space; however this causes the seeds
to be unevenly distributed across subjects with differing brains. We propose two differ-
ent approaches to seed distribution in this dissertation. The first method calculates seed
placement from template warping parameters via non-linear co-registration; the second
method utilizes the intracranial volume as a normalizing parameter.
2.2.3 Tract Normalization
DTI Tractography is normally conducted in the original, acquisition space of the data.
In many cases, tractography would be better compared and displayed in other spaces,
i.e. a template space. Current methods involve warping tensors and then conducting
tractography with the warped tensors; we created a method which allows for the tracts
7
to be generated using the most accurate acquisition space data and warp the tracts to the
template space.
2.2.4 Tractography Tools
In order to achieve quantitation, novel metrics needed to be derived. We offer the met-
rics of tract count and tract histogram as measures of tractography integrity. Whole
brain tractography routinely yields tens of thousands of tracts. In order to display only
selected tracts, we have created tract filtering and tract editing algorithms. Tract filtering
is the process in which tracts are selected using a region of interest (ROI). Tract editing
involves changing the overall length of a tract via factors such as FA threshold and stop
masks.
2.3 Human Data Processing
Processing human data using the methods described in the previous subsections aims
to provide both validation and utility for these methods in future clinical work. Human
data studies include human phantom scans, TBI studies, and AD studies.
8
Chapter 3
Background
In this chapter, we endeavor to introduce the theoretical background behind DTI and
DTT. We begin with diffusion MRI and the calculation of diffusion tensors. We con-
tinue with the derivation of the two most commonly used DTI metrics. We touch upon
imaging considerations due to eddy current artifacts. We conclude with a brief overview
of streamline tractography techniques.
3.1 Diffusion MRI
Diffusion magnetic resonance imaging is a relatively new addition to clinical MRI. The
effect of water diffusion in a magnetic field was described by Stejskal and Tanner [Ste65]
1965, but was not applied to human scanning until the 1990s. Stejskal and Tanner
formulated the signal attenuation for diffusion imaging in Equation 3.1 [Ste65, Tan78].
ln

S
S(0)

=
2
DG
2

2

3

(3.1)
whereS is the signal with the diffusion gradient applied,S
0
is the signal without gradi-
ent,
is the gyromagnetic ratio,D is the effective diffusion constant,G is the strength
gradient pulse, is the time between pulses, and is the duration of the pulse. If we
make the substitution where

2
G
2

2

3

=b (3.2)
9
we can use Equation 3.2 to reformulate Equation 3.1, where
ln

S
S(0)

=bD (3.3)
3.1.1 Calculating Tensors
In diffusion tensor imaging (DTI), we need multiple measurements to characterize a
three dimensional space. Equation 3.3 can be written as
ln

S(b)
S(0)

=
3
X
i=1
3
X
j=1
B
ij
D
ij
(3.4)
In clinical DTI sequences, B is a numerical value (b) entered into the scanner in the
course of specifying the parameters of the scan, combined with the gradient directions.
Diffusion is assumed to be a three-dimensional Gaussian and Basser et. al. [BMTLB93]
modeled diffusion as a rank-2 tensor, allowing us to represent three dimensional diffu-
sion as a positive, diagonalizable, symmetric 3x3 matrix representing D. Incorporating
the vectorx;y;z representing the direction of the diffusion gradient, Equation 3.4 can
be rewritten as
ln

S(b)
S(0)

=b
2
6
6
6
4
x
2
xy xz
xy y
2
yz
xz yz z
2
3
7
7
7
5
2
6
6
6
4
D
xx
D
xy
D
xz
D
xy
D
yy
D
yz
D
xz
D
yz
D
zz
3
7
7
7
5
(3.5)
10
Equation 3.5 can be expanded to
2
6
6
6
6
6
6
6
4
ln
h
S(1)
S(0)
i
1
ln
h
S(2)
S(0)
i
2
.
.
.
ln
h
S(n)
S(0)
i
n
3
7
7
7
7
7
7
7
5
=b
2
6
6
6
6
6
6
6
4
x
2
1
y
2
1
z
2
1
2x
1
y
1
2y
1
z
1
2x
1
z
1
x
2
2
y
2
2
z
2
2
2x
2
y
2
2y
2
z
2
2x
2
z
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
x
2
n
y
2
n
z
2
n
2x
n
y
n
2y
n
z
n
2x
n
z
n
3
7
7
7
7
7
7
7
5
2
6
6
6
6
6
6
6
6
6
6
6
6
6
4
D
xx
D
yy
D
zz
D
xy
D
yz
D
xz
3
7
7
7
7
7
7
7
7
7
7
7
7
7
5
(3.6)
The minimum number of required gradient directions n is 6, as determined from the 6
unknowns (D) in Equation 3.6. In general, we use 25 gradient directions to better sample
the three dimensional space.
(a) 6 diffusion gradient directions. (b) 25 diffusion gradient directions.
Figure 3.1: Graphical display of diffusion gradient directions distributed on a sphere.
11
3.2 DTI Value Maps
A rank-2 diffusion tensor can be fully characterized by its three eigenvalues () and
their corresponding eigenvectors () as seen in Equation 3.7.
D =
T
(3.7)
Equation 3.7 can be further expanded to Equation 3.8.
D =
2
6
6
6
4

1
1

2
1

3
1

1
2

2
2

3
2

1
3

2
3

3
3
3
7
7
7
5
2
6
6
6
4

1
0 0
0
2
0
0 0
3
3
7
7
7
5
2
6
6
6
4

1
1

1
2

1
3

2
1

2
2

2
3

3
1

3
2

3
3
3
7
7
7
5
(3.8)
From the eigenvalue and eigenvectors, a variety maps can be generated to quantify DTI.
Below, we focus on the two metrics most commonly used in DTI research; Fractional
Anisotropy and Mean Diffusivity.
(a) Anisotropic tensor. (b) Isotropic tensor.
Figure 3.2: 3D rendering of ideal diffusion tensors where
1
,
2
, and
3
are the eigen-
values and
1
,
1
, and
3
are the eigenvectors corresponding to the eigenvalues.
12
3.2.1 Fractional Anisotropy
The most widely used tensor derived metric in DTI research is fractional anisotropy
(FA). FA is used as a clinically used measurement of voxel-wise diffusion anisotropy in
DTI, and is also used as a stop threshold value in standard streamline tractography. FA
is mathematically defined in Equation 3.9.
FA =
s
3((
1
)
2
+ (
2
)
2
+ (
3
)
2
)
2(
2
1
+
2
2
+
2
3
)
(3.9)
where
1
,
2
, and
3
are the eigenvalues of the tensor and is the average of the eigen-
values. FA is normalized between 1 and 0, with 0 being completely isotropic (sphere)
and 1 being purely anisotropic (line).
A decrease in FA is believed to be linked to loss of white matter integrity
[MKD
+
02]. New studies have shown that an increase in FA might also denote white
matter damage in crossing areas [DJB
+
11].
3.2.2 Mean Diffusivity
Mean Diffusivity (MD) is another DTI value map that others have claimed to be more
sensitive in detecting axonal damage [ACWPN09, ACWPN10]. MD is defined as the
mean of the eigenvalues (see Equation 3.10):
MD =

1
+
2
+
3
3
(3.10)
where
1
,
2
, and
3
are the eigenvalues of the tensor. MD can be equated with diffusion
restriction. A high MD would appear in unrestricted materials such as cerebrospinal
fluid (CSF); a low MD would be present in areas with barriers to free diffusion, such as
white matter. Unlike FA, MD is unbounded.
13
3.3 Eddy Current Correction
Because of the powerful diffusion gradients used, eddy currents are a concern. Eddy
current artifacts manifest as distortions in the MR image. Others have compensated
for these artifacts by way of specialized pulse sequences [ATP97, PMP
+
00, RHWW03,
TCS08] or image correction using scanner obtained distortion maps [JBP98, Hor99,
CPGC99, Bas99, BA00, Bas01, PSBM05, CGS06]. For these studies eddy current
effects are already partially compensated for by the secondary application of an equal
and opposite diffusion gradient as part of the dual spin echo acquisition sequence.
Further correction was performed through our post-processing workflow when co-
registering the diffusion images with b=0. The method of correction chosen was
FMRIB’s Diffusion Toolbox (FDT), a freely distributed software written by the Anal-
ysis Group of the Oxford Centre for Functional MRI of the Brain (FMRIB). FDT
includes the function eddy correct, which corrects for the eddy current artifacts
by co-registering the images with a diffusion gradient applied to the b=0 images.
3.4 Streamline Tractography
The process of constructing connections from DTI data is known as tractography.
Though several different tractography methods have been created, both by our group
and others, the field is typically separated into two types of processes: deterministic and
probabilistic. Deterministic tractography is generated using point by point connection
through the diffusion tensor field; in contrast, probabilistic tractography is generated
from likelihood calculations based on the uncertainty of tensor fit [BWJ
+
03, PHWK03].
The focus of this section is streamline tractography, a type of deterministic tractogra-
phy. Below is a short overview of some of the most popular streamline tractography
algorithms.
14
3.4.1 FACT
One of the first tractography algorithms proposed was Fiber Assignment by Continuous
Tracking (FACT) [MCCVZ99]. FACT is a tractography algorithm which requires no
tensor interpolation; the algorithm works by having seeds generated at the center of
a voxel. From there, the direction of the tract is held constant in the direction of the
primary eigenvector until the tract exits the voxel. The tract then progresses in the
direction specified by the primary eigenvector of the new voxel. This process is repeated
until the tract reaches a point where the nearby eigenvectors are calculated to be too
dissimilar.
3.4.2 Tensor Interpolated Tractography
One limitation of FACT is the need for seeds to begin in the exact center of voxels.
According to Basser et. al. [BPP
+
00], white matter tracts could be represented as an arc
in three dimensions:
dr(s)
ds
="
1
(r(s)) (3.11)
where r(s) is a parameterized vector, s is the arc length, and"
1
is the principal eigenvec-
tor. Equation 3.11 can be solved using a variety of numerical methods. Euler’s method
is one of the simplest and can be applied to solve the differential equation yielding
Equation 3.12:
r(s
x+1
) = r(s
x
) +h"
1
(r(s
x
)) (3.12)
where the step size h is defined by the user.
Tractography can begin at seed point with interpolated directions from the clos-
est voxel principal eigenvector and the direction 180° from the principal eigenvector.
This allows for the tract to propagate in both directions. This also forces the seed point
to routinely lie in the middle of the tract rather than the ends. The primary eigenvector
15
is calculated from the interpolated tensor of the point. Tractography continues until one
of three criteria occur:
1. The FA of the point is lower than the set threshold.
2. The angle of deflection between steps is larger than the set threshold.
3. The angle of deflection of points in one centimeter is larger than the set threshold.
3.4.3 ICA Tractography
There are many multiple fiber tractography methods available, and many more in devel-
opment. Most involve the use of probabilistic tractography or scans that are beyond the
ability of current clinical MR scanners. For the scope of this dissertation, we focus on
our internally created tractography methodology.
Independent Component Analysis (ICA) tractography is an algorithm developed
in our lab, the USC Biomedical Imaging Lab [SW07, SW09, SW10]. Our method views
the diffusion data from each voxel as a mixture of source signals. We use ICA to separate
the data into one, two, and three sources. An f -test is conducted to choose the number of
sources per voxel. Once the sources are separated, the primary eigenvector is calculated
and used as the principal vector for each source.
Tractography is conducted using the same step-wise Euler numerical solution
as interpolated tractography (see Section 3.4.2); however, instead of interpolating the
tensor, the algorithm interpolates the principal vector. When approaching a voxel with
multiple vectors, our algorithm choses the vector which is closest to the previous vector
used.
Seed distribution is forced to be at the center of each voxel, with a seed point for
each vector. V oxels with multiple fibers are seeded multiple times, so there is a seed for
each fiber direction.
16
This method was created to reduce the partial volume/crossing fiber effects.
When a voxel contains multiple fibers the traditional tensor formulation fails, as it can
only provide one primary eigenvector. ICA tractography circumvents this problem, but
is an incomplete solution to this problem (see Chapter 12.2).
17
Chapter 4
Visualization of Tractography
Display of tractographical data is currently an open area of research. While different
groups have displayed tractography in varying custom displays, most agree that three
dimensional rendering is necessary to properly illustrate the complex geometries of the
axonal bundles. The level of complexity used for rendering varies from group to group.
MR slice display, anatomic structure rendering, light element placement, object trans-
parency, tract coloring, and surface shading are all options available in different software
packages; however, these packages offer differing amounts of control to the end user.
4.1 Software Feature Comparison
Each tractography visualization program provides a unique set of tools and capabilities.
The non-standardization of tractography has allowed for many different programs to
be developed on many different platforms; to highlight key differences between the
seven reviewed visualization software programs and TractRender, a feature comparison
is given in Table 4.1.
4.2 Tractography Visualization Software
Currently, there is not a standard display method in DTI tractography. While a variety
of software is available to visualize tractography, many options require the data to be
18
Features DTI-Query
DTI Studio
medInria
Track Vis
3D Slicer
FiberNavigator
Dipy
TractRender
Actively Updated X X X X X "
Windows X X X X X X X "
MacOS X X X X X X "
Linux X X X X X "
Brain Shell X X X "
Rendered Tracts X X X X X X X "
Lighting Control X "
Opacity Control X X X X X "
Video Scripting "
Multi-Resolution "
Table 4.1: Feature matrix of various tractography visualization programs. Actively
Updated was judged via the last update posted for the software; any software with con-
sistent version updates within the last year is considered Actively Updated. Brain Shell
refers to the cortical surface rendering (see Chapter 4.4); all programs can render tracts
as tubes. Lighting Control refers to the ability of a program to change the position and
intensity of lighting sources during rendering. Opacity Control refers to the ability to
adjust transparency of various objects during rendering. Video Scripting is the creation
of an automated camera path for video generation. Multi-Resolution refers to the ability
for the program to display images of differing resolutions all registered to the same real
world space.
pre-processed using their own specific methods of computing tensor information and
tractography. Below is a brief review of tractography visualization software.
DTI-Query [ASM
+
04, SAM
+
05] (http://http://graphics.
stanford.edu/projects/dti/software/) is designed to work with
AFNI [Cox96] formatted DTI data to calculate and display tractography. Released by
Stanford University, this software is available for Windows, MacOS, and Linux. In
addition to generation of tractography, the tracts are rendered as tube, and orthonormal
slices can be displayed as background. This software also has the ability to create
19
Figure 4.1: Screenshots of various tractography visualization software. DTI-Query
[Ope12] (top left); DTIStudio [RSD12] (top right); medInria [Asc12] (middle left);
Track Vis [WW12] (middle right); 3D Slicer [3DS12] (bottom left); Fiber Navigator
[LNN12] (bottom right).
volume of interest (VOI) filters to selectively display tracts. Initially released in
August of 2005 (updated most recently in November 2007), this software comes with a
20
graphical user interface (GUI) giving the user control over the tractography algorithm,
tract filtering abilities, and background anatomy display.
DTI Studio [JvZK
+
06] is part of the MRI Studio Software application suite
(http://mristudio.org). Developed for Windows, this software is presented
as a turnkey solution for diffusion tensor tractography. All DTI and tractography are
directly calculated by the program. DTI Studio calculates its tractography via either
Fiber Assignment by Continuous Tracking (FACT) or Diffusion Probability (DP), and
has built-in ROI-based tract filtering. This software allows for orthonormal background
slice display with rendered 3D tracts. DTI Studio was last updated to version 1.1 in May
2007.
medInria [FTP06] (http://med.inria.fr) is a medical visualization
software released for Windows, MacOS, and Linux. Developed and actively supported
by Inria teams (Asclepios, Athena, Parietal, and Visages), medInria 2.0.1 was released in
April 2012. medInria calculates the tractography using a custom algorithm, and has tract
filtering via ROI. The software displays orthonormal slices and renders tracts, which can
be color coded by tensor orientation.
Track Vis [WBSW07] is a visualization software for both DTI and DSI devel-
oped at Massachusetts General Hospital. Coded in C++, Track Vis uses Qt for the
graphical user interface (GUI) and Visualization Toolkit (VTK) for the visualization.
The first public version was released in May 2007, and the software was last updated
was in November 2011. Compiled for Windows, MacOS, and Linux, this software can
display orthonormal co-registered MRI slices as background and offers ROI-based tract
filtering. Tracts can be rendered as lines or tubes, and opacity for different 3D rendered
objects can be controlled.
21
3D Slicer [PHK04] is a medical imaging software suite first released in 2002;
this software includes tractography capabilities starting with version 2 in 2005. Com-
piled for Windows, MacOS, and Linux, the latest version of the software, known as
Slicer 4 (release 4.1.0) was published in April 2012. Slicer 4 is a multi-modality imag-
ing suite designed for the general rendering of three dimensional images. It has the
capability to render anatomy and generate brain shells. The software framework cal-
culates tensors and conducts streamline tractography, tracts can be rendered as lines or
tubes. In addition, opacity for different 3D rendered objects can be controlled.
FiberNavigator (http://code.google.com/p/fibernavigator/)
was created by a team from the Universit´ e de Sherbrooke (Qu´ ebec, Canada) under the
direction of Maxime Descoteaux, Ph.D. and is available for Windows and MacOS. Com-
piled in C++ and utilizing Open GL for 3D rendering, GLSL, and wxWidgets for GUI,
this program has the capability to render anatomy and generate brain shells. Tracts
can be rendered as lines or tubes, and opacity for different 3D rendered objects can be
controlled.
Dipy [GBA
+
11] is the newest entrant in tractography visualization software.
Released in 2011, Dipy is part of the Neuroimaging in Python (NIPY) community. Cur-
rently, Dipy is released as Python source code with instructions on how to compile in
Windows, MacOS, and Linux. It is a collaborative software development platform, and
not a polished distributed program.
While all of the tractography programs presented above allow for display of
tracts, most of these programs are inflexible and require data to be processed starting
from the acquired diffusion MR images. This limits the user’s ability to modify DTI
calculation and tractography methods. In addition, many of these programs are com-
piled programs, originally designed to be all-encompassing solutions; this makes them
22
difficult to adapt to new imaging ideas. We created Tract Render as a platform to gener-
ate rendered images and to develop new visualization methods.
4.3 Tract Render
Figure 4.2: TractRender: Matlab-based GUI driven tractography render program.
TractRender is a GUI driven program which began development in 2004, predat-
ing all of the currently available tract visualization software (see Figure 4.4). Developed
in Matlab, the software provides researchers with a level of customized automation and
image flexibility that other compiled program lack. TractRender is designed to work
23
with the custom tractography file format initially designed by Sungheon Kim, Ph.D.;
the format has subsequently been extended by multiple members of the USC Biomed-
ical Imaging Lab. The complete rendering process can be separated into six distinct
display sections which will be elaborated upon further in the following sections.
1. Anatomical Shell Rendering: This is the segmented brain shell which gives the
viewer the overall shape of the brain and aids in determining three dimensional
orientation. Generally it is rendered transparently to help 3D visualization of the
tracts inside the brain volume.
2. Tractography Display: Tracts can be shown in methods ranging from simple
line depictions of the tracts to complex tubular rendering of the data which allow
for lighting, shading and transparencies. The seed points of the tracts can also be
displayed.
3. ROI Visualization: ROIs are defined on a voxel basis, and are rendered as a
convex hull with inner surfaces removed.
4. Anatomical Background Rendering: Often, it is desirable to render slices of
anatomy to show details in the brain volume. Tract Render’s background render-
ing inherently accepts images of differing resolution to be displayed in the same
brain volume.
5. Variable Resolution Output: Rendered images can be produced at varying res-
olutions to match intended application.
6. Video Scripting: Single frame images are generated on a virtual camera, which
moves around the three dimensional rendering on user defined trajectories.
TractRender (see Figure 4.2) is designed to only display pre-calculated data. All
data must be preprocessed and formatted by other software (such as NeuroTract).
24
4.4 Anatomical Shell Rendering
4.4.1 Anatomy Preprocessing
The brain volume used in TractRender begins as b=0 co-registered anatomy images. In
most cases, the b=0 volume from the DTI data set or the T1-weighted anatomy set can
be used as the starting point. The brain volume is then segmented into gray matter,
white matter, and CSF. Only the gray and white matter volumes are used for display.
The volumes may still include non-brain portions, which can require user input editing.
The current MNI registered brain shell comes from a segmentation of the SPM
T1 single subject MNI template. TractRender provides the flexibility to display brain
shells from any reference frame as long it is registered to the tractography.
(a) Brain volume rendered with flat colors. (b) Brain volume rendered with lighting options.
Figure 4.3: Examples of rendered brain volumes from TractRender with and without
lighting options.
25
4.4.2 TractRender Rendering
The brain anatomic shell is constructed using a surface comprised of all points of a
similar value in a three dimensional volume. As an optional pre-processing step, the user
can smooth the data prior to surface generation. As a default, chosen for aesthetics, the
brain is rendered as a transparent blue color surface (see Figure 4.3a). More perceived
three-dimensionality can be obtained by controlling surface lighting and shading (see
Figure 4.3b).
4.5 Tractography Display
4.5.1 NeuroTract
Tractography generation is conducted in NeuroTract. NeuroTract is the current gener-
ation of the NTrack software created by Sungheon Kim, Ph.D. for use in our lab. In
NeuroTract, we completely renovated both the GUI and software architecture, employ-
ing a more object-oriented approach, which allows increased flexibilty when adding
new processing options, and integrated ICA tractography (see Section 3.4.3) into the
capabilities of NeuroTract.
NeuroTract generates a.mat file which consists of the following variables:
• tract—Stores the tracts first by seed point slice, then by seed point. The seed
point is in voxel dimension, however, the tract is stored in millimeter space.
• tractp—Stores the parameters of the tractography, including the following
fields:
– sf—Stores the voxel size in millimeters of x, y, and z of the acquired DTI
image.
26
– h—Stores the Euler step size in millimeters.
– fa—Minimum fractional anisotropy value used as a stop criteria.
– mintract—Minimum length of tracts measured in millimeters.
• img mode—Indicates whether the.mat file contains tractography data.
• vol dim—Stores the dimension of the image volume in x, y, and z voxels.
4.5.2 TractRender Rendering
Tract information can be displayed either as seed points or complete tracts. Since seeds
are stored in voxel space, they are rendered directly in TractRender’s three dimensional
space. Tracts are converted from millimeter space to voxel space before rendering.
Rendering is available in two varieties: point and surface. Point rendering allows
for relatively quick display. The seed points are marked by dots (see Figure 4.4a) while
tracts are displayed as lines (see Figure 4.4b). In this mode, color is the only property
which can be adjusted for seeds and tracts. Surface rendering is much more compu-
tationally intensive; seed points are calculated as spherical surfaces (see Figure 4.4c),
which become noticeable when zoomed in, and tracts are calculated as tubular surfaces
(see Figure 4.4d). The number of faces can be adjusted to speed up computation time.
Once a surface is created, color, opacity, and lighting can be manipulated. Using sur-
faces allows the user to add three-dimensional shading, providing additional visual cues
for tract and seed orientation.
27
(a) Simple point-based seed rendering. (b) Simple line-based tract rendering.
(c) Sphere-based tract rendering. (d) Tube-based tract rendering.
Figure 4.4: Examples of seed and tractography rendering from TractRender.
4.6 ROI Visualization
4.6.1 ROI Selector
ROIs are binary or thresholded grayscale images stored in the Analyze 7.5 or NIFTI
formats. ROIs can be generated in a variety of ways, including SPM and MRIcro;
however, for easy and uniform ROI file generation, ROI Selector was created for Matlab.
28
The program allows for the loading of anatomy files in Analyze 7.5 or NIFTI formats,
and then creates and displays ROIs in three 2-dimensional slice views (axial, coronal,
and sagittal).
ROI Selector generates drawings of ROIs in any of the slice views. When
changes are made in the selected pane, they are automatically made in the other two
views. Multiple ROIs can be displayed simultaneously in different colors/opacities, a
vital feature in checking the relative positions of ROIs.
4.6.2 TractRender Rendering
All ROIs are rendered as the outer shell of a series of stacked rectangular prisms. Elim-
ination of the inner surfaces of the ROI allows for unobstructed views through the ROI
when transparency is selected. In addition to handling binary ROI files, TractRender
accepts Analyze 7.5 and NIFTI formatted files that have a range of values, and gener-
ates ROIs based on threshold values. As with all surface data in TractRender, color,
opacity, and lighting options are user-adjustable. Additionally, included is an option to
render the ROI as a smoothed volume to soften the hard edges of orthonormal faces.
4.7 Anatomical Background Rendering
Often, the addition of an anatomical slice benefits the viewer’s understanding of the
three dimensional tractography (see Figure 4.5). To achieve this effect, the built-in Mat-
labslice function is used, allowing for a flat surface to be created in any orientation.
Typically, the anatomy is only needed in the orthogonal views; once the surface is cre-
ated, opacity and colormap control can be adjusted for optimal display.
In addition to tractography space anatomy images, TractRender has the ability to
display anatomical background images which are of differing resolution or orientation,
29
but are co-registered to the tractography space, without the need for reslicing the volume.
This procedure allows for lower resolution generated tractography to be displayed on
higher resolution anatomy images.
Figure 4.5: Example of brain rendered with anatomical slices.
4.8 Variable Resolution Output
Image rendering is a computationally intensive process. Part of the added complexity
is the resolution at which an image is finally rendered; images are generated for dif-
ferent purposes varying from relatively low-resolution images generated for slide pre-
sentations and the web, to high-resolution print images for publication (see Figure 4.6).
TractRender is capable of generating variable resolution images tailored to the specific
30
needs of the user. For final image manipulation, we recommend using an image editing
software such as Photoshop.
Figure 4.6: Example of high resolution image rendered for the cover for Keck Medicine
Magazine (Winter 2010)
31
4.9 Video Scripting
The programs reviewed in Section 4.2 are all geared towards user interaction as the
final display method. Users are expected to view the rendered image in the software,
and rotate it using sliders or mouse clicks. The few options that do have auto rotation
capabilities still require the software to be running during the duration of rotation.
Researchers often need to present information to an audience in a clear and con-
cise manner; videos allow researchers to prepare information for public consumption
prior to the presentation. Additionally, in recent years the Internet has made the distri-
bution of digital videos more available which has led to an even wider dissemination of
research.
Video generation is presented as a complement to user direct control methods.
Clinicians and researchers alike want the freedom to view tractography data in all man-
ners of custom ways. Tract Render provides its users the opportunity to present their
data in an easy to understand, streamlined format that is much easier to disseminate
than the software currently on the market. It is the difference between watching a movie
and playing a video game; both have stories, narratives, and can be visually stunning, but
movies have an added benefit in that they can be presented without a human operator.
TractRender provides researchers with the ability to generate the images needed
to create videos. This subroutine renders images while steadily moving both the camera
and lighting sources around the virtual three dimensional space. The camera follows a
set path determined in advance by the user.
Camera paths are designed to highlight the features of tractography in three
dimensions by changing the view perspective. The standard path we created that is
seen in most of the currently generated videos begins with the camera overhead, look-
ing down on the axial view of the superior side of the brain. Our DTI scans are acquired
as axial slice images. Beginning the video with an overhead view helps orient viewers,
32
who are accustomed to axial images, to the three dimensional display. The camera then
rotates down and around an oblique view of the right posterior side of the brain. At a
constant 20° from horizontal, the camera makes a complete 360° orbit around the brain.
Once this orbit is complete, the camera is rotated back to the original overhead view.
This circular path provides the option of a smooth looping video if the user so desires.
TractRender was created with customization in mind. All objects are individu-
ally accessible, and the user can adjust all properties during the image rendering routine.
This allows the researcher complete control over the canvas and the end result.
Any program which can generate videos from sequential images can be used
to construct the video file. We use Adobe Premiere; this program allows for input of
sequential images as frames for a movie, and has various transition effects which allow
the seamless joining of different image sequences. Premiere affords the ability to pro-
duce movies in different formats and optimization levels subject to the requirements of
the presentation. As most videos are created for integration into Microsoft PowerPoint
presentations or web-based viewing, the two most commonly used video container files
are the Windows Media Video (.wmv) format and the Apple Quicktime (.mov) format.
4.10 Software Distribution
It is the intention of the author to release this software for use under GNU General Public
License as part of an entire Tractography Suite which includes NeuroTract (our GUI
driven DTI and tractography calculation program), ROISelector (our region of interest
marking software), and TractRender in early 2013.
33
Chapter 5
Non-linear Registration of MR Images
One of the many challenges of group study work with DTI is the need for a common
reference frame. Human subjects are unique; no two persons have identical scans. To
confound the matter further, brains change in both size and shape in the normal course
of aging. To achieve meaningful results, we must warp subject data to a common space.
DTI registration, in essence, is the calculation of a transformation from one space
to another space. Registration appears to be a straight-forward process of warping DTI
data to a common space, but many problems exist which prevent proper execution. One
issue is the lack of definitive DTI atlases. Brain atlases have been developed as rep-
resentative averages; however, general application of atlases is problematic due to the
fact that they are only suitable for a specific population. Though work has begun on
templates based on DTI metrics [ZYRG09], no true standard has emerged. Due to this
lack of standardization, it is imperative that we create custom templates for our work.
The methodology of obtaining and applying the deformation transform changes
with different populations. The atlases most used by our work are the templates devel-
oped by the Montreal Neurological Institute (MNI). The MNI templates were created
from normal subjects with an average age in their 20’s; this population of “youthful”
brains varies greatly from that of Alzheimer Disease research, where the population is
typically between the ages of 60 and 80. Aged brains suffer from structural changes due
to atrophy, which prevents accurate warping.
To overcome this difficulty, we developed methods of non-linearly registering
DTI data to the MNI template space by using custom intermediate templates. These
34
methods allow us to generate custom warping transforms, giving us the ability to bring
all data to a common space for analysis. To calculate the non-linear warping parameters,
we used Statistical Parametric Mapping [AF97, Ash07].
5.1 Statistical Parametric Mapping
Statistical Parametric Mapping (SPM) is a freely distributed Matlab program developed
by the Wellcome Trust Centre for Neuroimaging based at the University College Lon-
don. SPM has undergone many updates since it was first introduced. Four versions are
currently available directly from the Wellcome Trust Centre for Neuroimaging: SPM99,
SPM2, SPM5, and the most recent release, SPM8. This software is one of the standard
research tools for functional magnetic resonance imaging (fMRI), positron emission
tomography (PET), single photon emission computed tomography (SPECT), electroen-
cephalogram (EEG), magnetoencephalogram (MEG), structural imaging, and diffusion
tensor imaging (DTI). A key component of SPM’s features is the warping of images to
template spaces. The developers of SPM are continuously working to refine the accu-
racy of their algorithms [AF97, Ash07]. For this dissertation, SPM2 was used during
the initial development. During the course of our research, our lab transitioned to SPM8
both for batch processing abilities and the DARTEL toolkit. SPM8’s DARTEL was a
new registration algorithm that improved on the functionality found in SPM2.
Spatial Normalization: SPM uses the term “normalization” to refer to non-
linear image co-registration and “co-registration” for linear affine image co-registration
methods. This dissertation adheres to this naming convention.
35
SPM2 has a template normalization function that takes an image in subject space
and non-linearly warps the image to template space [AF99], and allows the use of cus-
tom templates in this process. SPM normalization uses twelve-parameter affine trans-
formation followed by a nonlinear warping transformation to fit source images to the
specified template. SPM uses a linear combination of discrete cosine transform (DCT)
basis function, iteratively optimized to minimize the bending energies of the deforma-
tion fields, to model the non-linear warping.
SPM8 DARTEL Toolbox: The Diffeomorphic Anatomical Registration
Through Exponentiated Lie Algebra (DARTEL) [Ash07] Toolbox found in SPM8 is a
new normalization suite which shows marked improvement over the SPM2/SPM8 stan-
dard Spatial Normalization method. At this time, it has not been fully integrated into
SPM8. DARTEL registration compares favorably with other normalization methods
[KAA
+
09] and integrates well into our Matlab workflow.
Brain Segmentation: SPM2 has an automated segmentation process which sep-
arates the white matter, gray matter, and cerebral spinal fluid [AF05], and requires the
user to give a contrast type for the image. SPM8 advanced the SPM2 segmentation not
only by utilizing more advance segmentation techniques, but also by autodetecting the
type of image inputted by the user.
The b=0 EPI images are automatically segmented by SPM, where the white
and grey matter segments are used for the anatomic rendering detailed in section 4.4.
The gray matter, white matter, and cerebrospinal fluid (CSF) segmentations from SPGR
volumes are summed up to calculate intracranial volume (ICV), which is used for the
ICV seed distribution outlined in Section 6.2. Summation of thresholded segmentations
yields skull-stripped versions of T
1
.
36
In some cases, the segmentation is made more accurate by eliminating the
anatomy outside to the brain, and manual realignment of the image to correct for abnor-
mal head tilt.
5.2 Normalization Methods
We have developed new workflows for diffusion data normalization. The commonality
between both workflows is the creation of custom templates for the subject populations.
The two primary classifications of subjects we work with are young and aged popula-
tions; the major difference between these two populations is the presence of atrophy in
the aged populations, which can prevent accurate normalization. Below are two different
pipelines designed to work with the differing subject pools.
5.2.1 FA Template Normalization
The FA Template Normalization workflow is based on the creation of a custom FA atlas.
The process begins with selecting an appropriate number of age matched control DTI
scans. Using the b=0 volume, each data set is segmented using SPM. SPM segmentation
generates per-voxel percentages of white matter, grey matter, and CSF. White matter
files for each subject are created by thresholding voxels with a >50% probability of
being white matter. These images are co-registered using 12-parameter affine/non-linear
transformation to the MNI white matter template. The transform is then applied to
the b=0 images, and the transformed images are normalized to the MNI EPI template.
FA images are calculated for each subject, followed by the combined linear and non-
linear transform being applied to each subject. With all the FA images existing in the
same MNI template space, the values are averaged to form the final FA template, which
37
appears rather smooth. Internal tests have shown that normalization works poorly with
sharp templates; better registration is achieved with smoother templates.
After an age-matched FA template is created, FA maps are generated for all
subjects. The FA maps of each subject are non-linearly normalized to the FA template.
The resulting deformation file is used as the transform from individual acquisition space
to MNI template space.
5.2.2 DARTEL Template Normalization
When the study population has significant amounts of distortion, such as the atrophy
effects of aged populations, FA template normalization does not correctly normalize the
data. To compensate for these effects, a more flexible algorithm for normalization is
required. DARTEL was shown to be much more accurate than SPM’s standard normal-
ization algorithm (SPM is currently in the process of making DARTEL their standard
normalization algorithm). DARTEL has also been shown to work well with distortion
[KAA
+
09], placing highly when compared to other techniques.
Using DARTEL requires the creation of a custom template. We used SPGR
images from a number of normal control subjects and segmented them using the “New
Segmentation” function of SPM [AF05], creating the volumes used in the DARTEL
template generation function. The resulting template is warped to MNI space and the
warping deformations are saved.
Our DARTEL normalization workflow, shown in Figure 5.1, utilizes a multi-
step method to generate a final deformation. B=0 images are co-registered using 12-
parameter affine/non-linear transformation to SPGR images acquired during the same
scanning session. The co-registration transform is saved, and the SPGR images are
38
segmented into volumes, once again using the “New Segmentation” function. The seg-
mented volumes are used in the DARTEL normalization and warped to the DARTEL
template. The resulting transform is saved.
The finalized deformation from subject acquisition to MNI template space is the
concatenation of the co-registration transform, SPGR to DARTEL template deforma-
tion, and DARTEL template to MNI template deformations. The finalized deformation
is applied to our DTI metric maps (FA, MD, etc.) and is used as part of our Normalized
Tractography algorithm (see Chapter 7).
Image
Intensity
Correction
Intensity corrected
SPGR
DWI
DTI
Metrics
Calculation
FA MD
SPGR
DARTEL
Image
Normalization
DARTEL
Normalization
Transform
DARTEL to MNI
Template
Normalization
DARTEL to MNI
Template
Transform
12 Parameter
Affine
Registration
Affine
Registration
Transform
Image
Segmentation
DARTEL
Template
Generation
Images
from
Controls
Concatenated
Transform
Figure 5.1: Workflow for DARTEL template normalization. Dark gray sections com-
pleted in FreeSurfer [SZE98]. Light gray sections completed in SPM. Sections outlined
in dotted lines generated by custom Matlab code.
A similar normalization technique was published by Canu et. al. [CMF
+
10].
They removed the 12 parameter affine registration between the diffusion and SPGR
39
images. To register diffusion to T
1
-weighted images, they created a custom FA template
which was then registered to the DARTEL template. The registration of two templates
adds an extra transform compared to our method which introduces the possibility of
more computational rounding errors.
40
Chapter 6
Seed Distribution
Our next step towards quantitative tractography deals with seed distribution. Stream-
line tractography begins with the placement of the seed points, the starting points for
an individual tract in the diffusion tensor field. To conduct any kind of group statis-
tics with measures such as tract count, the number and location of the seed points for
tractography needs to be arranged in a consistent distribution throughout each person’s
brain. The methods must be adapted to the population being studied; when studying
brains with atrophy where portions of the brain no longer exist, the distribution of seeds
must be adjusted for the lack of anatomy. We have developed two approaches for seed
distribution which are targeted towards different study populations: template based, and
intracranial volume based. It is important to note that the seed distribution occurs in
acquisition space (See Chapter 7 for tract normalization).
6.1 Template Based Seed Distribution
SPM template normalization non-linearly warps the subject brain scan to the standard
space of the Montreal Neurological Institute (MNI). After normalization is conducted
using the Deformation Toolbox included in SPM, a one-to-one voxel correlation can be
generated. This correlation, in the form of a.mat file, encodes the corresponding posi-
tion in subject space for each voxel in template space. As long as SPM normalization
is accurate, this method works very well in distributing the tractographical seeds evenly
through the template space. Equal distribution in template space allows for the final
41
normalized tractography to be quantitatively compared. This method was used when
analyzing the traumatic brain injury (TBI) subjects (see section 10).
Figure 6.1 shows the seeds for the same slice in both normalized template space
(Fig. 6.1a) and subject space (Fig. 6.1b). The template space is seeded at one seed per
voxel. The subject space shows not only the smooth deformation of the distribution
process, but also the floating point nature of redistributed seeds. Template based seed
distribution can be used with populations where there is little concern for large scale
atrophy.
(a) Template space distribution. (b) Subject space distribution.
Figure 6.1: Seed distribution of slice 34 of 69.
6.2 ICV Based Seed Distribution
Alzheimer Disease presents many challenges in normalization. AD primarily affects
elderly subjects, and one of the challenges with this population is brain atrophy. Sub-
jects with AD show a greater degree of atrophy than those exhibiting normal aging.
The central conundrum is: which changes are attributed to normal aging and which
are attributed to AD. It is known that intracranial volume (ICV) is not affected by AD.
42
Using ICV , we redistribute the seed points to normalized coverage between all subjects
regardless of actual brain size. Our procedure accounts for changes for differing brain
sizes, but preserves changes due to atrophy. This method allows differences to be seen
as a measure of atrophy. To calculate the seed distribution, the ICV ratio between the
subject and the template volume is used.
R
ICV
=
ICV
S
ICV
N
(6.1)
where ICV
S
is the intracranial volume of the subject, and ICV
N
is the normal intracra-
nial volume (which can be taken as the average ICV of the normal population). The
corrected subject space volume is calculated as follows:
V
vox
S
=R
ICV
V
vox
N
(6.2)
whereV
vox
N
is the volume of the voxel in normalize space. Because we want the actual
size of each dimension, we take the cube root of the ICV ratio and multiply each indi-
vidual dimension of the normal voxel to obtain the dimensions in subject space.
x
vox
S
=
3
p
R
ICV
x
vox
N
(6.3)
y
vox
S
=
3
p
R
ICV
y
vox
N
(6.4)
z
vox
S
=
3
p
R
ICV
z
vox
N
(6.5)
The procedure scales the seed distribution as if each subject had the same ICV . This
provides a normalized seed distribution, which equalizes the effects of differing head
sizes for each subject, but keeps the seeds in a scaled grid distribution. This means
the seeds do not correspond to the same anatomic locations on each subject, but can
be used to study atrophy changes. Once the subject voxel size is determined, a three
43
dimensional grid of points is calculated, centered around (0, 0, 0) using thex
vox
S
,y
vox
S
,
andz
vox
S
values. The grid of points is then shifted so the center is the defined origin of
the volume. The final step is the removal of all points which fall outside of the mask
calculated during the tensor estimation phase in NeuroTract.
44
Chapter 7
Tractography Normalization
Many different approaches have been used to try to obtain tractography in spaces other
than the original acquisition space. Most involve a warping of the tensors to create a
newly interpolated tensor field in a template space, then conducting tractography in the
newly calculated space [PCF
+
00, JGA
+
02, XMS
+
03, KIN
+
07, MULK07, XHB
+
08].
Tensor warp methods suffer from severe interpolation artifacts. There can be no guar-
antee of continuity of tracts from pre-warping to post-warping; tracts might become
discontinuous, or new tracts could be formed by inaccurate interpolation.
Our novel tract warping procedure eliminates the need for tensor warping. All
tractography is calculated in individual subject acquisition space. Since tractography
is a series of points which are connected sequentially, each tract is treated as a simple
collection of points in the original data acquisition space. The tracts are then warped
point-wise into the template space, preserving continuity without the effects of data
interpolation (see Figure 7.1). Tract continuity is preserved as long as the registration
method used generates a smooth deformation field.
The mapping procedure begins either with a FA template-based normalization
(see Section 5.2.1) or the multi-stepped DARTEL Template Normalization (see Section
5.2.2). The end result of both methods is the calculation of a deformation field, which is
stored as a matrix. The acquisition space coordinates (A) and template space voxel (T )
are related to each other by a deformation transformation (M),
A =M(T ) (7.1)
45
Effectively, the deformation transformM is only known for the pointsA. The inverse
of Equation 7.1 also holds true, but only forT points in template space.
T =M
1
(A) (7.2)
T is only known whenA is a point which was calculated in the deformation file. The
known points in data acquisition space A are not defined rigidly on a grid, so sim-
ple trilinear interpolation cannot be conducted. The deformation fields are generally
smooth when generated by the affine and non-linear normalization procedures of SPM.
Smoothness is a result of calculating the warping as a linear combination of discrete
cosine transform basis functions. This results in the warping parameters varying only
slightly from one point to an adjacent point. If the point of interest is not one of the
known points A , that point, call it A
0
, is simply a known point (A) plus a correction
factor (dA).
A
0
=A +dA (7.3)
The corresponding unknown point in template space (T
0
) can also be seen as a known
pointT plus a correction factor (dT ).
T
0
=T +dT (7.4)
To solve forT
0
we have to relatedT anddA in Equation 7.5.
J =
2
6
6
6
4
@M
x
@M
x
@M
x
@M
y
@M
x
@M
z
@M
y
@M
x
@M
y
@M
y
@M
y
@M
z
@M
z
@M
x
@M
z
@M
y
@M
z
@M
z
3
7
7
7
5
(7.5)
46
We calculate the Jacobian for each point as a pair-wise partial derivative of theM matrix.
With the Jacobian, we can calculatedT .
dT =J(dA) (7.6)
Substituting into Equation 7.4 with Equations 7.2 and 7.6
T
0
=M
1
(A) +J(dA) (7.7)
This allows us to calculate any template space pointT
0
given a floating point in acquisi-
tion space (A
0
).
(a) Tractography in acquisition space. (b) Tractography in template space.
Figure 7.1: Tractography of slice 34 of 69.
7.1 Algorithm
Using the deformation file we calculated using the methods in Chapter 5, we calculate
a Jacobian file containing a Jacobian matrix for each known voxel. The process (see
Figure 7.2) is written in C and accessed through Matlab. The program begins with the
47
conversion of all tractography to voxel coordinates. The algorithm then finds the known
point which is closest to the tract segment being converted using the Approximate Near-
est Neighbor (ANN) library for C [MA12]. If the tract point is one of the known points,
then it is converted and saved. If it is not an exact match, the difference between the
known point and tract point being examined is calculated. The difference is converted
using the Jacobian file and used as a correction factor for the conversion.
After conversion all the tracts are in voxel space, which is subsequently con-
verted back to millimeter space. This process converts all tracts which were calculated
in subject space into atlas space. The end result is normalized tractography without the
discontinuities caused by spatial warping of tensors.
48
SPM2
NeuroTrac t
Tractography
File
Forward
Deformation
File (FDF)
DTI value
map
DTI Data
Calculate
Jacobian
Jacobian
File
Seed Point
(Voxel
Coordinates )
Tract Points
(mm Coordinates)
Find known point
in FDF using ANN
Covert mm
coordinates to
voxel coordinates
Equals known
point.
Di!erence
between known
point
No
Is seed point?
Covert voxel
coordinates
to mm
coordinate s
Yes
No
Calculate
correction to
known point value
Yes
Save value to
tractography "le
Write seed and
tract point
values
Figure 7.2: Novel tract warping method. The portion enclosed in the dotted line was
written in C.
49
Chapter 8
Tractography Metrics and Software
Toolkit
To further exam DTI and DTT as discriminators of disease progress, we created quality
measures and a software toolkit to better interact with the large amounts of data. In this
chapter, we introduce two new metrics for DTT quantitation and present methodologies
to isolate and prepare tracts of interest for quantification and comparison.
8.1 Tract Count Metric
There are many DTI metrics which have been created to measure changes in diffusion
(see Chapter 3.2), but there are no metrics to measure changes in tractography. Tractog-
raphy has been viewed as a visualization technique for the diffusion tensor field. The
tract count (TC) metric was proposed as a measure of streamline tractography quality
and tract integrity. Streamline tractography is dependent on both the orientation of the
tensor and the DTI metric of FA (see Section 3.4). TC requires the maintenance of
identical tracking parameters for the tractography and a method of normalizing the seed
point distribution (see Chapter 6) for all scans being studied. TC has been used previ-
ously with simple scalar compensation [SKWC04], but our corrected TC can be seen in
more recent publications [SH05, SHSV05, HTS09, SJ09, SJH
+
10].
50
8.2 Tract-length Histogram
In order to evaluate the relative quality of DTI protocols on the same scanner, we derived
a measure to assess the quality of the scan protocol. Initially, using average FA values
for the whole brain seemed to be the logical way of determining quality, the rationale
being the protocol with the highest FA was the best; however, this failed to address
the lack of directionality of FA. Eddy current effects can create a “sorting bias,” which
cause a switching among the eigenvalues of the tensor. In addition, noise and gradient
sampling can contribute to the deviation of the primary eigenvector [BP00]. Both of
these errors result in a change of tensor orientation which may be decoupled from any
effect in the tensor shape. FA gives an accounting of only the shape of the tensor, but
not directionality.
DTT requires the use of both FA and directionality for successful tractography.
Incorrect tensors tend to divert the tracts into wrong paths in whole brain tractography,
causing the tracts to terminate prematurely. When two different scans are performed
on the same phantom (for example, when testing a larger number of gradient direc-
tions verses repeated measures of a smaller number of gradient directions), we can rate
the relative quality of the scanned images by examining the tract-length histogram (see
Chapter 9).
8.3 Tract Filtering
Often, whole brain tractography generates an overwhelming volume of data. A sorting
method is required to filter out the tracts of interest. This begins by specifying a region
of interest (ROI). Once an ROI is designated in the brain volume, tract filtering, the
binary process where a tract is either included or excluded depending on whether or not
that tract intersects the ROI, can begin. Conturo et al. [CLC
+
99] used this method to
51
isolate pathways which joined regions. Wakana et al. [WJNP
+
04, WCP
+
07] extended
this technique by adding in Boolean logic. These concepts were adapted to the custom
tractography software of the USC Biomedical Imaging Lab and coded in Matlab
r
with
Boolean logic of AND and NOT (using OR logic with ROI filtering is essentially the
joining of two ROIs prior to filtering).
Tracts can be filtered by either tract location or seed point location. Though at
first glance the tract location filtering and seed point location filtering appear similar, the
applications are very different. Tract location filtering selects a tract if any portion of the
tract lies inside the ROI. The concept of seed point filtering of whole-brain tractography
is analogous to conducting tractography solely from the seeds found in an ROI, without
the need to recompute tractography.
Tractography results vary greatly depending on whether you are tracking out of
an ROI or filtering whole brain tractography with an ROI, as tractography follows the
path of least angular change. As an example, in the case where a weak tract merges
into a strong tract, if the seed point is placed upstream on the weaker tract, tractography
would propagate along the weaker tract and continue after merging with the stronger
tract. If the seed point is placed upstream of the merge point on the stronger tract, that
tractography would propagate along the stronger tract even after the merge point. Since
tractography is a bi-directional propagation from the seed point, if the seed point is
downstream in the stronger tract, as it tracts upstream it will only find the stronger tract,
and will never deviate into the weaker tract.
8.4 Tract Editing
Tractography has a tendency to create false connections in areas of poor tensor defini-
tion. For example, the area near the edges of the ventricles often causes pronounced
52
merging of the fronto-occiptal and splenium tracts. One correction method to eliminate
false positives is to eliminate these areas before tractography, but this causes a hard edge
which, when subjected to tensor interpolation, causes further erroneous tractographical
effects. An alternative post-processing solution presented below is retroactively termi-
nating the tracts when tracts enter these predefined ROIs.
Algorithm
ROI determined tract editing requires two inputs: tracts and the ROI. The algorithm
begins by examining the seed point of each tract to determine if the tract is to be saved
or discarded outright. If the seed point is located inside the ROI, the algorithm only
saves the seed point in millimeter space and discards the tract. If the seed is not in the
ROI, it is examined further. The tract is then examined point by point.
The tract points are stored as sequential points in millimeter space from one
end of the tract to the other. The seed point is generally not the first or last point, but
rather a point located in the middle of the sequence. For ROI filtering, the tracts are
converted into voxel coordinates. After conversion, the algorithm locates the seed point,
and starting from that point, progresses sequentially to the next point, checking if that
point is in the ROI. If it is, the algorithm marks the current point as the stop point for the
tract. If not, it goes to the next point and repeats the process. If the algorithm finishes
the tract without finding a point which lands in the ROI, the last point of the tract is
marked as the stop point. When the stop point is defined, the algorithm returns to the
seed point and repeats the process going in the opposite direction to mark the start point
of the tract. Once the start and stop points both are defined, the algorithm then saves all
points in millimeter space, beginning with the start point and ending with the stop point.
This method was the central focus of our work in retrospective processing to
reduce partial volume effects [HSRS07b, HSRS07a].
53
Tractography
File
ROI
*.mat
File
Seed Point
(Voxel
Coordinates)
Tract Points
(mm Coordinates)
Covert mm
coordinates to
voxel coordinates
Save tract seeded
from ROI?
Yes
Is the next point
in the ROI?
Yes
No
Save value to
tractography !le
Write seed and
tract point
values
Is seed point in
ROI?
No
Start at the tract point
which is the seed
point. Progress in the
positive direction.
Covert voxel
coordinates
to mm
coordinate s
Yes
Progress to
the next
tract point
Is it the last
tract point?
No
Progressing in
positive direction?
Yes
Reset to
seed point
and start in
negative
direction
Yes
No
Figure 8.1: ROI-based tract editing algorithm. The portion enclosed in the dotted line
written in C.
54
Chapter 9
Equal-time Protocol Evaluation
This was presented at SPIE Medical Imaging 2005 as “Evaluation of MRI DTI-
tractography by tract-length histogram”[SHSV05]
9.1 Abstract
In the absence of ground truth, there are very few methods available to evaluate the
accuracy of a specific tracking algorithm or the various data acquisition protocols for
DTI-tractography. The objective of this work was to develop methodology, based on
tract-length histograms, that could be used to evaluate whole-brain tractography with
data acquired under different conditions for a given subject, for example six versus 25
gradient directions, or use of an 8-element phased array versus quadrature head-coil.
Whole-brain DTI data were acquired from six healthy normal human volunteers on a
1.5 T GE scanner at TR=10.3s, field-of-view 26cm, 128x128 matrix, 28 contiguous
4mm thick slices from 25 isotropic gradient directions with b=1000s/mm
2
, one b=0
acquisition, and number of excitations (NEX)=1 for a total acquisition time of 3min
53s. Similarly, four sets of data were acquired from 6 non-colinear directions and com-
bined with two b=0 acquisitions to equalize the time for 25 and 6-directions acquisitions.
The tract-length histograms clearly show that at equal acquisition time, there are more
long tracts in the 25-direction acquisition than the 6-direction acquisition, suggesting
better estimation of the tensor with 25 directions. Tract-counts above a threshold pro-
vide an objective index to evaluate tractography. Also a comparison of the two coils
55
shows a higher tract-count for long tracts with the 8-element coil, consistent with the
demonstrated higher sensitivity and higher signal-to-noise ratio for EPI acquisitions by
the 8-element coil.
Keywords: Diffusion Tensor Imaging, Tractography, MRI, Tractography Evalu-
ation
9.2 Introduction
Tractography based on MRI Diffusion Tensor Imaging (DTI) is a rapidly growing field
to map axonal connections through white matter in the human brain. Either alone, or in
conjunction with functional MRI, the axonal connections provide important new infor-
mation to decipher brain structure and function. Though the basic concepts of diffu-
sion MRI have been known since the mid 1980s [LBBL
+
86] and diffusion weighted
imaging (DWI) has been correlated to early brain ischemia in the early 90s [MCK
+
90],
the formulation of DTI in a form applicable to white matter tractography is relatively
recent [BMTLB93, BMLB94]. The directional dependence of water-molecular diffu-
sion (anisotropic diffusion) forms the basis of mapping axonal tracts in white matter. In
general, diffusion is much faster along the fiber than orthogonal to the fiber. Thus by
measuring diffusion along different directions in 3D space, it becomes possible to detect
and infer the direction of axonal tracts.
Diffusion is a tensor, implying that it has a different magnitude in different direc-
tions. The tensor can be mapped by modifying a gradient echo pulse sequence to become
sensitive to diffusion along a particular gradient direction by applying a pair of diffusion
pulses (bipolar gradients) along that direction [HBTV99]. A commonly used diffusion
sensitive pulse sequence is the Pulsed Gradient Spin Echo (PGSE) sequence developed
by Stejaskal and Tanner in 1965.
56
The MRI signal acquired with a diffusion sensitive pulse sequence will attenuate
in proportion to diffusion along the direction of the diffusion gradients. Conceptually,
the attenuation ‘A’ in the MRI signal with diffusion ‘D’ can be expressed by:
A =
S(b)
S(0)
=e
bD
(9.1)
where b is a “b-factor” that depends on the timing, amplitude and shape of the
diffusion gradients,S(0) is the signal at b=0 (thus no effect from diffusion), andS(b) is
the signal for a particular value of the b-factor.
Diffusion is conveniently modeled as an ellipsoid or rank 2 tensor (3x3 com-
ponents) requiring a minimum of six non-collinear gradient directions to estimate the
diffusion tensor matrix per voxel.
Casting diffusion as a tensor and representing it by a matrix D, the above equa-
tion becomes:
log

S(b)
S(0)

=
3
X
i=1
3
X
j=1
B
ij
D
ij
(9.2)
where B is the matrix form of the b-factor calculated from the applied magnetic
gradient pulse [x y z]:
B =b
2
6
6
6
4
x
y
z
3
7
7
7
5
h
x y z
i
=b
2
6
6
6
4
x
2
xy xz
xy y
2
yz
xz yz z
2
3
7
7
7
5
(9.3)
57
Here we assume the matrix to be symmetric about the diagonal. The diffusion
tensor D can be written as:
D =
2
6
6
6
4
D
xx
D
xy
D
xz
D
xy
D
yy
D
yz
D
xz
D
yz
D
zz
3
7
7
7
5
(9.4)
whereD
ab
=D
ba
. Taking into account the symmetric nature of B and D, we can rewrite
Eqn 9.2 as
log

S
S
o

=b
h
x
2
D
xx
+y
2
D
yy
+z
2
D
zz
+ 2xyD
xy
+ 2yzD
yz
+ 2xzD
xz
i
(9.5)
In practice, gradients are applied in n independent directions to measure the diffusion
tensor, each with its own equation 9.5, which leads to the following system of linear
equations:
2
6
6
6
6
6
6
6
4
log
h
S
1
So
i
1
log
h
S
2
So
i
2
.
.
.
log
h
Sn
So
i
n
3
7
7
7
7
7
7
7
5
=b
2
6
6
6
6
6
6
6
4
x
2
1
y
2
1
z
2
1
2x
1
y
1
2y
1
z
1
2x
1
z
1
x
2
2
y
2
2
z
2
2
2x
2
y
2
2y
2
z
2
2x
2
z
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
x
2
n
y
2
n
z
2
n
2x
n
y
n
2y
n
z
n
2x
n
z
n
3
7
7
7
7
7
7
7
5
2
6
6
6
6
6
6
6
6
6
6
6
6
6
4
D
xx
D
yy
D
zz
D
xy
D
yz
D
xz
3
7
7
7
7
7
7
7
7
7
7
7
7
7
5
(9.6)
Though a minimum of six non-collinear gradient directions are required to estimate D,
increasing the number of gradient directions is likely to improve the estimate but the
tradeoffs between the total number of gradient directions and the data acquisition time
are not well understood at this time. For example, one report [HPA01] suggests that for
a given scan time “there does not appear to be a significant advantage to using more
58
than six directions”, whereas a previous report [JHS99] suggested that an optimized
scheme would be “3 measurements acquired with no diffusion weighting and 25 high b-
factor measurements (gradient vectors uniformly distributed in space)”. This issue was
re-investigated by Jones recently [Jon04] concluding that at least 20 unique orientations
were needed to estimate anisotropy and at least 30 orientations were required for a robust
estimation of the tensor-orientation. However it is also pointed out [Jon04] that there is
a diminishing return for more that 10 unique sampling directions.
After estimating D per voxel, tractography is commonly accomplished by a
“streamline tractography (SLT)” approach where a seed voxel is connected to surround-
ing voxels in piecewise linear steps along the direction indicated by the principal eigen-
vector of the diffusion tensor matrix. The vector associated with the first eigenvalue
of the matrix indicates the direction of fastest diffusion and is presumed parallel to the
orientation of a fiber lying in the voxel. This assumption, however, breaks down for
multiple fibers or crossing fibers within a voxel. Models that incorporate higher order
tensors [TRW
+
02] or other approaches, such as use of Independent Component Analy-
sis (ICA) [KJS05] instead of the commonly used Principal Component Analysis (PCA),
are being investigated to resolve the fiber-crossing problem.
The anisotropy per voxel can be expressed in many ways but usually the frac-
tional anisotropy (FA), which is an index of anisotropy normalized to the 0-1 range as
defined below, is computed from the differences among the three eigenvalues (
1
,
2
,

3
).
FA =
r
3
2
s
(
1
hi)
2
+ (
2
hi)
2
+ (
3
hi)
2
(
2
1
+
2
2
+
2
3
)
(9.7)
wherehi = (
1
+
2
+
3
)=3
59
The FA together with the angular change between two steps are commonly used
as stopping criteria to end tracts [CLC
+
99, JHS99, MKP
+
00, BPP
+
00]. A 3D interpo-
lation of the tensor matrix is usually employed to track at a sub-voxel resolution.
Using DTI, the feasibility of mapping long-cortical-association fibers in humans
and the reproducibility of mapping tracts between individuals has been demonstrated
recently [MKD
+
02], and it has been shown that these tracts could be mapped in a
highly reproducible manner between individuals. In the absence of ground truth, there
is a dearth of methods to evaluate the accuracy of a specific tracking algorithm or data
acquisition protocols. Errors in estimating the tensor matrix due to noise in the mea-
sured data and other factors such as movement related artifacts and distortions due to
eddy currents produced by the switching gradients will most likely perturb the eigen-
values [And01], introducing a “sorting bias” [BP00] that may cause a switching among
the three eigenvalues. Though perturbations in general may increase or decrease the FA,
a mere switching between eigenvalues would not change the FA. Thus FA maps may
not be a sensitive indicator to evaluate tractography. However, the noise and gradient
sampling configuration also affect the estimate of the propagation angle between vox-
els [BP00], which is determined by the vector associated with the largest eigenvalue
(
1
). An inaccurate estimate of the propagation angle would cause tracts to be deflected
along a wrong path, making them likely to terminate sooner than if they had continued
along the correct path. Consequently, tensor matrix estimation errors would be mani-
fested in fewer long tracts. Drawing on this rationale, the objective of this work was to
use tract-length histograms to evaluate whole-brain tractography under different condi-
tions, for example 6 gradient directions versus 25 directions acquisitions, or MRI signal
acquisition with an 8-element phased array coil versus a quadrature head coil.
60
9.3 Methodology
Whole-brain single shot EPI DTI data were acquired from six healthy normal human
volunteers on a 1.5 T GE EXCITE scanner at TR=10.3s, field-of-view 26cm, 128x128
matrix, 28 contiguous 4mm thick slices using 25 isotropic gradient directions with
b=1000s/mm
2
, one b=0 acquisition, and number of excitations (NEX)=1 for a total
acquisition time of 3min 53s. Similarly, four sets of data were acquired from 6 non-
collinear directions and combined with two b=0 acquisitions to equalize the time for
25 and 6-directions acquisitions. An 8-element phased array RF coil was used for all
subjects except in one subject where a quadrature head coil was also used to compare
coils.
The b=0 and b=1000s/mm
2
data from the different gradient directions (6 or 25)
were input in Eqn 9.6 above to estimate D per voxel and the eigenvalues of the D-
matrix were estimated using Matlab’s single value decomposition function. The three
eigen-values were then used to compute the FA per voxel. An example of the FA map
is presented in Figure 9.1. Color-coding can be used to demar-cate the three main
directions (Anterior-Posterior (green), Superior-Inferior (blue), Left-Right(red)) in these
maps. Using the RGB color scale, tensors favoring non-orthog-onal directions would be
shown as a combination of red, green, and blue. An intensity threshold mask is applied
to each slice to mask out non-brain areas. Tensors are only calculated for the relevant
brain area of each slice. Each voxel inside the masked volume becomes a seed point for
tracking. Using Euler’s method and a step size of 0.1 mm, the tracking begins at each
seed point and tracks according to the direction prescribed by the eigenvector associated
with the primary eigenvalue and also in the direction completely opposite to the eigen-
vector. The tracking continues until the minimum FA threshold is not met or the angle
of change in one step becomes greater that 45 degrees.
61
Figure 9.1: Fractional Anisotropy map of a 28-slice brain volume
After tracking from every voxel in each slice, the length of each tract was com-
puted in terms of millimeters. A histo-gram of tract-length was then generated using
201 bins, which maps to tract-lengths between 0 and 400 millimeters.
62
9.4 Results and Discussion
A typical result of the 25-direction tractography is presented in Figure 9.2a where all
tracts lying on, or intersecting the selected slice are shown. Below the montage, tracts
for a specified slice are shown superposed within a 3-D rendered brain (Figure 9.2b).
These images are best visualized in a movie shown at http://nmr.usc.edu. Results for
the 6-direction tractography for the same subject, including the rendered version for the
same slice selected for the 25-direction, are presented in Figure 9.3a and 9.3b. The trac-
tography results appear very similar and it becomes difficult to evaluate by visual inspec-
tion alone whether there is any significant difference between the 25 and 6-direction
acquisitions.
On the other hand, the tract-length histograms reveal significant differences
between the two cases. The tract-length histograms for the same subject whose tracts
are shown in Figures 9.2 and 9.3 , are presented in Figure 9.4 (left) and the differences in
the tract-length per histogram bin (25-directions minus the 6-directions lengths) are dis-
played at the right. Similarly, the tract-length histograms and the difference histograms
for two other subjects are presented in Figures 9.5 and 9.6 respectively. The tract-length
histograms clearly show that at equal acquisition or scan time, there are more long tracts
in the 25-direction acquisition than the 6-direction acquisition, suggesting better esti-
mation of the tensor with 25 directions. The tract-count above a threshold could thus be
used as an objective index to evaluate tractography. Table 9.1 lists the tract-count for 25
versus 6 gradient directions with a threshold tract length set at 10 cm for the six subjects
studied. The table shows that 5 out of the 6 subjects show a higher number of tracts in
the case of 25 gradients directions as compared to the 6 gradient direction data.
Results comparing the two coils are presented in Figure 9.7 and Table 9.2, clearly
showing a higher tract-count for long tracts with the 8-element coil. This is consis-
tent with the demonstrated higher sensitivity and higher signal-to-noise ratio for EPI
63
(a) Tractography with 25 gradient directions. Tracts lying on, or
intersecting each slice are shown for 28 slices.
(b) Tracts from 25 gradient direction data that intersect slice 14
rendered in 3D brain volume.
Figure 9.2: Tractography from data obtained using 25 gradient directions.
64
(a) Tractography with 6 gradient directions. Tracts lying on, or
intersecting each slice are shown for 28 slices.
(b) Tractography with 6 gradient directions. Tracts lying on, or
intersecting each slice are shown for 28 slices.
Figure 9.3: Tractography from data obtained using 6 gradient directions.
65
Figure 9.4: Track-length histograms for 25 gradient direction and 6 gradient direction
data sets (left) and difference between the two histograms (25 - 6) (right).
Figure 9.5: Tract-length histograms (left) and difference between histogram (25 gradi-
ents - 6 gradients) (right) for second subject.
25 6
Sub # Length Length
1 3638 3152
2 4293 4143
3 4124 3768
4 4347 3969
5 4762 3962
6 3789 4182
Table 9.1: Tract-lengths for 25 vs. 6 gradients.
66
Figure 9.6: Tract-length histograms (left) and difference between histogram (25 gradi-
ents - 6 gradients) (right) for third subject.
Figure 9.7: Tract-length histogram (left) and difference (right) between histograms for
data obtained with an 8-element coil and a quadrature coil.
8-Element Quad
Sub # Length Length
1 5361 4472
Table 9.2: Tract-lengths for 8-element vs. quadrature coil.
acquisitions by the 8-element coil and provides an indirect validation of the tract-length
histogram to evaluate tractography.
67
9.5 Conclusion
Tract-length histograms provide a reliable and quantitative approach to evaluate 3D trac-
tography where data are acquired under different acquisition protocols or using differ-
ent coils. For a given subject, we have demonstrated that in a fixed scan time, 25-
gradient directions produce more long tracts, suggesting a higher signal-to-noise ratio
than acquisition with 6-gradient directions. Tract-length histograms also demonstrated
longer tracts when an 8-element coil was used compared to a standard quadrature head
coil. Tract-counts above a specified threshold could thus be used as a metric to evaluate
global tractography.
9.6 Acknowledgment
This work is supported in part by grants NIA-NIH P50 AG05142 and NIA-NIH
1PO1AG 12453.
68
Chapter 10
Application 1: Traumatic Brain Injury
This was published in Magnetic Resonance Imaging as “Novel diffusion tensor imag-
ing methodology to detect and quantify injured regions and affected brain pathways in
traumatic brain injury”[SJH
+
10]
10.1 Abstract
Purpose: To develop and apply diffusion tensor imaging (DTI)-based normalization
methodology for the detection and quantification of sites of traumatic brain injury (TBI)
and the impact of injury along specific brain pathways in (a) individual TBI subjects
and (b) a TBI group. Materials and Methods: Normalized DTI tractography was con-
ducted in the native space of 12 TBI and 10 age-matched control subjects using the
same number of seeds in each subject, distributed at anatomically equivalent locations.
Whole-brain tracts from the control group were mapped onto the head of each TBI sub-
ject. Differences in the fractional anisotropy (FA) maps between each TBI subject and
the control group were computed in a common space using a t test, transformed back
to the individual TBI subject’s head space, and thresholded to form regions of inter-
est (ROIs) that were used to sort tracts from the control group and the individual TBI
subject. Tract counts for a given ROI in each TBI subject were compared to group
mean for the same ROI to quantify the impact of injury along affected pathways. The
same procedure was used to compare the TBI group to the control group in a common
space. Results: Sites of injury within individual TBI subjects and affected pathways
69
included hippocampal/fornix, inferior fronto-occipital, inferior longitudinal fasciculus,
corpus callosum (genu and splenium), cortico-spinal tracts and the uncinate fascicu-
lus. Most of these regions were also detected in the group study. Conclusions: The
DTI normalization methodology presented here enables automatic delineation of ROIs
within the heads of individual subjects (or in a group). These ROIs not only localize and
quantify the extent of injury, but also quantify the impact of injury on affected pathways
in an individual or in a group of TBI subjects. Keywords: Diffusion tensor imaging;
Tractography; Normalization; Traumatic brain injury
10.2 Introduction
This article presents a novel methodology to use diffusion tensor imaging (DTI) in
the detection and quantification of traumatic brain injury (TBI) and brain pathways
(tracts) affected by the injury. TBI is a growing health problem in the US, commonly
attributed to motor vehicular accidents and sports injuries but recently also to war-
related injuries sustained near an explosion. One of the serious consequences of TBI
is diffuse axonal injury (DAI), or white matter injuries, induced by sudden accelera-
tion/deceleration and/or rotational/vibrational forces causing a shearing of nerve fibers
[Str61, AGMS82, AHC
+
02, HSS
+
04]. In addition to diffuse injury, local shearing of
axons at the gray/white interface is also possible. Diffuse and local axonal shearing
both disrupt axonal connections critical to brain function, and DAI commonly refers to
both types of injury. DAI has been identified as one of the key reasons for permanent
disability or death. In general, DAI can be very debilitating, leading to a wide range
of neurological impairments from mild memory deficits to persistent vegetative states.
Though CT and MRI are routinely employed to evaluate trauma and DAI, several reports
suggest that CT and commonly used T
1
- and T
2
-weighted MRI protocols are unable to
70
detect the full extent of injury and likely to underestimate the consequences of DAI,
resulting in a poor correlation between diagnosis and final outcome [GGT88, KZS
+
88].
Typically, these patients exhibit a high rate of morbidity without any evidence of lesions
on CT or MRI. It has even been suggested that “DAI can only be definitely diagnosed
postmortem” [HMAT04].
Histologically, DAI is characterized by disruption of the cytoskeletal network
and axonal membranes, leading to impaired axonal transport [AHC
+
02]. Thus a thor-
ough evaluation of DAI requires an imaging method able to quantify the integrity of
axons and the network of connections required to support normal brain function. For
example, a person may look fully recovered months after surviving a nearby explo-
sion, but cognitive function may be abnormal due to loss of connectivity among key
brain regions. Some of the sites frequently damaged in TBI include the corpus callo-
sum, fornix and cingulum [RTC
+
08]. Because of these locations, memory and language
deficits are common as are certain prefrontal syndromes such as dysexecutive, disinhib-
ited and apathetic behavior. Often, these deficits are present without any gross lesions on
the CT scan and the etiology is attributed to DAI. Moreover, approaches to identify DAI
through antibodies targeting amyloid precursor protein (APP) [SP02] may not always
succeed because some axons demonstrate cytoskeletal alteration and detachment with-
out axonal swelling and are thus not identifiable by markers of axonal swelling such as
APP [SSP01].
DTI has been used over the past several years to detect injured regions in TBI
(e.g., [AHC
+
02, HSS
+
04, HMAT04, RTC
+
08, PSR
+
03, IMJ
+
05, SMC
+
06, BZB
+
07]).
DTI is sensitive to the biological diffusion of water molecules, hindered by extracellular
and restricted by intracellular components [AFRB04]. When there is no obstruction to
diffusion, the diffusion tensor is isotropic. However, in the presence of axons and their
myelinated sheath, diffusion becomes anisotropic and quantifiable, which reveals the
71
direction and integrity of axons. Though myelin is not necessary, it influences several
diffusion anisotropy metrics. Thus damage to the myelin sheath and/or axons is poten-
tially detectable by DTI. Maps of several diffusion anisotropy metrics [e.g., fractional
anisotropy (FA) - the normalized differences among the three eigen values of the tensor;
mean diffusivity (MD) - the average of the three eigen values; axial diffusivity (DA) -
the first eigen value oriented along the long axis of the axons; radial diffusivity (DR) -
the average of second and third eigen values; volume ratio (VR) - the normalized prod-
uct of the three eigen values; and the ratio DA/DR] have been evaluated in human and
animal DTI studies [MIZ
+
01, NMMH02, SSR
+
02, ZZZ
+
03, NTL
+
05, TRB
+
06].
The exact model of axonal damage leading to changes in the diffusivity and
resulting diffusion anisotropy measures such as FA is not well understood. There is con-
verging opinion that DAI represents a progressive injury, beginning with local swelling
of axons, followed by cytoskeletal perturbations including misalignment of fibers and
eventual disconnection [AHC
+
02, HSS
+
04, BZB
+
07, BP06]. It has been hypothesized
[AHC
+
02] that the consequences of TBI on DTI would be a decrease in FA and an
increase in MD, attributed mainly to an increase in DR and a decrease in diffusivity
along the principal direction (i.e., a reduction in DA). Indeed, significant FA reductions
but less significant changes in MD have been reported in the internal capsule and cor-
pus callosum during the first 24 h of injury [AHC
+
02]. Changes in FA, however, were
significantly less at 1 month, suggesting the dynamic nature of the injury. A 20-subject
study of changes in the FA maps in DAI [HSS
+
04] with significant correlation with
the Glasgow Coma Scale (r=0.65-0.74,P <.001) also concluded that FA values were
significantly reduced within the internal capsule and splenium. MD was not analyzed
in that study but no significant changes were found in the apparent diffusion coefficient,
which is a less sensitive marker of diffusivity. A previous study also reported an increase
72
in MD but unchanged FA [RGSB
+
01], and another study found reduced FA in the cor-
pus callosum, internal capsule and centrum semiovale, with significant increase in MD
in the corpus callosum and internal capsule [IMJ
+
05]. Similar results have also been
reported very recently [WBD
+
08] for the fornix body, all subregions of the corpus cal-
losum and peduncular projections, with the additional observation that not only did DTI
detect loss of white matter integrity at the beginning of the injury process but the DTI
metrics also correlate strongly to functional outcome.
A recent study also compared mild TBI to moderate-to-severe TBI and found
that the FA in moderate-to-severe TBI was reduced in the posterior corona radiata,
cortico-spinal tracts, cingulum, external capsule, forceps minor and major, genu, body
and splenium of the corpus callosum, inferior fronto-occipital (IFO) fasciculus, superior
longitudinal fasciculus (SLF) and sagittal stratum. FA reduction in mild TBI, however,
was detected only in the cortico-spinal tract, sagittal stratum and the SLF [KSC
+
07].
This study also examined DR and DA and found that both DR and DA were increased
in moderate-to-severe TBI, which could explain the decrease in FA. However, only DA
was found to increase in some regions in mild TBI, while DR did not show any sig-
nificant increase in any region. It is not clear how the FA would decrease under these
circumstances as FA is likely to increase if DA increases and DR remains unchanged.
This specific issue was not addressed by the authors, but the authors suggest that their
results are consistent with the notion that damage to myelin in mild TBI is less com-
mon, whereas both the axons and myelin are likely to be damaged in moderate to severe
injury. Contrary to the above studies, an increase in MD and an increase in FA in an
infant with severe TBI has also been reported [JHS99], though the authors suggest that
the increase in FA is probably a transient effect due to an orderly disruption of cellu-
lar membrane in combination with cellular and vasogenic edema that could temporarily
increase both the FA and diffusivity. Also, an increase in FA and a decrease in MD
73
have been reported very recently [BZB
+
07] in a six-subject study of acute mild TBI. A
significant increase in FA in the posterior corpus callosum and a significant decrease in
MD in the left anterior internal capsule within 72 h of injury were observed, and though
these results were highly correlated with postconcussive symptoms (PCSs) and neurobe-
havioral tests, the authors [BZB
+
07] acknowledged that their results were not consistent
with previous studies of mild TBI [AHC
+
02, IMJ
+
05]. The explanation offered by the
authors was that as most DTI measurements observe diffusion in water contained in the
space between axons (intercellular space) rather than the axons themselves (intracellular
space), the swelling of axons in the acute injury phase would constrict the intercellular
space, leading to a decrease in MD and an increase in FA. This model is plausible but
awaits validation with a larger number of patients. Other possible explanations include
coregistration/normalization errors and the presence of multiple fibers within a voxel. If
more than one fiber were present in a voxel, and one of them was selectively damaged,
the diffusion anisotropy around the remaining intact fibers would be enhanced, leading
to a higher FA for that voxel. Also, a recent longitudinal study of severe TBI found that,
although the FA was reduced in all investigated regions, mainly due to a decrease in DA
and an increase in DR, the FA had increased in the internal capsule and the centrum
semiovale at a mean 12-month follow-up [SES
+
08]. The increase in FA, which now
reached normal to above normal values, primarily in patients with favorable outcome,
was attributed to an increase in DAwith no change in DR. The FA remained depressed
in patients with unfavorable outcome.
Although the model for observed FA changes in TBI is not well established,
it is apparent that maps of anisotropy changes can be, and have been, used to detect
injured regions. However, these maps alone are inadequate to isolate specific brain
pathways disrupted by the injury or to quantify the loss of brain connectivity along
74
affected pathways. DTI tractography, which creates 3D maps of axonal connections,
provides a mechanism to localize and quantify pathways affected by the injury.
Although there are many computational schemes to conduct tractography, the
commonly used streamline tractography procedure tracks the direction of the first eigen
value of an assumed rank-2 diffusion tensor per voxel until either the FA falls below
the threshold or the orientation shows an abrupt angular change exceeding a speci-
fied threshold [CLC
+
99, JHS99, MKP
+
00, BPP
+
00]. A 3D interpolation of the tensor
matrix is usually employed to display tracts with subvoxel resolution.
Although human DTI resolution limitations (approximately 2 mm) limit tractog-
raphy to mapping bundles of tightly packed axons (tracts), this is not expected to be a
major weakness since DAI suffered in head trauma is likely to disrupt tracts and not
merely single axons. This capability to directly visualize axonal connections and quan-
tify connectivity among specified regions is a particular strength of DTI tractography
not achievable with any other imaging modality at the present time. When combined
with methods to detect voxel-based anisotropy changes, DTI including tractography
provides a unique capability to localize and quantify injured regions and brain pathways
affected by the injury. Anisotropy changes are detected by first normalizing the maps
of anisotropy metrics (scalars) from all subjects (subject space) to a common space
(normalized space) and then performing a t test or similar statistical analysis. There
are numerous publications on voxel-based whole-brain anisotropy comparisons, par-
ticularly FA in TBI [RTC
+
08, WBD
+
08, SES
+
08, GSA
+
05, WMH
+
08]. Statistical
comparisons of whole-brain tractography, however, are more complex as normalization
of tensors or tracts requires additional eigen vector corrections. A previous approach
to normalize tractography relied on averaging tensors in normalized space [JGA
+
02],
but this does not fully account for inherent primary eigen vector variations in individ-
uals. The commonly used current approaches to normalize whole-brain tractography
75
incorporate a rotational correction factor that realigns the eigen vectors (or gradients
before estimating the eigen vectors) after mapping all images into a standard or nor-
malized space, consistent with the nonlinear mapping transformations for every voxel
[APBG01, XMS
+
03, MULK07]. However, there are several inherent problems with
tractography when it is conducted after spatial transformations, namely, (a) spatial map-
ping exacerbates partial volume problems due to the required averaging of neighboring
voxels during spatial interpolation, (b) it may not maintain tract topography when non-
linear normalization is used and (c) frequently known continuous tracts are divided into
noncontiguous segments. To avoid performing whole-brain tractography in a common
space, an alternative approach is to first identify ROIs in normalized space by, for exam-
ple, an FA comparison [RTC
+
08], and based on the anatomical locations of these ROIs,
or based on a priori hypotheses [WBD
+
08], subjectively draw these ROIs in each sub-
ject’s head space. These ROIs are then used to sort tracts from whole-brain tractography
conducted in each subject’s native space. However, manually outlining ROIs is subjec-
tive and tracts sorted by such ROIs are prone to relatively large errors as even small
errors in the location of seed points propagate and accumulate along tracts spanning a
large number of voxels.
In an attempt to improve tract normalization and quantification methodology for
application to TBI, the primary objectives of this work were to identify injured regions
in a TBI patient, quantify the injury in terms of DTI anisotropy metrics, identify brain
pathways (tracts) affected by the injury and quantify the effect on impacted pathways.
A preliminary study where affected pathways in four TBI subjects were identified in a
standard normalized space was reported by us recently at an MRI conference [SJ08].
However, as neurosurgical or other individualized interventions rely on the anatomy of
the patient’s own brain, one of the key goals was to quantify injury-related tractography
changes in the patient’s own 3D head space, which requires mapping of whole-brain
76
tractographies of a control group to the 3D brain coordinates of an individual patient. In
addition, a prerequisite was not to require any a priori hypotheses or manual outlining
of ROIs to reduce errors of subjectivity. A secondary objective was to conduct a similar
study comparing a group of TBI patients to a group of control subjects in a common 3D
space to detect and quantify patterns of injury and affected pathways in TBI patients,
also without the need for any a priori hypotheses or manually drawn ROIs. Both objec-
tives require novel tractography normalization and quantification methods. Details of
the methodology and results from a small sample of TBI patients (n=12) and normal
subjects (n=10) are presented in this article.
10.3 Method
10.3.1 Patients and Control Subjects
Patients with mild to moderate brain injuries (loss of consciousness for<1 h) primarily
due to falls, assault or traffic accidents, referred to the LAC+USC Brain Injury Clinic
consented to a DTI scan as part of their medical workup and care. IRB approval was
obtained to process existing DTI data from 12 such TBI subjects who had been iden-
tified by a neurosurgeon based on their clinical history and no evidence of additional
central nervous system disease. The mean and standard deviation of the TBI and control
subjects’ ages were 28±11.4 and 27±8.1 years, respectively, and the mean time interval
between TBI and the DTI scan was approximately 1 month. All 12 patients were ambu-
latory (without assistance) and all could communicate well enough to provide a history
of their injury and care. Patients were interviewed in their primary language (Spanish
or English) regarding symptoms of post-concussive syndrome (headache, dizziness, tin-
nitus, depression, irritability, sexual dysfunction, etc.) and every patient who agreed to
DTI had at least one PCS. None of the patients had hemiparesis or plegia. There were no
77
Figure 10.1: Block diagram of the procedure to detect FA changes between an individual
TBI subject and a group of controls following normalization of all FA images to an FA
template in MNI space. The t map of the FA differences in MNI space was inverse
mapped to the TBI subject’s space and thresholded to generate regions of interest (ROIs)
showing significant FA changes due to injury.
aphasic patients, nor any patients who were blind or deaf. The only cranial nerve deficit
was anosmia in two patients. Existing DTI data that had been acquired on the same
scanner using the same protocol as the TBI patients from a group of 10 age-matched
78
volunteers, with no known history of central nervous system disease and no evidence of
neurological disease on MR imaging, were used as controls.
Figure 10.2: Superposition of fronto-occipital tracts from 10 control subjects, obtained
after mapping whole-brain tracts from all control subjects onto the head of a TBI subject
and sorting in the TBI subject’s head space using a common set of frontal and occipital
regions as filters. (Left) Axial view; (right) sagittal view.
10.3.2 MRI Protocol
The available DTI data from 12 TBI and 10 normal control subjects had been acquired
with a whole-brain single-shot echo planar imaging (EPI) DTI pulse sequence on a 1.5-T
GE EXCITE scanner with an eight-element phased array RF coil at TR=10.3 s, field-of-
view=26 cm, 128 x128 matrix, 28 contiguous 4-mm-thick slices using 25 isotropic gra-
dient directions with b=1000 s/mm
2
, one b=0 acquisition and number of excitations=2
for a total acquisition time of 7 min 50 s. The voxel size was 2.03x2.03x4.0 mm
3
. A
512x512 axial fluid attenuated inversion recovery (FLAIR) had also been acquired from
the TBI patients using TR/TE/inversion time of 8402/146.3/2100 ms, 20 slices, slice
thickness of 7 mm, planar resolution of 0.43x0.43 mm
2
.
79
Figure 10.3: The procedure used to normalize and sort tracts in each TBI subject’s
space. (Top row) Seeds from the standard MNI space were first distributed within each
control and TBI subject’s head using inverse normalization to maintain the same number
of seeds and anatomical equivalency of seed distribution in each subject. (Left to right
blocks) Whole-brain tractography was conducted in each subject and all tracts from
all control subjects were first mapped to MNI space and then mapped onto the head
space of individual TBI subjects. ROIs were identified in the TBI subject’s head using
the procedure outlined in Fig. 10.1. Individual ROIs were then extracted one by one
from the FA-difference map between the TBI subject and controls, and used to sort
tracts. An example of a particular ROI (green) is shown at the bottom of the second
column. Tracts were sorted in the TBI subject’s space using this ROI from all control
subjects and the TBI subject, and the mean number of tracts from the control subjects
was compared to that from the TBI subject to generate the TC metric for this ROI. The
sorted and superimposed tracts for this ROI from two normal subjects (in red and purple,
respectively) and the TBI subject (blue) are shown at the bottom of the third column.
80
10.3.3 DTI Data Processing
The b=1000s/mm
2
images were corrected for distortions and misregistration due to eddy
currents and any movements during the DTI acquisition using the eddy current correc-
tion module of FSL (The Oxford Centre for Functional Magnetic Resonance Imaging of
the Brain Software Library: http://www.fmrib.ox.ac.uk/fsl/fdt). The correction relies on
a 12-parameter affine transformation. Maps of diffusion anisotropy parameters includ-
ing FA, DA, DR and MD were reconstructed using in-house developed DTI software
incorporating a signal-to-noise-ratio weighted multivariate least-square fitting approach
[BMLB94]. Whole-brain tractography was conducted using an in-house developed C-
code version of streamline tractography.
Figure 10.4: Comparison of previous tensor reorientation-based normalized tractogra-
phy (left, in blue) to our normalized approach where all subject space tracts are indi-
vidually mapped to normalized (MNI space) using point-to-point transformation (right,
in red). Sorting of both sets of normalized tractography was conducted with identi-
cal ROIs, located in the frontal and occipital areas (shown in green) in template space.
The tracts generated with the previous voxel-based normalization and reoriented tensor
method tend to suffer from discontinuity toward the ends. Compared to our approach,
the previous method also appears to increase the confounds of partial volume effects
(arrows indicate areas of tract discontinuity and redirection around the posterior right
corner of the ventricle) due to the interpolation inherent to voxel-based normalization.
81
10.3.4 Detection and Quantification of Anisotropy Changes
As several previous studies suggest FA is altered significantly in TBI, identification of
the injured regions was based on changes in an individual’s FA map with respect to
a group of normal subjects. This comparison was carried out in the standard Mon-
treal Neurological Institute (MNI) space using SPM (Statistical Parametric Mapping,
http://www.fil.ion.ucl.ac.uk/spm)-based forward and inverse normalization procedures
[AF99, FAK
+
07]. In forward normalization, each pointx within subject space image
f(x) is mapped by a unique warping functionM(
:
) onto its equivalent point y of a tem-
plate in MNI space to generate the normalized image,g(y). Inverse normalization does
the reverse, i.e., maps from MNI space back to an individual subject’s space. As each
point within the template space maps to exactly one point in subject space, a unique
inverse mapping from the MNI space to the subject space, M
1
(
:
), exists between all
points of both spaces. SPM normalization relying on a 12-parameter affine transforma-
tion to register voxels of the subject space to those of the template space followed by
a nonlinear warping transformation to estimate the 3D deformation field at each point
x was used. SPM uses a linear combination of 3D discrete cosine transform (DCT)
basis functions in three orthogonal directions to model the deformation. The DCT coef-
ficients are iteratively optimized via Gauss-Newton strategy to minimize the bending
energies of the deformation fields as well as the residual squared difference between
f(x) and g(y). We used fourth-order DCT basis functions, 16 iterations and trilinear
interpolation in this work. An FA template was constructed from the 10 normal control
subjects using a two-step procedure where segmented white matter voxels within each
normal subject’s b=0 image were first mapped to the MNI white-matter template using
SPM 12-parameter affine/nonlinear transformation. Segmentation of the b=0 images
into white-matter voxels was also done by SPM retaining only those voxels that dis-
played >50% probability of being classified as white matter. The forward mapping
82
parameters obtained in the first step were applied to whole-brain b=0 images and refined
in a second step by mapping the resulting images to the MNI EPI template. The result-
ing parameters were applied to individual FA maps of normal control subjects to create
the FA template in MNI space froman average of the 10 control subjects. The FA maps
of all control and TBI subjects were then individually remapped to the FA template, and
data were analyzed to determine the voxel-based FA differences between (a) each TBI
subject and the control group as described below, and (b) the TBI group and control
group in MNI space.
10.3.4.1 Individual TBI Subject vs. Control Group
A block diagram illustrating the procedure to determine the differences between the
control group and individual TBI subjects is presented in Fig. 10.1.
The FA map of an individual was first compared voxel by voxel to the FA maps
of the control group in MNI space by computing a modified t score defined as:
t
i
= [FA
i
(control)FA
i
(individual)]=
i
(control) (10.1)
whereFA
i
(control),FA
i
(individual) and
i
represent mean FA of the control group,
FA of an individual and the standard deviation in the FA values of the control group,
respectively, for theith voxel. Thet score in Eq. 10.1 thus reflects the number of stan-
dard deviations by which the FA of an individual is reduced with respect to the control
group FA mean, equivalent to a one-samplet test. Thet-score distribution in MNI space
for an individual was then mapped back to the individual subject’s head by using inverse
normalization and thresholded (t3.0, cluster sizek12) to localize the injured regions.
V oxels satisfying the threshold criteria were clustered automatically into ROIs relying
on contiguity of voxels along any direction. In addition to FA, other diffusion anisotropy
83
metrics, e.g., DA, DR, MD, were also computed for these ROIs. Thet-score map of an
individual in MNI space was also mapped onto FLAIR images using the inverse of the
FLAIR to MNI T
1
-template normalization done by a 12-parameter affine transformation
followed by nonlinear warping. After mapping, thet-score distribution was thresholded
as before (t3.0,k12) to show FA-changed regions superimposed on FLAIR images.
10.3.4.2 TBI Group vs. Control Group
Standard SPM group comparison procedure was used to determine FA differences
between the TBI group and the control group in MNI space. A two-samplet test was
used to detect ROIs where FA had changed significantly (P
FDR
.05 andk12) in MNI
space, whereP
FDR
denotes the false detection rate correctedP value.
10.3.5 Seed Placement and Tractography
As the number and characteristics of tracts are critically dependent on the number of
seeds and how seeds are distributed in the head, it is critical to account for these varia-
tions among subjects to normalize tracts. We achieved a normalized distribution of seeds
in individuals by inverse mapping seeds from the FA template onto each subject’s native
space, thereby distributing the same number of seeds in each subject at anatomically
equivalent locations throughout the brain to account for variations in head size, shape
and white-matter distribution. The MNI space was masked to create a volume that corre-
sponded to the common brain volume scanned in each subject’s head. The center coordi-
nates of voxels in the common brain volume were considered to be seeds for all subjects.
The millimeter-space coordinates of these seeds were then mapped on a point-to-point
basis to the native space of each subject in “floating point” using inverse mapping from
MNI space to subject space, as described in Section 10.3.4, without any interpolation.
The mapped MNI space seed points were used to conduct whole-brain tractography
84
in each subject. An in-house developed streamline tractography code based on Euler’s
method and a step size of 0.2 mm with tensor interpolation [BPP
+
00, MvZ02, SHSV05]
was used to generate tracts from all seeds (whole-brain tractography), where the num-
ber of seeds and their distribution were normalized in each subject’s head as described
above. Tracts were propagated along the direction indicated by the primary eigen vector
and continued until either an FA threshold was not met or the deflection angle exceeded
45°.
A similar procedure was first proposed by Clayden et al. [CMB
+
05, CBS06]
but apparently not pursued due to registration errors in their approach. With our FA
template-based normalization, we have been able to achieve very good registration
among the 10 control subjects used in this work suggesting the viability of this approach.
An example of the registration accuracy is presented in Fig. 10.2, where whole-brain
tractography was first conducted in each control subject using an equal number of seeds
obtained by inverse mapping, and then all tracts from all control subjects were mapped
onto a TBI subject’s head (see Section 10.3.6.1 below) and sorted in the TBI subject to
extract fronto-occipital tracts using planar filters as suggested by Mori et al. [MKD
+
02].
It is visually apparent that the overlap among the 10 subjects is very good in the middle
portion of the tracts while there are intersubject variations (to be expected) as the tracts
propagate toward cortical boundaries.
10.3.6 Detection and Quantification of Affected Tracts
10.3.6.1 Individual TBI Subject vs. Control Group
Whole-brain tractographs of each TBI subject were filtered by FA-reduced ROIs identi-
fied in the subject’s native space as described above (Section 10.3.4.1) to isolate corre-
sponding pathways affected by each injured region. Quantification of connectivity along
85
Figure 10.5: FA-reduced regions corresponding tot3.0 (see color bar) superimposed
on the FLAIR images of TBI Subject 1. An extent threshold k12 was also used to
identify clusters. Some FA-reduced regions overlap completely, others partially and
some do not overlap with the FLAIR spots. (In this montage, the left hemisphere L
appears at the right in each image.)
the affected pathways was then performed by comparing the number of tracts traversing
a given ROI in the TBI subject to the mean number of tracts from the control group also
traversing the same ROI in the TBI subject’s space after mapping wholebrain tracts of
the entire control group to the individual TBI subject’s space as described below. Con-
nectivity here is defined by the number of tracts traversing an ROI along a specified
pathway. When small step sizes (such as the 0.2-mm step size used in this work) are
used with interpolation after each step in streamline tractography, and when only one
seed-one eigen vector is used per voxel, then each of the n voxels lying on a tract ema-
nating from the ith seed will generally also “send” a tract back to the ith voxel. Thus the
86
number of tracts intersecting a voxel is a reflection of how many voxels are connected to
it, which provides a convenient metric to quantify connectivity. In this work, the number
of tracts (referred to as the tract count) traversing voxels within an ROI along a specific
pathway was used as the metric to quantify the reduction in connectivity in a TBI subject
with respect to a control group along the same pathway at a given FA threshold.
A two-step procedure was implemented to map whole-brain tracts from each of
the control subjects to the TBI subject’s space. Tracts from all control subjects were
first mapped to the MNI space and then the TBI subject-specific inverse parameters
were used to map all tracts from the MNI space to the TBI subject’s space. Point-to-
point mapping of all points lying on each tract was accomplished in the following way.
SPM normalization relies on one-to-one mapping of the vertices of tetrahederal volume
elements [AAF00] from one space to another. Let us assume thaty =M(x) represents
the mapping, via FA template-based spatial normalization, of a seed x = [x
1
;x
2
;x
3
],
lying at the vertex of a tetrahedral in subject space, to pointy = [y
1
;y
2
;y
3
], also lying on
a vertex of a tetrahedral in standard MNI space. Then the mapped pointy
0
corresponding
the pointx
0
on a tract following a small step can be obtained from the following equation,
y
0
=
2
6
6
6
4
y
0
1
y
0
2
y
0
3
3
7
7
7
5
=M(x
0
)

=M(x) +J
x

2
6
6
6
4
@x
1
@x
2
@x
3
3
7
7
7
5
(10.2)
where the step in any direction is represented by [@x
1
;@x
2
;@x
3
], andJ
x
defines
the Jacobian of the deformation at the point x = [x
1
;x
2
;x
3
]. The elements of the
Jacobian were determined by the multiplication of partial derivatives of the deformation
field along pairs of two orthogonal axes. Similarly, the tract pointy
0
= [y
1
+@y
1
;y
2
+
87
Figure 10.6: Three pathways, hippocampal/fornix (HC/FX), inferior fronto-occipital
(IFO) and inferior longitudinal fasciculus (ILF), identified as crossing the voxels of the
highest t-score ROI in a TBI subject (Subject 1). (Top panel) Coronal, sagittal and
axial views of the ROI with color coding of the t-score as indicated in the color bar.
(Middle row) The three pathways (HC/FX—red, IFO—magenta, ILF—black) in a nor-
mal subject shown in axial and sagittal views. (Bottom row) The same three pathways
(HC/FX—blue, IFO—magenta, ILF—black) in a TBI subject. (The yellow and green
colors in the middle and bottom row pictures were used to identify sub-ROIs within the
ROI as described in the text.)
88
@y
2
;y
3
+@y
3
] in the standard MNI space can be transformed back to the point x
0
=
[x
0
1
;x
0
2
;x
0
3
] in a specific subject space via the following equation,
x
0
=
2
6
6
6
4
x
0
1
x
0
2
x
0
3
3
7
7
7
5
=M(y
0
)

=M
1
(y) +J
y

2
6
6
6
4
@y
1
@y
2
@y
3
3
7
7
7
5
(10.3)
whereM
1
(y) indicates a one-to-one mapping of the vertex point y in the stan-
dard MNI space to the vertex point x in the subject space. The deformation deriva-
tives at the point y and its increments are represented by the Jacobian matrix J
y
and
[@y
1
;@y
2
;@y
3
], respectively.
Figure 10.7: Tract count-based sensitivity (effect size) to differentiate TBI subjects from
controls as a function of FA.
Based on the forward and inverse mapping described in Eqs. 10.2 and 10.3, all
points on any tract in any Subject 1 can be transformed to the standard MNI space and
then inverse mapped to the subject space of any other Subject 2 using the inverse map-
ping function of Subject 2, thereby making it possible to map all tracts from a group of
control subjects onto the subject space of an individual TBI subject. In conjunction with
89
the normalized seed tractography described in Section 10.3.5, it now becomes possible
to compare quantitatively a specific tract identified in a TBI subject to corresponding
tracts in a control group, without requiring any manual drawing of any ROI.
The overall approach to normalize and map tracts from control subjects onto the
head of a TBI subject is conceptualized in Fig. 10.3. Seeds were first distributed in each
subject’s head using inverse mapping of MNI seeds as shown in the top row, and whole-
brain streamline tractography was conducted in each subject’s native space from these
mapped seeds. All tracts from each of the 10 control subjects (left box in Fig. 10.3)
were first mapped onto MNI space using nonlinear warping and then mapped to the
head of a TBI subject (right box) where they were sorted by ROIs obtained from the FA
comparison (Section 10.3.4.1). An example of tracts sorted by a particular ROI (shown
as a green cluster) in the TBI subject space is also shown in Fig. 10.3, where in this
particular case, the ROI cluster identified the fronto-occipital and hippocampal/fornix
pathways. In situations like this where more than one pathway was identified by an
ROI, the ROI was subdivided into multiple sub-ROIs by identifying voxels along each
pathway and grouping voxels (within the ROI) that were common to at least 50% of the
control subjects along a given pathway. This procedure was used to mitigate effects of
intersubject tract variations and coregistration errors remaining after our normalization
procedure.
The number of tracts passing through each ROI (or sub-ROI) was counted in the
TBI subject and compared to the mean and standard deviation of the normalized tracts
from all control subjects passing through the same ROI (or sub-ROI) in the TBI sub-
ject space to quantify the reduction along a specific pathway. When multiple pathways
were identified within an ROI, the FA and other diffusion anisotropy metrics were also
computed for the sub-ROIs.
90
ROI t score Pathway FA Effect size TC Effect size
Control TBI Control TBI
1 5.35 (1.68) HC/FX 0.32 (0.10) 0.17 (0.11) 1.43 367.80 (88.58) 135 2.63
IFO-L 0.47 (0.08) 0.21 (0.12) 2.55 421.50 (114.29) 119 2.65
ILF-L 0.49 (0.08) 0.24 (0.09) 2.94 404.30 (180.02) 121 1.57
2 5.29 (1.59) IFO-L 0.28 (0.06) 0.11 (0.04) 3.33 409.60 (150.97) 62 2.3
3 5.00 (0.83) CC-g 0.47 (0.03) 0.33 (0.01) 6.26 238.00 (55.04) 171 1.22
UF-R 0.42 (0.04) 0.31 (0.02) 3.48 113.80 (41.11) 55 1.43
4 4.84 (0.85) CC-s 0.45 (0.05) 0.28 (0.07) 2.79 375.80 (116.83) 169 1.77
5 4.80 (1.14) FX-L 0.29 (0.10) 0.16 (0.09) 1.37 253.50 (110.68) 91 1.47
IFO-L 0.47 (0.07) 0.28 (0.07) 2.71 347.80 (108.94) 74 2.51
ILF-L 0.48 (0.06) 0.29 (0.07) 2.91 410.10 (181.88) 115 1.62
6 4.77 (0.93) CC-g 0.44 (0.09) 0.27 (0.08) 2 589.70 (184.80) 571 0.1
IFO-L 0.38 (0.06) 0.25 (0.05) 2.35 526.20 (120.94) 81 3.68
7 4.32 (0.53) CG/FX 0.25 (0.07) 0.10 (0.04) 2.63 215.40 (64.76) 10 3.17
8 4.51 (0.64) IFO-L 0.35 (0.07) 0.18 (0.04) 2.98 285.80 (102.36) 51 2.29
9 4.67 (1.22) SLF-L 0.43 (0.09) 0.30 (0.10) 1.37 957.10 (146.47) 730 1.55
10 4.58 (0.76) CST 0.30 (0.12) 0.16 (0.09) 1.32 594.50 (205.11) 15 2.83
11 4.18 (0.45) UF-L 0.22 (0.05) 0.11 (0.03) 2.67 75.80 (46.05) 16 1.3
12 3.13 (0.61) HC/FX-R 0.32 (0.09) 0.22 (0.09) 1.11 232.90 (70.22) 8 3.2
ROI Pathway MD (x10
3
mm
2
/s) Effect size DA (x10
3
mm
2
/s) Effect size
Control TBI Control TBI
1 HC/FX 0.78 (0.20) 1.91 (0.55) 2.73 1.05 (0.23) 2.16 (0.48) 2.95
IFO-L 0.83 (0.20) 1.83 (0.59) 2.27 1.28 (0.20) 2.14 (0.49) 2.3
ILF-L 0.74 (0.11) 1.51 (0.44) 2.4 1.18 (0.14) 1.84 (0.39) 2.25
2 IFO-L 0.72 (0.06) 1.21 (0.18) 3.65 0.94 (0.07) 1.34 (0.18) 2.93
3 CC-g 0.74 (0.03) 0.81 (0.01) 3.13 1.17 (0.04) 1.13 (0.01) -1.37
UF-R 0.75 (0.03) 0.81 (0.01) 2.68 1.12 (0.06) 1.10 (0.02) -0.45
4 CC-s 0.64 (0.03) 0.72 (0.03) 2.67 0.97 (0.07) 0.92 (0.04) -0.88
5 FX-L 0.97 (0.24) 1.64 (0.51) 1.68 1.24 (0.24) 1.85 (0.50) 1.56
IFO-L 0.76 (0.07) 1.11 (0.24) 1.98 1.17 (0.11) 1.41 (0.23) 1.33
ILF-L 0.73 (0.07) 1.02 (0.21) 1.85 1.15 (0.11) 1.31 (0.22) 0.92
6 CC-g 0.82 (0.20) 1.18 (0.51) 0.93 1.23 (0.25) 1.48 (0.54) 0.59
IFO-L 0.72 (0.06) 0.97 (0.22) 1.55 1.04 (0.09) 1.21 (0.20) 1.1
7 CG/FX 1.22 (0.43) 2.25 (0.42) 2.42 1.54 (0.48) 2.44 (0.42) 2
8 IFO-L 0.72 (0.04) 0.97 (0.06) 4.9 1.01 (0.06) 1.14 (0.06) 2.17
9 SLF-L 0.62 (0.03) 0.83 (0.17) 1.72 0.94 (0.10) 1.08 (0.16) 1.05
10 CST 0.94 (0.21) 1.81 (0.43) 2.57 1.23 (0.24) 2.04 (0.38) 2.55
11 UF-L 0.76 (0.07) 1.45 (0.14) 6.23 0.94 (0.07) 1.60 (0.12) 6.72
12 HC/FX-R 0.87 (0.21) 1.32 (0.48) 1.21 1.15 (0.21) 1.58 (0.48) 1.16
ROI Pathway DR (x10
3
mm
2
/s) Effect size
Control TBI
1 HC/FX 0.65 (0.20) 1.78 (0.58) 2.6
IFO-L 0.61 (0.21) 1.67 (0.64) 2.23
ILF-L 0.53 (0.11) 1.35 (0.47) 2.4
2 IFO-L 0.61 (0.07) 1.15 (0.18) 3.95
3 CC-g 0.52 (0.03) 0.66 (0.01) 6.26
UF-R 0.56 (0.04) 0.67 (0.02) 3.48
4 CC-s 0.47 (0.03) 0.62 (0.05) 3.64
5 FX-L 0.83 (0.24) 1.54 (0.52) 1.75
IFO-L 0.55 (0.07) 0.96 (0.26) 2.15
ILF-L 0.52 (0.07) 0.87 (0.22) 2.14
6 CC-g 0.61 (0.19) 1.02 (0.50) 1.08
IFO-L 0.56 (0.07) 0.85 (0.24) 1.64
7 CG/FX 1.07 (0.41) 2.15 (0.42) 2.6
8 IFO-L 0.58 (0.06) 0.89 (0.07) 4.76
9 SLF-L 0.46 (0.05) 0.70 (0.19) 1.73
10 CST 0.80 (0.23) 1.69 (0.46) 2.45
11 UF-L 0.67 (0.08) 1.38 (0.16) 5.61
12 HC/FX-R 0.73 (0.21) 1.20 (0.49) 1.25
Table 10.1: List of affected pathways and diffusion metrics for 12 ROIs identified by
FA reduction in a TBI subject (Subject 1) compared to a group of 10 normal controls
(t3.00,k12)
91
Figure 10.8: Similar to Fig. 10.6, pathways associated with two other ROIs within TBI
Subject 1 are shown in (A) and (B), respectively. Coronal, sagittal and axial views of
the two ROIs with color coding of the t-score as indicated in the color bar are shown in
the top row. The ROI in (A) identified tracts through the posterior portion of the corpus
callosum, whereas (B) identified the right HC/FX tracts. Like Fig. 10.6, tracts for a
normal subject are shown in red and those in the TBI subject are shown in blue. The
reduction of tract counts in the TBI subject with respect to the control subject is obvious
for these ROIs.
The key feature of our tract normalization approach is to conduct tractography
in subject space using a unique seed distribution and then transform each tract indi-
vidually, first to MNI space and then to the head space of another subject. Contrary
92
to previous approaches [APBG01, MULK07], our approach does not warp voxels but
instead takes each individual tract in subject space composed of a string of points located
0.2 mm apart, transforms these points to normalized space using a point-to-point map-
ping and connects these mapped points to regenerate the new tract in normalized space.
Mapping these individual 0.2-mm spaced points and reconnecting them in normalized
space is equivalent to reorienting the primary eigen vectors along each tract from sub-
ject to normalized space consistent with the nonlinear transformation, thereby overcom-
ing the problem of correcting eigen vectors used in previous normalization approaches
[APBG01, MULK07]. To validate this key feature, a study was performed to com-
pare tract normalizations obtained with our approach and the commonly used previous
approach [APBG01, MULK07] where the b=0 and all b=1000 s/mm
2
images are first
transformed to MNI space and then the perturbations in eigen vectors are corrected to
account for spatial transformations. A common set of sorting filters were employed
in both cases to compare the integrity of known fronto-occipital connections after nor-
malization. The results of this comparison are presented in Fig. 10.4. Although both
approaches show recovery of the IFO and SLF by the filters used to sort these tracts, our
method shows better continuity and branching near the ends (consistent with subject-
space mapping), whereas the previous approach shows discontinuities apparently caused
by the voxel-based normalization. Also, the inherent smoothing in voxel-based normal-
ization increases partial volume artifacts causing a portion of the right IFO tracts to turn
around the ventricle, whereas this effect is much reduced in our approach (Fig. 10.4).
10.3.6.2 TBI Group vs. Control Group
The group analysis was done in MNI space. The same procedure as described above
was used to map all tracts from all subjects including the TBI and control subjects into
the common MNI space, and ROIs defined from the FA group comparison in MNI
93
Figure 10.9: Similar to Fig. 10.5, the FA-reduced regions for TBI Subject 2 superim-
posed on the FLAIR images of the subject. (In this montage, the left hemisphere L
appears at the right in each image.)
space (Section 10.3.4.2) were used to isolate pathways affected by specific ROIs in each
subject. Tract counts along these pathways were compared between the two groups to
quantify the reduction in connectivity.
10.4 Results
10.4.1 Individual TBI Subjects
An example of t score maps corresponding to significant reduction in FA (t 3:0, see
color bar) and cluster extentk 12 between a TBI subject (Subject 1) and 10 controls
94
is presented in Fig. 10.5 where the FA-reduced regions have been superimposed on TBI
subject’s FLAIR images. It can be seen that there is partial overlap between the FLAIR
spots (which mainly highlight edema resulting from cellular injury) and the FA-reduced
regions, although there are regions highlighted in FLAIR but not observed in the FA
reduced maps and vice versa, consistent with previous studies [RTC
+
08, WBD
+
08,
SES
+
08, WMH
+
08, DHF
+
05]. These results are also consistent with a recent study
[ZZL
+
08] where partial overlap between the FLAIR and FA-reduced regions has been
explained by a model of the contrast mechanism in each modality.
With the use of the FA-reduced regions of Fig. 10.5 as ROIs, the sorted tracts
in TBI Subject 1 were compared to sorted tracts obtained from each of the 10 control
subjects for the same ROIs after mapping all tracts of all control subjects to the head of
the TBI subject as described in Section 10.3.6.1. As an example, pathways detected by
the ROI corresponding to the highest t score in TBI Subject 1 are shown in Fig. 10.6 at
FA0.15.
The FA threshold has a significant impact on tract counts. FA thresholds ranging
from 0.10 to 0.3 have been commonly used in streamline tractography, representing a
tradeoff between noise (which increases as the FA threshold is lowered) and sensitivity
(which decreases as the FA threshold is raised). A previous study where the influence of
FA threshold on the accuracy of DTI tractography was evaluated by comparing histology
(ground truth) to DTI tractography at FA thresholds varying from 0 to 0.60 for every 0.05
[DPB
+
07] reported that the best visual agreement was achieved at an FA threshold of
0.10 in that work. To determine an optimal FA for this study, we measured the sensitivity
of tract counts to differentiate between controls and individual TBI subjects as a function
of FA threshold. A tract-count (TC)-based effect size was defined as:
Effectsize
i
= [TC
i
(control)TC
i
(individual)]=
i
(control) (10.4)
95
Figure 10.10: Similar to Fig. 10.6, pathways associated with two ROIs in TBI Subject 2
are shown in (A) and (B), respectively. The ROI in (A) identified the left HC/FX tracts,
whereas (B) identified the right HC/FX tracts. Like Fig. 10.6, tracts for a normal subject
are shown in red and those in the TBI subject are shown in blue. The reduction of tract
counts in the TBI subject with respect to the control subject is obvious in these ROIs.
whereTC
i
(control),TC
i
(individual) and
i
represent mean TC of the control group,
TC of an individual TBI subject and the standard deviation in the TC values of the con-
trol group, respectively, for the ith ROI. The effect size was computed for 28 FA-reduced
ROIs identified in two TBI subjects at thet 3:0,k 12 threshold. A plot of the mean
96
effect size over these 28 ROIs as a function of FA is presented in Fig. 10.7, showing
that the effect size peaks at an FA threshold of about 0.15. Thus an FA threshold of 0.15
was used in this work, though the results would not change significantly for FA thresh-
olds ranging from 0.1 to 0.2. Examples of pathways identified by two other relatively
high t-score FA-reduced regions for TBI Subject 1 are shown in Fig. 10.8 at FA0.15.
Though most ROIs identified only one pathway or tract, some ROIs, for example the
ROI shown in Fig. 10.6, identified three distinct pathways going through it, namely,
hippocampal/fornix (HC/FX), inferior fronto-occipital (IFO) and inferior longitudinal
fasciculus (ILF). In situations like this where the ROI contained multiple pathways, or
in the situation where some voxels within an ROI did not show any pathway as their
FAvalues were below the FA threshold, the ROI was partitioned into multiple sub-ROIs
by grouping voxels within each sub-ROI that were common to tracts along specific path-
ways in at least 50% of the control subjects. The diffusion anisotropy and TC metrics
were then computed within these sub-ROIs. A summary of these metrics and the path-
ways identified by the ROIs/sub-ROIs is presented in Table 10.1. The values of FA (and
other diffusion metrics) for all voxels contained within the ROIs were pooled for the 10
control subjects to compute the mean and standard deviation. When an ROI indicated
multiple pathways, only those voxels within the ROI that contained tracts from at least
50% of the subjects along a specific pathway were considered. In addition to FA, the
anisotropy metrics DA, DR and MD were also computed as listed in Table 10.1.
The FA-reduced regions of another TBI subject (Subject 2), detected att3.0,
k12, superimposed on the subject’s FLAIR images, are presented in Fig. 10.9, and,
as an example of quantifying pathways affected by injury, the tracts traversing two of
the FA-reduced regions in this subject are shown in Fig. 10.10. The anisotropy metrics
and the tract counts for all ROIs att3.0,k12 are summarized in Table 10.2 for TBI
Subject 2.
97
ROI t score Pathway FA Effect size TC Effect size
Control TBI Control TBI
1 4.13 (0.91) CC-b 0.29 (0.08) 0.09 (0.02) 3.43 225.50 (111.24) 2 2.01
2 4.09 (0.82) CC-s 0.49 (0.07) 0.33 (0.05) 2.63 422.50 (136.31) 249 1.27
3 4.03 (0.93) CC-s 0.37 (0.08) 0.17 (0.08) 2.5 543.80 (141.35) 174 2.62
4 3.89 (0.53) ILF-L 0.46 (0.07) 0.26 (0.09) 2.48 277.10 (117.08) 8 2.30
5 3.79 (0.54) CC-g 0.35 (0.04) 0.23 (0.02) 3.79 316.20 (118.45) 317 -0.01
6 3.73 (0.56) CST 0.48 (0.05) 0.38 (0.05) 2 433.00 (151.25) 355 0.52
7 3.71 (0.62) CC-b 0.51 (0.03) 0.39 (0.01) 5.37 108.80 (94.95) 30 0.83
SLF-L 0.31 (0.07) 0.25 (0.07) 0.86 45.00 (41.46) 15 0.72
8 3.67 (0.47) SLF-L 0.49 (0.05) 0.36 (0.07) 2.14 536.00 (179.76) 177 2.00
9 3.58 (0.38) CC-s 0.47 (0.11) 0.16 (0.03) 3.85 658.50 (109.39) 206 4.14
10 3.54 (0.45) SLF-L 0.39 (0.05) 0.27 (0.03) 2.91 307.00 (146.48) 168 0.95
11 2.67 (0.59) HC/FX-R 0.33 (0.06) 0.27 (0.05) 1.09 117.40 (53.08) 78 0.74
ROI Pathway MD (x10
3
mm
2
/s) Effect size DA (x10
3
mm
2
/s) Effect size
Control TBI Control TBI
1 CC-b 0.65 (0.05) 0.92 (0.09) 3.71 0.85 (0.07) 1.00 (0.10) 1.74
2 CC-s 0.67 (0.03) 0.79 (0.03) 4 1.08 (0.09) 1.03 (0.05) -0.69
3 CC-s 0.72 (0.06) 0.85 (0.09) 1.7 1.02 (0.10) 0.99 (0.07) -0.35
4 ILF-L 0.68 (0.03) 0.83 (0.05) 3.64 1.06 (0.07) 1.05 (0.06) -0.15
5 CC-g 0.68 (0.04) 0.81 (0.02) 4.11 0.95 (0.05) 1.01 (0.04) 1.33
6 CST 0.62 (0.02) 0.69 (0.01) 4.43 0.98 (0.07) 1.00 (0.06) 0.31
7 CC-b 0.63 (0.03) 0.71 (0.02) 3.14 1.03 (0.04) 1.02 (0.02) -0.32
SLF-L 0.62 (0.05) 0.75 (0.14) 1.24 0.82 (0.06) 0.95 (0.11) 1.47
8 SLF-L 0.69 (0.04) 0.82 (0.03) 3.68 1.09 (0.09) 1.15 (0.08) 0.7
9 CC-s 0.89 (0.23) 0.88 (0.10) -0.06 1.38 (0.28) 1.04 (0.12) -1.58
10 SLF-L 0.73 (0.04) 0.88 (0.03) 4.24 1.07 (0.08) 1.14 (0.05) 1.05
11 HC/FX-R 0.83 (0.09) 0.93 (0.08) 1.17 1.12 (0.11) 1.19 (0.10) 0.67
ROI Pathway DR (x10
3
mm
2
/s) Effect size
Control TBI
1 CC-b 0.54 (0.07) 0.88 (0.09) 4.22
2 CC-s 0.47 (0.04) 0.66 (0.04) 4.75
3 CC-s 0.57 (0.07) 0.78 (0.11) 2.28
4 ILF-L 0.49 (0.06) 0.72 (0.07) 3.53
5 CC-g 0.55 (0.05) 0.72 (0.03) 4.12
6 CST 0.44 (0.02) 0.54 (0.02) 5.00
7 CC-b 0.43 (0.03) 0.55 (0.02) 4.71
SLF-L 0.52 (0.06) 0.66 (0.15) 1.23
8 SLF-L 0.48 (0.04) 0.66 (0.05) 3.98
9 CC-s 0.64 (0.23) 0.81 (0.10) 0.96
10 SLF-L 0.56 (0.04) 0.75 (0.04) 4.75
11 HC/FX-R 0.69 (0.10) 0.79 (0.08) 1.10
All analysis was done in the TBI subject’s head space. CC-b: Corpus callosum body.
Table 10.2: List of affected pathways and diffusion metrics for 11 ROIs identified by
FA reduction in a TBI subject (Subject 2) compared to a group of 10 normal controls
(t3.0,k12)
10.4.2 Group Study: Control vs. TBI Group
The results of the group comparison between 10 normal controls and 12 TBI subjects
are presented in Figs. 10.11 and 10.12. The FA-reduced regions in the TBI group at
P
FDR
.05,k12 are shown in Fig. 10.11, where these FA-reduced regions have been
superimposed on our FA template in MNI space to show their relative location with
98
Figure 10.11: A superposition of the FA-reduced regions (in color) detected by an SPM-
based statistical comparison of the TBI group (n=12) to the normal group (n=10) at
P
FDR
.05, k12 on the FA template in MNI space. Arrows point to the three ROIs
(genu of the corpus callosum, hippocampal region, splenium of the corpus callosum)
used in Fig. 10.12.
respect to some of the affected pathways. As an example, the pathways obtained by
using two of these ROIs to sort tracts from whole-brain tracts are shown in Fig. 10.12.
Just like individual TBI subjects, when multiple pathways were detected within an ROI,
for example as indicated in Fig. 10.12A where the HC/FX, IFO and a portion of the ILF
were detected, the anisotropy and TC metrics were computed within sub-ROIs formed
by voxels lying at the intersection of tracts from at least 50% of the subjects. Also, as
99
in single TBI subject studies, the same procedure was used to refine those ROIs where
some voxels did not contain any tracts as their FA values were below the threshold.
The anterior and inferior portions of the corpus callosum were frequently injured
in this cohort of TBI subjects. An example of the ROI associated with injury to the
anterior (genu) of the corpus callosum and affected tracts is shown in Fig. 10.12B. All
the detected ROIs with their associated pathways, and the anisotropy and TC metrics for
these ROIs, are listed in Table 10.3.
10.5 Discussion
This article presents a methodology not only to identify and quantify the location and
extent of brain regions injured in TBI, but more importantly also to quantify the impact
of injury on specific pathways in an individual. This information should be vital to
diagnose and monitor objectively the progression of the injury in TBI as well as to
assess the outcome of treatment.
Identification of injured regions was based on detecting voxel-based changes in
FAvalues between an individual TBI subject and a group of age-matched control subjects
after normalization of all FA maps onto a common FA template in MNI space. The
FA-reduced regions were subsequently mapped back to each subject’s head space and
used as objective ROIs to sort tracts in each subject’s head. A unique seed-placement
procedure where the number of seeds and their corresponding anatomical locations were
equalized in all subjects was used to attain tractography normalization for quantitative
comparisons. A unique procedure was also devised to map tracts from any subject
onto the head of any other subject to perform quantitative comparisons within the head
space of any individual subject. Thus it now becomes possible to visualize, quantify
and compare any specific tract or pathway in an individual patient to corresponding
100
Figure 10.12: Similar to Fig. 10.6, pathways associated with three ROIs identified as
FA-reduced regions in the group study between 10 controls and 12 TBI subjects are
shown in (A), (B), and (C), respectively. The ROI in the middle row of (A) identified
the left IFO, HC/FX and a portion of the ILF for a normal control subject (IFO—purple;
ILF—black, HC/FX—red); (B) identified the anterior portion of the corpus callosum
(red) and (C) identified the posterior portion of the corpus callosum (red). The bottom
row shows corresponding tracts identified by the same ROIs in a TBI subject. (Bottom
row, A) IFO—purple; ILF—black; HC/FX—blue. (Bottom row, B) Anterior corpus
callosum—blue. (Bottom row, C) Posterior corpus callosum—blue. The reduction of
tract counts in the TBI subject with respect to the control subject is obvious in these
ROIs.
tracts or pathways in a group of normal subjects, all normalized and co-registered to
the individual patient’s brain anatomy, which should be of particular significance to an
individual based interventional procedure such as neurosurgery.
After normalization and mapping of tracts from a control group onto the head
space of an individual, a given voxel in the individual should ideally show the same
101
ROI t score Pathway FA Effect size TC Effect size
Control TBI Control TBI
1 4.13 (0.91) CC-b 0.29 (0.08) 0.09 (0.02) 3.43 225.50 (111.24) 2 2.01
2 4.09 (0.82) CC-s 0.49 (0.07) 0.33 (0.05) 2.63 422.50 (136.31) 249 1.27
3 4.03 (0.93) CC-s 0.37 (0.08) 0.17 (0.08) 2.5 543.80 (141.35) 174 2.62
4 3.89 (0.53) ILF-L 0.46 (0.07) 0.26 (0.09) 2.48 277.10 (117.08) 8 2.3
5 3.79 (0.54) CC-g 0.35 (0.04) 0.23 (0.02) 3.79 316.20 (118.45) 317 -0.01
6 3.73 (0.56) CST 0.48 (0.05) 0.38 (0.05) 2 433.00 (151.25) 355 0.52
7 3.71 (0.62) CC-b 0.51 (0.03) 0.39 (0.01) 5.37 108.80 (94.95) 30 0.83
SLF-L 0.31 (0.07) 0.25 (0.07) 0.86 45.00 (41.46) 15 0.72
8 3.67 (0.47) SLF-L 0.49 (0.05) 0.36 (0.07) 2.14 536.00 (179.76) 177 2
9 3.58 (0.38) CC-s 0.47 (0.11) 0.16 (0.03) 3.85 658.50 (109.39) 206 4.14
10 3.54 (0.45) SLF-L 0.39 (0.05) 0.27 (0.03) 2.91 307.00 (146.48) 168 0.95
11 2.67 (0.59) HC/FX-R 0.33 (0.06) 0.27 (0.05) 1.09 117.40 (53.08) 78 0.74
ROI Pathway MD (x10
3
mm
2
/s) Effect size DA (x10
3
mm
2
/s) Effect size
Control TBI Control TBI
1 CC-b 0.65 (0.05) 0.92 (0.09) 3.71 0.85 (0.07) 1.00 (0.10) 1.74
2 CC-s 0.67 (0.03) 0.79 (0.03) 4 1.08 (0.09) 1.03 (0.05) -0.69
3 CC-s 0.72 (0.06) 0.85 (0.09) 1.7 1.02 (0.10) 0.99 (0.07) -0.35
4 ILF-L 0.68 (0.03) 0.83 (0.05) 3.64 1.06 (0.07) 1.05 (0.06) -0.15
5 CC-g 0.68 (0.04) 0.81 (0.02) 4.11 0.95 (0.05) 1.01 (0.04) 1.33
6 CST 0.62 (0.02) 0.69 (0.01) 4.43 0.98 (0.07) 1.00 (0.06) 0.31
7 CC-b 0.63 (0.03) 0.71 (0.02) 3.14 1.03 (0.04) 1.02 (0.02) -0.32
SLF-L 0.62 (0.05) 0.75 (0.14) 1.24 0.82 (0.06) 0.95 (0.11) 1.47
8 SLF-L 0.69 (0.04) 0.82 (0.03) 3.68 1.09 (0.09) 1.15 (0.08) 0.7
9 CC-s 0.89 (0.23) 0.88 (0.10) -0.06 1.38 (0.28) 1.04 (0.12) -1.58
10 SLF-L 0.73 (0.04) 0.88 (0.03) 4.24 1.07 (0.08) 1.14 (0.05) 1.05
11 HC/FX-R 0.83 (0.09) 0.93 (0.08) 1.17 1.12 (0.11) 1.19 (0.10) 0.67
ROI Pathway DR (x10
3
mm
2
/s) Effect size
Control TBI
1 CC-b 0.54 (0.07) 0.88 (0.09) 4.22
2 CC-s 0.47 (0.04) 0.66 (0.04) 4.75
3 CC-s 0.57 (0.07) 0.78 (0.11) 2.28
4 ILF-L 0.49 (0.06) 0.72 (0.07) 3.53
5 CC-g 0.55 (0.05) 0.72 (0.03) 4.12
6 CST 0.44 (0.02) 0.54 (0.02) 5
7 CC-b 0.43 (0.03) 0.55 (0.02) 4.71
SLF-L 0.52 (0.06) 0.66 (0.15) 1.23
8 SLF-L 0.48 (0.04) 0.66 (0.05) 3.98
9 CC-s 0.64 (0.23) 0.81 (0.10) 0.96
10 SLF-L 0.56 (0.04) 0.75 (0.04) 4.75
11 HC/FX-R 0.69 (0.10) 0.79 (0.08) 1.1
Table 10.3: List of affected pathways and diffusion metrics for 12 ROIs identified by
SPM group analysis as FA-reduced regions in the TBI group (n=12) compared to the
normal control group (n=10) (P
FDR
.05,k12)
number of tracts and the same pathways for any control subject. However, coregistration
is never perfect and one would also expect some level of intersubject variability in brain
connections. Indeed, in this work, we found that tracts sorted by a given ROI in the
TBI subject’s head space did not overlap completely among the 10 control subjects used
in this work. To mitigate intersubject variability, we refined the ROIs by subdividing
them into smaller ROIs representing regions where specific tracts from at least 50%
102
of the control subjects overlapped completely. All anisotropy and TC metrics were
subsequently computed within these sub-ROIs.
As specific brain pathways are correlated to specific brain function and behavior,
identification and quantification of damage to specific pathways due to TBI provide a
metric to correlate DTI findings to neurocognitive and other behavior/clinical test scores.
For example, consistent with previous studies [SMC
+
06, WBD
+
08], we observed that
the HC/FX pathway was frequently affected by TBI injury in the subjects reported here.
This observation in individuals was validated by the group study where the HC/FX
pathway was identified by one of the high t-score FA-reduced ROIs (ROI 2 in Table
10.3) as a commonly injured region in this cohort of TBI subjects.
Consistent with injury to the HC/FX pathways, many of the TBI subjects also
subjectively reported short-term memory loss during their clinical interviews. Future
work should investigate the correlation between neuro-psychological tests designed
to probe specific cognitive function, such as memory, and the DTI metrics including
anisotropy metrics and tract counts along specific pathways, such as HC/FX, to validate
the DTI findings. If validated, the DTI metrics would provide a unique and powerful
approach to objectively monitor TBI progression.
The ROIs used in this work derive from a significant reduction in FA between an
individual TBI subject and a group of age-matched controls. MD was found to increase
significantly in these regions. We also investigated whether there were regions in the
TBI subjects where FA had increased significantly with respect to the controls but were
unable to identify any such regions in our TBI cohort. As described in Section 10.2,
though many other investigators have reported a decrease in FA and an increase in MD
in injured regions, an increase in FA and a decrease in MD in acute TBI have also been
reported with plausible explanations for how FA could increase initially and decrease
subsequently following injury. However, the exact model of how FA or other anisotropy
103
metrics would change in TBI is not known and future studies are expected to shed more
light on the dynamics of FA changes in TBI.
In addition to FA, we investigated the difference maps for MD, DR and DA,
and found that except for FA, the other measures showed relatively diffuse and nonspe-
cific changes, distributed across the entire brain and appeared to affect almost all brain
pathways. This observation is consistent with a recent report [BRL
+
08]. As our key
objective in this work was to localize injured regions and specific pathways affected by
the injury, and FA changes were much better localized than the other diffusion mea-
sures, we chose to identify injured regions from their FA changes and then compute the
MD, DR and DA changes over these regions. A detailed study of all of these diffusion
measures would be desirable in the future.
It is interesting to note that though DR was significantly higher in TBI compared
to controls (which is expected as there would be more free space for water molecules to
diffuse radially in the absence of barriers posed by intact axons and their myelin sheath),
DA was also consistently higher in TBI than in normal subjects in our study. The ratio
DA/DR, however, was significantly lower in TBI than in control subjects. The increase
in DA in TBI is puzzling at first glance but is consistent with the observation that intact
axons and their myelin sheath would present some restriction to free diffusion along the
axial direction, which would be lifted following axonal degeneration in TBI, leading to
higher DA. The significantly higher values of the ratio DA/DR in controls compared to
TBI clearly imply that diffusion along the radial direction is much more restricted when
intact axons are present.
It is important to note that the number of tracts is a function of the FA and
deflection thresholds in streamline tractography and hence the TC metric computed here
is valid only for the FA threshold of 0.15 and deflection angle of<45°used in this work.
104
The FA threshold was determined objectively to maximize the TC difference-related
effect size between the TBI and controls.
Previous reports have suggested [JSCH05] that the distribution of FA (and other
diffusion parameters) is not strictly Gaussian. Thus the precise statistical significance
of the single subject as well as the group study is not established in this work. A non-
parametric comparison such as the permutation-based approaches [NH02] would lead
to more accurate statistical comparisons in future work.
Finally, the accuracy of DTI metrics including tract counts is ultimately limited
by the signal-to-noise ratio of the raw data; partial volume problems due to the limited
spatial resolution of the EPI images; susceptibility artifacts at high fields, micromove-
ment and motion artifacts; corrections for eddy currents; co-registration between the EPI
and highresolution anatomical images to identify anatomical regions; and the accuracy
of spatial normalization procedures. In addition, multiple fibers within a voxel present
challenges to quantification. It is very likely that several voxels contain multiple cross-
ing fibers, resulting in erroneous estimation of the DTI tensor’s eigen values and eigen
vectors that are based on a single rank-2 tensor. Errors in the eigen values would affect
the anisotropy metrics, and errors in the primary eigen vector would propagate tracts
incorrectly leading to early termination or misinterpretation. Work is in progress in our
lab and other institutions to reduce these errors.
10.6 Acknowledgment
The authors would like to thanks Ms. Amrita Rajagopalan and Mr. Sinchai Tsao for their
assistance with the tract normalization code. This research was supported by Contract
No. W81XWH-07-1-0015 awarded by USA Medical Research and Material Command,
TATRC, and partially supported by grant NIA-NIH P50 AG05142.
105
Chapter 11
Application 2: Alzheimer Disease
11.1 Introduction
Alzheimer Disease (AD) is a progressive neurodegenerative disease afflicting 5.4 mil-
lion Americans in 2012 in total, and 5.2 million aged 65 or older [Ass12]. With only 4%
of the affected population being under the age of 65, most AD research involves aged
populations.
The cause of AD is currently unknown. Two of the most common events
linked to AD are the formation of plaques and tangles. Plaques refer to build up of
-amyloid (A) protein in-between nerve cells which disrupt communication and may
trigger immune responses [HA91]. Tangles are malformations of tau proteins inside
the nerve cells which disrupt the microtubular transport system, leading to neuron death
[IdCAC
+
05]. Both phenomena should theoretically change the diffusion characteristics
in their affected areas. With a clinical scan b-value of 1000 sec/mm
2
, our DTI should
mostly sense the effects of neuronal fiber demylination or other physical fiber disrup-
tions.
The neuropathology of AD is believed to begin in the entorhinal cortex, spread to
the hippocampus, and then progress further to the rest of the brain [BB97]. Hippocampal
volume reduction has been observed in AD patients [CLT
+
97] and is one of the most
accurate imaging predictors of AD [JPOT92, JPX
+
99, DSA
+
01, SWB
+
09]. Much of
our AD research are concentrated on the entorhinal cortex, hippocampus, and the white
matter tracts connected to those structures, the fornix and cingulum.
106
11.2 Dataset
The University of Southern California Alzheimer Disease Research Center (USC-
ADRC) is an interdisciplinary group tasked with longitudinal studies focusing on
Alzheimer Disease and cerebrovascular disease. As part of the Imaging Core, under the
direction of Manbir Singh, we were tasked with developing brain imaging techniques
for use in disease study and with imaging support of investigators, with emphasis on
MRI systems. Since the passing of Dr. Singh, the USC Biomedical Imaging Lab is
continuing work with guidance from Helena Chui, Meng Law, and Natasha Lepore.
MR imaging data acquisition and subject categorization for the AD project
spanned from 2006 to 2009. Subject categorization was received in batches after the
MR scans were obtained.
11.2.1 Subjects
These scans were part of a comprehensive examination involving neurological tests
and neuropsychological tests, including the Mini Mental State Examination (MMSE)
and the Clinical Dementia Rating Scale (CDR), for participants in longitudinal studies
conducted by USC-ADRC under Institutional Review Board approval. Subjects were
categorized as Normal Control, Mild Cognitively Impaired, and Probably AD through
consensus diagnosis of USC-ADRC neurologists.
11.2.2 MR Scanners
MRI data was acquired on two 3T GE Signa HDx MRI scanners. From January 2006
to February 2008, images were acquired at USC University Hospital. In July 2008,
additional scans were acquired at Healthcare Consultation Center 2 (HCC2) Outpatient
107
Imaging Center. Scanners produced comparable data using the techniques highlighted
in Chapter 9.
11.2.3 MR Image Data
Diffusion scans used a multislice, twice refocused sequence with 128 x 128 single-shot
EPI (echo planar imaging) readout, 25 gradient directions with b = 1000 sec/mm
2
, one
b = 0 image, TR = 8300 msec, field of view (FOV) 26 cm, two averages (NEX=2), and 4
mm thick 28 contiguous axial slices covering the entire head with 2.04 x 2.04 x 4 mm
3
voxels in7 min for a total of 728 images. Although isotropic voxels are generally
preferable, 4 mm thick slices were acquired to shorten the data acquisition time by a
factor of 2 and reduce movement artifacts.
Anatomic scans were 3D T1-Weighted images with 256 x 256 SPoiled Gradient
Recalled (SPGR) readout, field of view (FOV) 25 cm, one average (NEX=1), and 1.5
mm thick coronal slices coving the entire head with 0.977 x 0.977 x 1.5 mm
3
voxels
for a total of 136 images (including one case of 150 images). In July 2006 the slice
thickness was reduced to 1 mm which lead to 0.977 x 0.977 x 1 mm
3
voxels for a total
of 252 images.
Images were received on compact disk, which were processed by custom Matlab
code to automatically parse the information into DICOM images grouped by scan type.
Scans were archived in original form on RAID 10 servers. A second copy was stored on
a custom password secured image archive, created by the author, using a PHP front end
and MySQL back end.
108
11.3 DTT Difference in AD Populations
A loss of coherence between the frontal and occipital lobes in AD was detected using
electroencephalography (EEG) [LNC
+
92], which can be interpreted as a disruption
in the fronto-occipital white matter tracts connecting the frontal and occipital lobes.
We applied normalized DTT to quantitate these changes and other tract-based changes
caused by AD.
11.3.1 Implementation
Data normalization on the AD study population began with the application of the FA
Template Normalization technique detailed in Section 5.2.1. For this study, 8 normal
controls were used to create the FA template.
Seed spacing was normalized using the template-based seed distribution (see
Section 6.1). Normalization parameters were the same as the ones used to warp the
FA maps to template space. Streamline tractography was conducted using NeuroTract
with the individually calculated seed distributions in the original diffusion acquisition
space. The tractography settings used include tensor interpolation, an FA minimum
value of .15, 45° maximum angle deflection, and Euler method numerical approximation
with a .2 mm step size. The tracts were then warped to template space using the tract
normalization method detailed in Chapter 7. Tracts were filtered using the ROI shown
in Figure 11.1 using the tract filtering algorithms discussed in Section 8.3.
11.3.2 Results
The resulting tracts (shown in Figure 11.2) show the same general trend of thinning with
AD. Table 11.1 displays the tract count for each subject corrected using ICV .
109
Figure 11.1: Fronto-Occipital ROIs. The ROIs are used pair-wise (shown by use of
different colors) to isolate tracts first in the left hemisphere, then in the right.
Right Hemisphere Left Hemisphere
Normal Prob AD Normal Prob AD
540.58 151.51 411.76 211.10
392.31 129.97 332.37 135.17
198.72 221.66 307.91 95.00
383.56 394.45 232.04 359.10
301.33 204.59 345.64 273.14
410.33 176.87 104.34 244.90
338.07 128.29 466.81 155.56
327.26 279.10
Standard Deviation 98.37 92.40 110.87 90.58
Mean 361.52 201.05 310.00 210.57
Effect Size 1.68 0.98
Table 11.1: Fronto-Occipital tract count for 8 normal control and 7 AD subjects
110
Figure 11.2: Fronto-Occipital Tracts. The images to the left are from a normal control
and the right are from a probable AD.
11.3.3 Discussion
The working hypothesis of the DTI work being done with respect to Alzheimer Dis-
ease is that disease effects in white matter tracts of the brain can be detected by DTI.
An initial study was conducted which found detectable changes in DTT between AD
and controls[SKWC04]. The results in Section 11.3.2 show significant changes in trac-
tography between normal and AD subjects using our new seed and tract normalization
methods. This preliminary work shows DTI to be a promising imaging method to track
the disease progression of AD.
111
11.4 Pathway Isolation Using Objective FA Difference
ROIs
Fractional Anisotropy (FA) can be seen as a measure of white matter integrity. With a
reduction in integrity, FA should decrease as a result of voxels becoming more isotropic
as there is less ordered restriction of diffusion. Using normalized tractography and FA
decrease areas as ROI, we are able to delineate the tracts affected by disease progres-
sion. This research was presented at ISMRM 17th Scientific Meeting and Exhibition
[HSS08].
11.4.1 Implementation
For this study, 10 normal control subjects were used to create the FA template used in
the FA Template Normalization technique detailed in section 5.2.1.
After creating the FA template, the FA maps from all subjects were normalized
to the FA template via SPM. The resulting FA maps of normal controls were then com-
pared against MCI and AD probable groups using two-value t-test statistical analysis in
SPM. Two comparisons were done: normal versus MCI (see Figure 11.3a) and normal
versus probable AD (see Figure 11.3b). With a threshold of p>0.005 and a minimum
voxel connection level of 8 voxels, the intersection of the clusters from the two groups
provided ROIs for tractographical filtering (see Figure 11.4). The ROIs were clustered
by connected voxels and numbered sequentially.
Seed spacing was normalized using the template-based seed distribution (see
Section 6.1). Normalization parameters from FA map to template space warping were
used to distribute the seeds. Streamline tractography was conducted using NeuroTract
with the individually calculated seed distributions in the original diffusion acquisition
space. The tractography settings used include tensor interpolation, an FA minimum
112
(a) Normal versus MCI.
(b) Normal versus probable AD.
Figure 11.3: Two value t-test results for FA where subject groups had lower FA than the
normal control group.
value of .15, 45° maximum angle deflection, and Euler method numerical approximation
with a .2 mm step size. The tracts were then warped to template space using the tract
normalization method detailed in Chapter 7. Tracts were filtered based on the ROIs
obtained by the FA group comparisons and using the ROI tract filtering method detailed
in Section 8.3.
11.4.2 Results
Tracts identified by this method include the fornix, fronto-occipital, temporal-occipital,
and parietal-temporal tracts.
113
(a) Axial view.
(b) Sagittal view.
Figure 11.4: Intersection between normal/probable AD and normal/MCI populations.
11.4.3 Discussion
FA has been studied as a possible metric to measure axon degeneration. The methods
described in Section 11.4.1 should end in the isolation of tracts known to be involved in
AD. The results in Section 11.4.2 identified fornix, fronto-occipital, temporal-occipital,
and parietal-temporal tracts as being affected by Alzheimer Disease, which is consistent
114
(a) Frontal-Occipital and fornix tracts
(b) Parietal-Temporal and fornix tracts
Figure 11.5: Tracts isolated from ROIs
with tracts identified through histology and other techniques. This gave us a working
methodology to find affected tracts: isolate ROIs of DTI metric change, and then use
those ROIs to find tracts. This concept was applied to traumatic brain injury and pre-
sented in the work detailed in Chapter 10.
115
11.5 Fornix Tract Count Using Objective Hippocampus
ROI
Numerous studies have shown a pattern of progressive hippocampal shrinkage as a result
of mild cognitive impairment (MCI) and Alzheimer Disease (AD) [JPOT92, JPX
+
99,
DSA
+
01, SWB
+
09]. The purpose of this study was to (a) objectively define hippocam-
pal ROIs as most previous studies rely on subjectively delineated ROIs to perform DTI
analysis, (b) correct for varying subject head sizes using the intracranial volume (ICV)
of subjects in a novel way to equalize the number of seed points and their anatomical dis-
tribution so that total number of tracts becomes independent of the subject’s head size,
and (c) examine the differences in the equalized streamline fornix tracts. This research
was presented at ISMRM 17th Scientific Meeting and Exhibition [HTS09].
11.5.1 Implementation
The right and left hippocampi were segmented using Freesurfer [FSB
+
02, FSvdK
+
04,
FdKD
+
04]. The ROIs were then co-registered to the diffusion scans using SPM. By
using Freesurfer and SPM to segment the hippocampus, the ROI delineation process
was semi-automated and more objective than manual delineation. The standard method
of hippocampal ROI delineation is subjective human outlining, which is both time con-
suming and requires expert intervention.
Seed spacing was normalized using the ICV-based seed distribution (see sec-
tion 6.2). The ICV was calculated for each subject and seeds were normalized to a
standard value of a voxel size 2 x 2 x 2 mm for an ICV of 1:5x10
6
mm
3
. Streamline
tractography was conducted using NeuroTract with the individually calculated seed dis-
tributions. The tractography settings used include tensor interpolation, an FA minimum
116
value of .05, 45° maximum angle deflection, and Euler method numerical approximation
with a .2 mm step size.
11.5.2 Results
The hippocampal volume segmentation shows a decrease in volume indicative of the
progressive degeneration associated with MCI and AD (11.6a). The fornix is the major
tract associated with the hippocampus; the effects of MCI and AD on the fornix are
relatively unknown. The equalized tract count shows a measurable change between the
normal controls and the two conditions of MCI and AD (11.6b). Visually, the fornix
(a) Hippocampal volume
(b) Fornix tract count
Figure 11.6: Study trend charts
tracts show a degenerative pattern. As seen in Figure 11.7, the tracts show a noticible
decrease in density from normal to MCI to probable AD.
117
(a) Normal control (b) MCI
(c) Probable AD
Figure 11.7: Progressive degeneration of the fornix tract.
11.5.3 Discussion
In section 11.5.2, the hippocampal volume segmentation showed a decrease in volume
indicative of the progressive degeneration associated with MCI and AD (Figure 11.6a).
The fornix is the major tract associated with the hippocampus. The effects of MCI and
AD in the fornix are relatively unknown. The equalized tract count shows a measurable
change between normal controls and the conditions of MCI and AD (Figure 11.6b).
Visually, the fornix tracts show a degenerative pattern. As seen in Figure 11.7, the tracts
show a noticeable decrease in density from normal, to MCI, to probable AD populations.
11.6 Identification of Damaged Tracts Using Improved
Normalization
Once the dataset was gathered and diagnoses were received, we repeated the study
detailed in Section 11.4 with more subjects and an improved normalization procedure.
118
11.6.1 Implementation
For this study, 10 normal control subjects were used to create the DARTEL template
used in the DARTEL Template Normalization technique detailed in Section 5.2.2.
After obtaining the concatenated transform, all FA and MD maps from all sub-
jects were normalized to the MNI template. The resulting normal control FA and MD
maps and the two affected groups of MCI and AD probable were then compared using
two-value t-test statistical analysis in SPM. Two comparisons were done: normal ver-
sus MCI (see Figure 11.8a) and normal versus probable AD (see Figure 11.8b). With a
threshold of p>0.005 and a minimum voxel connection level of 8 voxels.
(a) Normal versus MCI.
(b) Normal versus probable AD.
Figure 11.8: Two value t-test results for FA where subject groups had lower FA than the
normal control group.
119
11.6.2 Results
The SPM8 DARTEL suite provided a much improved normalization of the aged brain
populations compared to the previous FA template normalization workflow (Section
5.2.1). Percentage overlap of filtered fornix and cingulum tracts confirmed greatly
improved alignment when compared to previous results (Section 11.4). ROIs gener-
ated by the procedure were non-specific resulting little to no correlation with known
disease state.
11.6.3 Discussion
Section 11.6.2 yielded nonspecific results. It is possible that a combination of poor
resolution and smaller atrophied brains led to ambiguous changes in FA. We believe
that FA change signaling damage could be either positive or negative. FA increase as a
sign of neuronal disruption has recently been observed by others [DJB
+
11]. This effect
is postulated as coming as a result of crossing fibers where one fiber is disrupted, leading
to the overall increase in a voxel’s anisotropy. The ambiguity of FA change may have
led to a dilution of statistical power when looking at changes between disease states.
11.7 Isolation of Damage Along the Fornix and Cingu-
lum Tracts
With the lack of specificity from the results of Section 11.6.2, we decided to approach
the problem from another perspective. Given a predefined tract, we can look for damage
along it that may result from AD. The fornix and cingulum (limbic regions) are sus-
pected of being affected by mild cognitive impairment (MCI) and Alzheimer Disease
(AD). Most studies resort to manual tracing of the anatomy in patient acquisition space.
120
Objective isolation of the fornix and cingulum has remained elusive. The purpose of
this study was to (a) bring DTI including tractography to a common template space, (b)
define the fornix and cingulum objectively using tractography, and (c) examine the FA
changes among normal control, MCI, and probable AD populations. This research was
presented at ISMRM 18th Scientific Meeting and Exhibition [HTS10].
11.7.1 Implementation
For this study, 10 normal control subjects were used to create the DARTEL template
used in the DARTEL Template Normalization technique detailed in Section 5.2.2.
A single set of ROIs were defined in MNI space and applied to the normal con-
trol group whole-brain tractography to isolate the fornix and cingulum. Four processing
masks were defined (left fornix, right fornix, left cingulum, and right cingulum) contain-
ing voxels where tracts overlapped in at least 75% of the control group. A voxel-based
comparison of the t-score statistics among the FA map difference of the three different
populations was conducted within the four masks to highlight changes observed in the
fornix and cingulum tracts.
11.7.2 Results
The SPM8 DARTEL suite provided a much improved normalization of the aged brain
populations compared to the previous FA template normalization workflow (Section
5.2.1). Visual inspection of the filtered fornix and cingulum tracts confirmed greatly
improved alignment when compared to previous results (Section 11.4). The t-score
comparisons (Figure 11.9 and 11.10) visually illustrate areas where FA in one population
is higher than the other.
121
Figure 11.9: N>MCI: Fornix (top row) and cingulum (bottom row) FA t-Score Com-
parison. Colors denote the t-score difference between the mean FA of the various popu-
lations (blue=0, red=3", line marksp<0.005 on the color scale).
11.7.3 Discussion
Section 11.7.2 shows a progressive degeneration of the fornix and the cingulum. The
right and left fornix branches both exhibit a pattern of damage which begins at the hip-
pocampal end of the fornix and propagates toward the hypothalamus with right reduc-
tion greater than left reduction in FA. The cingulum shows less damage, but does appear
to show progressive damage extending to the posterior end, more on the right than on
the left side. The complete posterior cingulum portion could be reliably identified in
the common space, suggesting the DARTEL-based normalization made vast improve-
ments in normalization of AD as our initial attempts yeiled the cingulum in several
pieces. These preliminary results show promise in tracking axonal damage with the
122
Figure 11.10: MCI>AD: Fornix (top row) and cingulum (bottom row) FA t-Score Com-
parison. Colors denote the t-score difference between the mean FA of the various popu-
lations (blue=0, red=3", line marksp<0.005 on the color scale).
progression of Alzheimer Disease. The stronger right reduction than left reduction of
FA in the fornix and cingulum is consistent with reports of white matter atrophy and
hypometabolism in AD [VDV
+
08].
123
Chapter 12
Conclusion and Future Work
This dissertation has been focused on diffusion tensor imaging and tractography, which
are quickly-maturing imaging methods just that are just beginning to acquire clini-
cal relevance. We have created new tools with researchers in mind. In addition, we
have applied these tools to study two neurological deficits: Traumatic Brain Injury and
Alzheimer Disease.
12.1 Original Contributions
12.1.1 USC Biomedical Imaging Lab DTI Software Suite
We have created a DTI software suite to process and visualize DTI and DTT data. Neu-
roTract is the DTI and tractography generation program. The original version of the
program, Ntrack, was written by Sungheon Kim during his time in the USC-BMIL.
NeuroTract was completely redesigned to accept new features such as ICA tractography
[SW06, SW07, SJ09, SW10]. ROISelector was created to easily mark ROIs in three
dimensional space. Visualization is handled by TractRender, which is a flexible tool
for researchers to create high-resolution images and output video-ready image stacks.
All three are written in Matlab and will be made publicly available in the near future.
Additional software for tract filtering and tract editing will be included as utility tools
for working with tractography.
124
12.1.2 DTT Normalization
Different populations of subjects, such as the elderly or our youthful adult TBI subjects,
have required different degrees of warping flexibility. We created two new normalization
procedures which allow diffusion data to be warped into a common space. Using the
output of these procedures, we can normalize the seed start points for tractography.
In addition, we have created a novel method of warping tracts from acquisition space
to template space, all the while maintaining the continuity of tracts generated in the
non-interpolated subject space. These algorithms form the foundation of quantitative
tractography.
12.1.3 DTI Metrics
We have proposed new metrics to quantify tractography. Tract count is used in con-
junction with normalized seed distribution as a measure of tract integrity. Tract-length
histograms provide an objective manner of choosing between different scan parameters
when designing protocols before long-term studies commence. These metrics are the
beginnings of quantitative tractography.
12.1.4 Human Studies
We applied our techniques on human datasets to explore the changes which occur as a
result of injuries and disease. We discovered patterns of difference between the afflicted
and healthy groups, which suggest the feasibility of using DTI and DTT to examine
changes in Traumatic Brain Injury and Alzheimer Disease. Our research adds to the
understanding of how diffusion imaging can be used to study both disorders.
125
12.2 Future Work
Through all this work, we found that partial volume effects were an important confound
for DTT analysis. Areas isolated as being damaged in our DTI metric comparisons could
also be interpreted as areas of crossing fiber. We addressed this issue in our laboratory
with the adoption of ICA tractography [SW06, SW07, SW09, SW10]. ICA tractography
uses independent component analysis to separate multiple diffusion signals from cross-
ing fibers. The method only recovers fiber direction and not full tensor information. We
are also investigating the separation of a CSF component from white matter voxels near
the ventricles [THS10] to further correct for partial volume effects.
The fundamental single diffusion tensor model can be formulated as Equation
12.1.
S
S
0
=e
bD
(12.1)
where D is the diffusion tensor and b is the gradient weighting of the diffusion scan.
When there is more than one fiber, this model fails to differentiate between multiple
fibers. To address this fundamental error in model selection, we can change the model
to Equation 12.2
S
S
0
=fe
bD
1
+ (1f)e
bD
2
(12.2)
wheref is a weighting fraction, b is the diffusion gradient weighting, andD
1
andD
2
are the diffusion tensors for two fibers. Alternatively, two fiber voxels could also be
modeled as Equation 12.3.
S
S
0
=e
b(fD
1
+(1f)D
2
)
(12.3)
Both models have been proposed for use in multi-tensor calculation [AHL
+
01]. Equa-
tion 12.3 is a representation of fast intracellular diffusion, while equation 12.2 models
126
slower extracellular diffusion, which is the range of diffusion detected by the clinical
sequences used in the USC Biomedical Imaging Lab.
We believe that a non-linear fitting solution is possible using the ICA calculated
principal direction as a starting point, and this is one of the new topics for my continuing
research.
127
Reference List
[3DS12] 3DSlicer. Screenshot of 3D Slicer. http://www.slicer.org/
slicerWiki/images/b/b6/Roi_tract.jpg, Accessed on Aug
18, 2012.
[AAF00] J. Ashburner, J.L. Andersson, and K.J. Friston. Image registration using
a symmetric prior–in three dimensions. Hum Brain Mapp, 9(4):212–225,
Apr 2000.
[ACWPN09] J. Acosta-Cabronero, G.B. Williams, G. Pengas, and P.J. Nestor. Diffu-
sivity changes are predominantly proportional along all axes with early
neurodegeneration in alzheimer’s disease. In Proceedings 17th Scien-
tific Meeting, International Society for Magnetic Resonance in Medicine,
Honolulu, page 732, 2009.
[ACWPN10] J. Acosta-Cabronero, G.B. Williams, G. Pengas, and P.J. Nestor. Abso-
lute diffusivities define the landscape of white matter degeneration in
alzheimer’s disease. Brain, 133(2):529–539, 2010.
[AF97] J. Ashburner and K.J. Friston. Multimodal image coregistration and
partitioning–a unified framework. Neuroimage, 6(3):209–217, Oct 1997.
[AF99] J. Ashburner and K.J. Friston. Nonlinear spatial normalization using basis
functions. Hum Brain Mapp, 7(4):254–266, 1999.
[AF05] J. Ashburner and K.J. Friston. Unified segmentation. Neuroimage,
26(3):839–851, Jul 2005.
[AFRB04] Y . Assaf, R.Z. Freidlin, G.K. Rohde, and P. J. Basser. New modeling
and experimental framework to characterize hindered and restricted water
diffusion in brain white matter. Magn Reson Med, 52(5):965–978, Nov
2004.
[AGMS82] J.H. Adams, D.I. Graham, L.S. Murray, and G. Scott. Diffuse axonal
injury due to nonmissile head injury in humans: an analysis of 45 cases.
Ann Neurol, 12(6):557–563, Dec 1982.
128
[AHC
+
02] K. Arfanakis, V .M. Haughton, J.D. Carew, B.P. Rogers, R.J. Dempsey,
and M.E. Meyerand. Diffusion tensor MR imaging in diffuse axonal
injury. AJNR Am J Neuroradiol, 23(5):794–802, May 2002.
[AHL
+
01] A.L. Alexander, K.M. Hasan, M. Lazar, J.S. Tsuruda, and D.L. Parker.
Analysis of partial volume effects in diffusion-tensor MRI. Magn Reson
Med, 45(5):770–780, May 2001.
[And01] A.W. Anderson. Theoretical analysis of the effects of noise on diffusion
tensor imaging. Magn Reson Med, 46(6):1174–1188, Dec 2001.
[APBG01] D.C. Alexander, C. Pierpaoli, P.J. Basser, and J.C. Gee. Spatial transfor-
mations of diffusion tensor magnetic resonance images. IEEE Trans Med
Imaging, 20(11):1131–1139, Nov 2001.
[Asc12] Team Asclepios. Screenshot of MedINRIA. http://raweb.
inria.fr/rapportsactivite/RA2007/asclepios/1.png,
Accessed on Aug. 18, 2012.
[Ash07] J. Ashburner. A fast diffeomorphic image registration algorithm. Neu-
roimage, 38(1):95–113, Oct 2007.
[ASM
+
04] D. Akers, A. Sherbondy, R. Mackenzie, R. Dougherty, and B. Wandell.
Exploration of the brain’s white matter pathways with dynamic queries.
In Visualization, 2004. IEEE, pages 377–384, Oct 2004.
[Ass12] Alzheimer’s Association. 2012 Alzheimer’s disease facts and figures.
Alzheimers Dement, 8(2):131–168, Mar 2012.
[ATP97] A.L. Alexander, J.S. Tsuruda, and D.L. Parker. Elimination of eddy cur-
rent artifacts in diffusion-weighted echo-planar images: the use of bipolar
gradients. Magn Reson Med, 38(6):1016–1021, Dec 1997.
[BA00] M.E. Bastin and P.A. Armitage. On the use of water phantom images
to calibrate and correct eddy current induced artefacts in MR diffusion
tensor imaging. Magn Reson Imaging, 18(6):681–687, Jul 2000.
[Bas99] M. E. Bastin. Correction of eddy current-induced artefacts in diffusion
tensor imaging using iterative cross-correlation. Magn Reson Imaging,
17(7):1011–1024, Sep 1999.
[Bas01] M.E. Bastin. On the use of the FLAIR technique to improve the correction
of eddy current induced artefacts in MR diffusion tensor imaging. Magn
Reson Imaging, 19(7):937–950, Sep 2001.
129
[BB97] H. Braak and E. Braak. Staging of Alzheimer-related cortical destruction.
Int Psychogeriatr, 9 Suppl 1:257–61; discussion 269–72, 1997.
[BJ02] P.J. Basser and D.K. Jones. Diffusion-tensor MRI: theory, experimental
design and data analysis - a technical review. NMR Biomed, 15(7–8):456–
467, 2002.
[BMLB94] P.J. Basser, J. Mattiello, and D. Le Bihan. MR diffusion tensor spec-
troscopy and imaging. Biophys J, 66(1):259–267, Jan 1994.
[BMTLB93] P.J. Basser, J. Mattiello, R. Turner, and D. Le Bihan. Diffusion tensor
echo-planar imaging of human brain. In Proceedings of the SMRM, page
584, 1993.
[BP00] P.J. Basser and S. Pajevic. Statistical artifacts in diffusion tensor MRI
(DT-MRI) caused by background noise. Magn Reson Med, 44(1):41–50,
Jul 2000.
[BP06] A. Bki and J.T. Povlishock. All roads lead to disconnection?–Traumatic
axonal injury revisited. Acta Neurochir (Wien), 148(2):181–193; discus-
sion 193–194, Feb 2006.
[BPP
+
00] P.J. Basser, S. Pajevic, C. Pierpaoli, J. Duda, and A. Aldroubi. In vivo
fiber tractography using DT-MRI data. Magn Reson Med, 44(4):625–632,
Oct 2000.
[BRL
+
08] B.B. Bendlin, M.L. Ries, M. Lazar, A.L. Alexander, R.J. Dempsey,
H.A. Rowley, J.E. Sherman, and S.C. Johnson. Longitudinal changes
in patients with traumatic brain injury assessed with diffusion-tensor and
volumetric imaging. Neuroimage, 42(2):503–514, Aug 2008.
[BWJ
+
03] T.E.J. Behrens, M.W. Woolrich, M. Jenkinson, H. Johansen-Berg, R.G.
Nunes, S. Clare, P.M. Matthews, J.M. Brady, and S.M. Smith. Characteri-
zation and propagation of uncertainty in diffusion-weighted MR imaging.
Magnetic Resonance in Medicine, 50(5):1077–1088, 2003.
[BZB
+
07] J.J. Bazarian, J. Zhong, B. Blyth, T. Zhu, V . Kavcic, and D. Peterson.
Diffusion tensor imaging detects clinically important axonal damage after
mild traumatic brain injury: a pilot study. J Neurotrauma, 24(9):1447–
1459, Sep 2007.
[CBS06] J.D. Clayden, M.E. Bastin, and A.J. Storkey. Improved segmentation
reproducibility in group tractography using a quantitative tract similarity
measure. Neuroimage, 33(2):482–492, Nov 2006.
130
[CGS06] B. Chen, H. Guo, and A.W. Song. Correction for direction-dependent dis-
tortions in diffusion tensor imaging using matched magnetic field maps.
Neuroimage, 30(1):121–129, Mar 2006.
[CLC
+
99] T.E. Conturo, N.F. Lori, T.S. Cull, E. Akbudak, A.Z. Snyder, J.S. Shi-
mony, R.C. McKinstry, H. Burton, and M.E. Raichle. Tracking neuronal
fiber pathways in the living human brain. Proc Natl Acad Sci U S A,
96(18):10422–10427, Aug 1999.
[CLM
+
06] S. Correia, S. Lee, P. Malloy, N. Mehta, S. Zhang, S. Salloway, and D.H.
Laidlaw. Diffusion-tensor MRI tractography methods for assessing white
matter health and its relationship to cognitive functioning. In Proceedings
of the International Neurophisiological Society, Boston, February 2006.
http://www.cs.brown.edu/research/vis/docs/doc/Correia-2006-DTM.doc.
[CLT
+
97] A. Convit, M.J. De Leon, C. Tarshish, S. De Santi, W. Tsui, H. Rusinek,
and A. George. Specific hippocampal volume reductions in individuals
at risk for Alzheimer’s Disease. Neurobiology of Aging, 18(2):131–138,
1997.
[CMB
+
05] J.D. Clayden, D.K.S. Marjoram, M.E. Bastin, E.C. Johnstone, and S.M.
Lawrie. Towards an automated method for white matter integrity compar-
ison between populations. In Proceedings of the ESMRMB 22nd Annual
Meeting, page 508, Basel, Sqitzerland, 2005.
[CMF
+
10] E. Canu, D.G. McLaren, M.E. Fitzgerald, B.B. Bendlin, G. Zoccatelli,
F. Alessandrini, F.B. Pizzini, G.K. Ricciardi, A. Beltramello, S.C. John-
son, and G.B. Frisoni. Microstructural diffusion changes are independent
of macrostructural volume loss in moderate to severe Alzheimer’s disease.
J Alzheimers Dis, 19(3):963–976, 2010.
[Cox96] R.W. Cox. AFNI: software for analysis and visualization of functional
magnetic resonance neuroimages. Comput Biomed Res, 29(3):162–173,
Jun 1996.
[CPGC99] F. Calamante, D.A. Porter, D.G. Gadian, and A. Connelly. Correction for
eddy current induced B
o
shifts in diffusion-weighted echo-planar imag-
ing. Magn Reson Med, 41(1):95–102, Jan 1999.
[CWG
+
07] X. Chen, D. Weigel, O. Ganslandt, M. Buchfelder, and C. Nimsky. Diffu-
sion tensor imaging and white matter tractography in patients with brain-
stem lesions. Acta Neurochir (Wien), 149(11):1117–1131; discussion
1131, Nov 2007.
131
[DCM
+
08] C. DeCarli, O. Carmichael, D. Mungas, B. Reed, O. Martinez,
M. Perisianinova, M. Ortega, and E. Fletcher. Ic-p1-017: Evidence
for transynaptic degeneration of fornix fibers in Alzheimer’s disease.
Alzheimer’s and Dementia, 4(4, Supplement 1):T15, 2008. Alzheimer’s
Association International Conference on Alzheimer’s Disease.
[DHF
+
05] D. Ducreux, I. Huynh, P. Fillard, J. Renoux, M.C. Petit-Lacour,
K. Marsot-Dupuch, and P. Lasjaunias. Brain MR diffusion tensor imag-
ing and fibre tracking to differentiate between two diffuse axonal injuries.
Neuroradiology, 47(8):604–608, Aug 2005.
[DJB
+
11] G. Douaud, S. Jbabdi, T.E.J. Behrens, R.A. Menke, A. Gass, A.U. Mon-
sch, A. Rao, B. Whitcher, G. Kindlmann, P.M. Matthews, and S. Smith.
Dti measures in crossing-fibre areas: Increased diffusion anisotropy
reveals early white matter alteration in MCI and mild Alzheimer’s dis-
ease. NeuroImage, 55(3):880–890, 2011.
[DLNC94] J.J. Dunkin, A.F. Leuchter, T.F. Newton, and I.A. Cook. Reduced
EEG coherence in dementia: State or trait marker? Biol Psychiatry,
35(11):870–879, Jun 1994.
[DPB
+
07] J. Dauguet, S. Peled, V . Berezovskii, T. Delzescaux, S.K. Warfield,
R. Born, and C.-F. Westin. Comparison of fiber tracts derived from in-vivo
DTI tractography with 3D histological neural tract tracer reconstruction
on a macaque brain. Neuroimage, 37(2):530–538, Aug 2007.
[DSA
+
01] A.T. Du, N. Schuff, D. Amend, M.P. Laakso, Y .Y . Hsu, W.J. Jagust,
K. Yaffe, J.H. Kramer, B. Reed, D. Norman, H.C. Chui, and M.W. Weiner.
Magnetic resonance imaging of the entorhinal cortex and hippocampus in
mild cognitive impairment and Alzheimer’s disease. Journal of Neurol-
ogy, Neurosurgery & Psychiatry, 71(4):441–447, 2001.
[FAK
+
07] K.J. Friston, J. Ashburner, S.J. Kiebel, T.E. Nichols, and W.D. Penny, edi-
tors. Statistical Parametric Mapping: The Analysis of Functional Brain
Images. Academic Press, 2007.
[FdKD
+
04] B. Fischl, A. der Kouwe, C. Destrieux, E. Halgren, F. S´ egonne, D.H.
Salat, E. Busa, L.J. Seidman, J. Goldstein, D. Kennedy, V . Caviness,
N. Makris, B. Rosen, and A.M. Dale. Automatically parcellating the
human cerebral cortex. Cereb Cortex, 14(1):11–22, Jan 2004.
[FSB
+
02] B. Fischl, D.H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove,
A. van der Kouwe, R. Killiany, D. Kennedy, S. Klaveness, A. Montillo,
132
N. Makris, B. Rosen, and A.M. Dale. Whole brain segmentation: auto-
mated labeling of neuroanatomical structures in the human brain. Neuron,
33(3):341–355, Jan 2002.
[FSvdK
+
04] B. Fischl, D.H. Salat, A.J.W. van der Kouwe, N. Makris, F. S´ egonne, B.T.
Quinn, and A.M. Dale. Sequence-independent segmentation of magnetic
resonance images. Neuroimage, 23 Suppl 1:S69–S84, 2004.
[FTP06] P. Fillard, N. Toussaint, and X. Pennec. MedINRIA: DT-MRI processing
and visualization software., Nov 2006. Guest paper at the Similar Tensor
Workshop, Las Palmas, Spain.
[GBA
+
11] E. Garyfallidis, M. Brett, B. Amirbekian, C. Nguyen, F.-C. Yeh,
E. Olivetti, Y . Halchenko, and I. Nimmo-Smith. Dipy–a novel software
library for diffusion MR and tractography. In HBM 2011: 17th Annual
Meeting of the Organization for Human Brain Mapping, Qubec City,
2011.
[GFGG08] C.B. Goodlett, P.T. Fletcher, J.H. Gilmore, and G. Gerig. Group analysis
of DTI fiber tract statistics with application to neurodevelopment. Neu-
roimage, Nov 2008.
[GGT88] L. R. Gentry, J. C. Godersky, and B. Thompson. MR imaging of head
trauma: review of the distribution and radiopathologic features of trau-
matic lesions. AJR Am J Roentgenol, 150(3):663–672, Mar 1988.
[GMB
+
05] Kevin M. Guskiewicz, Stephen W. Marshall, J. Bailes, M. McCrea, R.C.
Cantu, C. Randolph, and B.D. Jordan. Association between recurrent con-
cussion and late-life cognitive impairment in retired professional football
players. Neurosurgery, 57(4):719–726, 2005.
[GSA
+
05] R.K. Gupta, S. Saksena, A. Agarwal, K.M. Hasan, M. Husain, V . Gupta,
and P.A. Narayana. Diffusion tensor imaging in late posttraumatic
epilepsy. Epilepsia, 46(9):1465–1471, Sep 2005.
[GVB05] E. Geuze, E. Vermetten, and J. D. Bremner. MR-based in vivo hippocam-
pal volumetrics: 2. findings in neuropsychiatric disorders. Mol Psychia-
try, 10(2):160–184, Feb 2005.
[HA91] J. Hardy and D. Allsop. Amyloid deposition as the central event in the
aetiology of Alzheimer’s disease. Trends in Pharmacological Sciences,
12(0):383–388, 1991.
[HBTV99] E.M. Haacke, R.W. Brown, M.R. Thompson, and R. Venkatesan. Mag-
netic Resonance Imaging: Physical Principles and Sequence Design.
John Wiley & Sons, 1999.
133
[HF07] X. Han and B. Fischl. Atlas renormalization for improved brain MR
image segmentation across scanner platforms. IEEE Trans Med Imaging,
26(4):479–486, Apr 2007.
[HKI
+
09] K.M. Hasan, A. Kamali, A. Iftikhar, L.A. Kramer, A.C. Papanicolaou,
J.M. Fletcher, and L. Ewing-Cobbs. Diffusion tensor tractography quan-
tification of the human corpus callosum fiber pathways across the lifes-
pan. Brain Res, 1249:91–100, Jan 2009.
[HMAT04] R.A. Hurley, J.C. McGowan, K. Arfanakis, and K.H. Taber. Traumatic
axonal injury: novel insights into evolution and identification. J Neu-
ropsychiatry Clin Neurosci, 16(1):1–7, 2004.
[HMT
+
08] C.W. Hoge, D. McGurk, J.L. Thomas, A.L. Cox, C.C. Engel, and C.A.
Castro. Mild traumatic brain injury in U.S. soldiers returning from Iraq.
New England Journal of Medicine, 358(5):453–463, 2008.
[Hor99] M.A. Horsfield. Mapping eddy current induced fields for the correc-
tion of diffusion-weighted echo planar images. Magn Reson Imaging,
17(9):1335–1345, Nov 1999.
[HPA01] K.M. Hasan, D.L. Parker, and A.L. Alexander. Comparison of gradi-
ent encoding schemes for diffusion-tensor MRI. J Magn Reson Imaging,
13(5):769–780, May 2001.
[HSRS07a] D.H. Hwang, A. Shetty, A. Rajagopalan, and M. Singh. Reduction of
partial volume artifacts in DTI tractography by post-processing. In Pro-
ceedings 15th Scientific Meeting, International Society for Magnetic Res-
onance in Medicine, Berlin, page 1547, 2007.
[HSRS07b] D.H. Hwang, A. Shetty, A. Rajagopalan, and M. Singh. Retrospective
processing of DTI tractography to compensate for partial volume effects.
In Armando Manduca and Xiaoping P. Hu, editors, Proceedings of SPIE,
volume 6511, page 651124. SPIE, 2007.
[HSS
+
04] T.A.G.M. Huisman, L.H. Schwamm, P.W. Schaefer, W.J. Koroshetz,
N. Shetty-Alva, Y . Ozsunar, O. Wu, and A.G. Sorensen. Diffusion tensor
imaging as potential biomarker of white matter injury in diffuse axonal
injury. AJNR Am J Neuroradiol, 25(3):370–376, Mar 2004.
[HSS08] D.H. Hwang, W. Sungkarat, and M. Singh. Location of affected pathways
in MCI and AD through FA comparison. In Proceedings 16th Scien-
tific Meeting, International Society for Magnetic Resonance in Medicine,
Toronto, page 2010, 2008.
134
[HTS09] D.H. Hwang, S. Tsao, and M. Singh. Quantification of fornix tracts in
MCI and AD. In Proceedings 17th Scientific Meeting, International Soci-
ety for Magnetic Resonance in Medicine, Honolulu, page 1181, 2009.
[HTS10] D.H. Hwang, S. Tsao, and M. Singh. Comparison of limbic regions
FA using tractography-defined ROIs in AD and MCI. In Proceedings
18th Scientific Meeting, International Society for Magnetic Resonance in
Medicine, Stockholm, page 2428, 2010.
[HZW
+
08] K. Hua, J. Zhang, S. Wakana, H. Jiang, X. Li, D.S. Reich, P.A. Calabresi,
J.J. Pekar, P.C.M. van Zijl, and S. Mori. Tract probability maps in stereo-
taxic spaces: analyses of white matter anatomy and tract-specific quan-
tification. Neuroimage, 39(1):336–347, Jan 2008.
[IdCAC
+
05] K. Iqbal, A. del C. Alonso, S. Chen, M.O. Chohan, E. El-Akkad,
C.-X. Gong, S. Khatoon, B. Li, F. Liu, A. Rahman, H. Tanimukai,
and I. Grundke-Iqbal. Tau pathology in Alzheimer disease and other
tauopathies. Biochimica et Biophysica Acta (BBA) - Molecular Basis of
Disease, 1739:198–210, 2005.
[IMJ
+
05] M. Inglese, S. Makani, G. Johnson, B.A. Cohen, J.A. Silver, O. Gonen,
and R.I. Grossman. Diffuse axonal injury in mild traumatic brain injury: a
diffusion tensor imaging study. J Neurosurg, 103(2):298–303, Aug 2005.
[JBP98] P. Jezzard, A.S. Barnett, and C. Pierpaoli. Characterization of and cor-
rection for eddy current artifacts in echo planar diffusion imaging. Magn
Reson Med, 39(5):801–812, May 1998.
[JGA
+
02] D.K. Jones, L.D. Griffin, D.C. Alexander, M. Catani, M.A. Horsfield,
R. Howard, and S.C.R. Williams. Spatial normalization and averaging of
diffusion tensor MRI data sets. Neuroimage, 17(2):592–617, Oct 2002.
[JHS99] D.K. Jones, M.A. Horsfield, and A. Simmons. Optimal strategies for mea-
suring diffusion in anisotropic systems by magnetic resonance imaging.
Magn Reson Med, 42(3):515–525, Sep 1999.
[JJA
+
00] V . Jelic, S.-E. Johansson, O. Almkvist, M. Shigeta, P. Julin, A. Nordberg,
B. Winblad, and L.-O. Wahlund. Quantitative electroencephalography in
mild cognitive impairment: longitudinal changes and possible prediction
of Alzheimer’s disease. Neurobiology of Aging, 21(4):533–540, 2000.
[Jon04] D.K. Jones. The effect of gradient sampling schemes on measures derived
from diffusion tensor MRI: a Monte Carlo study. Magn Reson Med,
51(4):807–815, Apr 2004.
135
[JPOT92] C.R. Jack, R.C. Petersen, P.C. O’Brien, and E.G. Tangalos. MR-based
hippocampal volumetry in the diagnosis of Alzheimer’s disease. Neurol-
ogy, 42(1):183–188, Jan 1992.
[JPX
+
99] C.R. Jack, R.C. Petersen, Y .C. Xu, P.C. O’Brien, G.E. Smith, R.J. Ivnik,
B.F. Boeve, S.C. Waring, E.G. Tangalos, and E. Kokmen. Prediction of
AD with MRI-based hippocampal volume in mild cognitive impairment.
Neurology, 52(7):1397–1403, 1999.
[JSCH05] D.K. Jones, M.R. Symms, M. Cercignani, and R.J. Howard. The effect
of filter size on VBM analyses of DT-MRI data. Neuroimage, 26(2):546–
554, Jun 2005.
[JvZK
+
06] H. Jiang, P.C.M. van Zijl, J. Kim, G.D. Pearlson, and S. Mori. DtiStu-
dio: resource program for diffusion tensor computation and fiber bundle
tracking. Comput Methods Programs Biomed, 81(2):106–116, Feb 2006.
[KAA
+
09] A. Klein, J. Andersson, B.A. Ardekani, J. Ashburner, B. Avants, M.-C.
Chiang, G.E. Christensen, D.L. Collins, J. Gee, P. Hellier, J.H. Song,
M. Jenkinson, C. Lepage, D. Rueckert, P. Thompson, T. Vercauteren, R.P.
Woods, J.J. Mann, and R.V . Parsey. Evaluation of 14 nonlinear deforma-
tion algorithms applied to human brain MRI registration. Neuroimage,
46(3):786–802, Jul 2009.
[KIN
+
07] A. Kunimatsu, D. Itoh, Y . Nakata, N. Kunimatsu, S. Aoki, Y . Masu-
tani, O. Abe, M. Yoshida, M. Minami, and K. Ohtomo. Utilization of
diffusion tensor tractography in combination with spatial normalization
to assess involvement of the corticospinal tract in capsular/pericapsular
stroke: Feasibility and clinical implications. J Magn Reson Imaging,
26(6):1399–1404, Dec 2007.
[KJS05] S. Kim, J.-W. Jeong, and M. Singh. Estimation of multiple fiber orienta-
tions from diffusion tensor MRI using independent component analysis.
Nuclear Science, IEEE Transactions on, 52(1):266–273, Feb 2005.
[KKP
+
08] D.-J. Kim, J.-J. Kim, J.Y . Park, S.Y . Lee, J. Kim, I.Y . Kim, S.I. Kim, and
H.-J. Park. Quantification of thalamocortical tracts in schizophrenia on
probabilistic maps. Neuroreport, 19(4):399–403, Mar 2008.
[KNZ
+
08] V . Kavcic, H. Ni, T. Zhu, J. Zhong, and C.J. Duffy. White matter integrity
linked to functional impairments in aging and early Alzheimer’s disease.
Alzheimers Dement, 4(6):381–389, Nov 2008.
[KSC
+
07] M.F. Kraus, T. Susmaras, B.P. Caughlin, C.J. Walker, J.A. Sweeney, and
D.M. Little. White matter integrity and cognition in chronic traumatic
136
brain injury: a diffusion tensor imaging study. Brain, 130(Pt 10):2508–
2519, Oct 2007.
[KZS
+
88] A.B. Kelly, R.D. Zimmerman, R.B. Snow, S.E. Gandy, L.A. Heier, and
M.D. Deck. Head trauma: comparison of MR and CT–experience in 100
patients. AJNR Am J Neuroradiol, 9(4):699–708, 1988.
[LBBL
+
86] D. Le Bihan, E. Breton, D. Lallemand, P. Grenier, E. Cabanis, and
M. Laval-Jeantet. MR imaging of intravoxel incoherent motions: appli-
cation to diffusion and perfusion in neurologic disorders. Radiology,
161(2):401–407, Nov 1986.
[LNC
+
92] A.F. Leuchter, T.F. Newton, I.A. Cook, D.O. Walter, S. Rosenberg-
Thompson, and P.A. Lachenbruch. Changes in brain functional connec-
tivity in Alzheimer-type and multi-infarct dementia. Brain, 115(5):1543–
1561, 1992.
[LNN12] J. Lamirande-Nadeau and R. Nazrati. Screenshot of Fiber Naviga-
tor. https://imn697-fibernavigator.googlecode.com/
files/Fiber%20Navigator%28IMN638%29.pdf, Accessed on
Aug 18, 2012.
[MA12] D. Mount and S. Arya. ANN: A library for approximate nearest neighbor
searching. http://www.cs.umd.edu/
˜
mount/ANN/, Accessed
on Aug 18, 2012.
[MBA07] S.M. Maniega, M.E. Bastin, and P.A. Armitage. A quantitative compar-
ison of two methods to correct eddy current-induced distortions in DT-
MRI. Magn Reson Imaging, 25(3):341–349, Apr 2007.
[MCCVZ99] S. Mori, B.J. Crain, V .P. Chacko, and P.C.M. Van Zijl. Three-dimensional
tracking of axonal projections in the brain by magnetic resonance imag-
ing. Annals of Neurology, 45(2):265–269, 1999.
[MCK
+
90] M.E. Moseley, Y . Cohen, J. Kucharczyk, J. Mintorovitch, H.S. Asgari,
M.F. Wendland, J. Tsuruda, and D. Norman. Diffusion-weighted MR
imaging of anisotropic water diffusion in cat central nervous system.
Radiology, 176(2):439–445, Aug 1990.
[MIZ
+
01] S. Mori, R. Itoh, J. Zhang, W.E. Kaufmann, P.C. van Zijl, M. Solaiyappan,
and P. Yarowsky. Diffusion tensor imaging of the developing mouse brain.
Magn Reson Med, 46(1):18–23, Jul 2001.
[MKD
+
02] S. Mori, W.E. Kaufmann, C. Davatzikos, B. Stieltjes, L. Amodei, K. Fred-
ericksen, G.D. Pearlson, E.R. Melhem, M. Solaiyappan, G.V . Raymond,
137
H.W. Moser, and P.C.M. van Zijl. Imaging cortical association tracts in the
human brain using diffusion-tensor-based axonal tracking. Magn Reson
Med, 47(2):215–223, Feb 2002.
[MKP
+
00] S. Mori, W.E. Kaufmann, G.D. Pearlson, B.J. Crain, B. Stieltjes,
M. Solaiyappan, and P.C. van Zijl. In vivo visualization of human neural
pathways by magnetic resonance imaging. Ann Neurol, 47(3):412–414,
Mar 2000.
[Mor07] S. Mori. Introduction to Diffusion Tensor Imaging. Elsevier, 2007.
[MTG
+
08] G.K. Malik, R. Trivedi, A. Gupta, R. Singh, K.N. Prasad, and R.K. Gupta.
Quantitative DTI assessment of periventricular white matter changes in
neonatal meningitis. Brain Dev, 30(5):334–341, May 2008.
[MULK07] H.-P. M
¨
(u)ller, A. Unrath, A.C. Ludolph, and J. Kassubek. Preservation of
diffusion tensor properties during spatial normalization by use of tensor
imaging and fibre tracking on a normal brain database. Phys Med Biol,
52(6):N99–109, Mar 2007.
[MvZ02] S. Mori and P.C.M. van Zijl. Fiber tracking: principles and strategies - a
technical review. NMR Biomed, 15(7–8):468–480, 2002.
[NH02] Thomas E. Nichols and Andrew P. Holmes. Nonparametric permutation
tests for functional neuroimaging: a primer with examples. Hum Brain
Mapp, 15(1):1–25, Jan 2002.
[NMMH02] J. Neil, J. Miller, P. Mukherjee, and P. S. Hppi. Diffusion tensor imaging
of normal and injured developing human brain - a technical review. NMR
Biomed, 15(7-8):543–552, 2002.
[NTL
+
05] Govind Nair, Yusuke Tanahashi, Hoi Pang Low, Susan Billings-Gagliardi,
William J. Schwartz, and Timothy Q. Duong. Myelination and long dif-
fusion times alter diffusion-tensor-imaging contrast in myelin-deficient
shiverer mice. Neuroimage, 28(1):165–174, Oct 2005.
[Ope12] OpenWetWare. Screenshot of DTI-Query. http://openwetware.
org/images/a/a2/TA314_N27AF.jpg, Accessed on Aug 18,
2012.
[PB96] C. Pierpaoli and P.J. Basser. Toward a quantitative assessment of diffusion
anisotropy. Magn Reson Med, 36(6):893–906, Dec 1996.
138
[PCF
+
00] C. Poupon, C.A. Clark, V . Frouin, J. Rgis, I. Bloch, D. Le Bihan, and
J. Mangin. Regularization of diffusion-based direction maps for the track-
ing of brain white matter fascicles. Neuroimage, 12(2):184–195, Aug
2000.
[PHK04] S. Pieper, M. Halle, and R. Kikinis. 3D Slicer. In Biomedical Imag-
ing: Nano to Macro, 2004. IEEE International Symposium on, volume 1,
pages 632–635, Apr 2004.
[PHWK03] G.J.M. Parker, H.A. Haroon, and C.A.M. Wheeler-Kingshott. A frame-
work for a streamline-based probabilistic index of connectivity (PICo)
using a structural interpretation of MRI diffusion measurements. Journal
of Magnetic Resonance Imaging, 18(2):242–254, 2003.
[PMB
+
05] S.C. Partridge, P. Mukherjee, J.I. Berman, R.G. Henry, S.P. Miller,
Y . Lu, O.A. Glenn, D.M. Ferriero, A.J. Barkovich, and D.B. Vigneron.
Tractography-based quantitation of diffusion tensor imaging parameters
in white matter tracts of preterm newborns. J Magn Reson Imaging,
22(4):467–474, Oct 2005.
[PMP
+
00] N.G. Papadakis, K.M. Martin, J.D. Pickard, L.D. Hall, T.A. Carpenter,
and C.L. Huang. Gradient preemphasis calibration in diffusion-weighted
echo-planar imaging. Magn Reson Med, 44(4):616–624, Oct 2000.
[PSBM05] N.G. Papadakis, T. Smponias, J. Berwick, and J.E.W. Mayhew. k-space
correction of eddy-current-induced distortions in diffusion-weighted
echo-planar imaging. Magn Reson Med, 53(5):1103–1111, May 2005.
[PSR
+
03] T. Ptak, R.L. Sheridan, J.T. Rhea, A.A. Gervasini, J.H. Yun, M.A. Cur-
ran, P. Borszuk, L. Petrovick, and R.A. Novelline. Cerebral fractional
anisotropy score in trauma patients: a new indicator of white matter injury
after trauma. AJR Am J Roentgenol, 181(5):1401–1407, Nov 2003.
[RGSB
+
01] F.J. Rugg-Gunn, M.R. Symms, G.J. Barker, R. Greenwood, and J.S. Dun-
can. Diffusion imaging shows abnormalities after blunt head trauma when
conventional magnetic resonance imaging is normal. J Neurol Neurosurg
Psychiatry, 70(4):530–533, Apr 2001.
[RHWW03] T.G. Reese, O. Heid, R.M. Weisskoff, and V .J. Wedeen. Reduction of
eddy-current-induced distortion in diffusion MRI using a twice-refocused
spin echo. Magn Reson Med, 49(1):177–182, Jan 2003.
[RSD12] A. Riggall, M. Steven, and K. Doron. Screenshot of DTIS-
tudio. http://dbic.dartmouth.edu/wiki/images/9/91/
150_6.5.jpg, Accessed on Aug 18, 2012.
139
[RTC
+
08] D.R. Rutgers, F. Toulgoat, J. Cazejust, P. Fillard, P. Lasjaunias, and
D. Ducreux. White matter abnormalities in mild traumatic brain injury: a
diffusion tensor imaging study. AJNR Am J Neuroradiol, 29(3):514–519,
Mar 2008.
[SAM
+
05] A. Sherbondy, D. Akers, R. Mackenzie, R. Dougherty, and B. Wandell.
Exploring connectivity of the brain’s white matter with dynamic queries.
IEEE Trans Vis Comput Graph, 11(4):419–430, 2005.
[SES
+
08] A. Sidaros, A.W. Engberg, K. Sidaros, M.G. Liptrot, M. Herning,
P. Petersen, O.B. Paulson, T.L. Jernigan, and E. Rostrup. Diffusion tensor
imaging during recovery from severe traumatic brain injury and relation
to clinical outcome: a longitudinal study. Brain, 131(Pt 2):559–572, Feb
2008.
[SH05] M. Singh and D.H. Hwang. Comparison of tract-length and FA his-
tograms to evaluate global tractography. In Proceedings 13th Scien-
tific Meeting, International Society for Magnetic Resonance in Medicine,
Miami, page 2731, 2005.
[SHSV05] M. Singh, D.H. Hwang, W. Sungkarat, and K. Veera. Evaluation of
MRI DTI-tractography by tract-length histogram. In Amir A. Amini and
Armando Manduca, editors, Progress in Biomedical Optics and Imaging:
Physiology, Function and Structure from Medical Images, volume 5746,
pages 138–147. SPIE, 2005.
[SJ06] M. Singh and J. Jeong. Use of independent component analysis to esti-
mate multiple fibers within a voxel. In Proceedings 14th Scientific Meet-
ing, International Society for Magnetic Resonance in Medicine, Seattle,
page 3168, 2006.
[SJ08] M. Singh and J. Jeong. DTI-tractography to detect and quantify brain
pathways affected in traumatic brain injury. In Proceedings 16th Scien-
tific Meeting, International Society for Magnetic Resonance in Medicine,
Toronto, page 2269, 2008.
[SJ09] M. Singh and J. Jeong. Localization and quantification of injured regions
and affected pathways in the 3D head-space of individual TBI subjects
using DTI tractography with automatically generated ROIs. In Proceed-
ings 17th Scientific Meeting, International Society for Magnetic Reso-
nance in Medicine, Honolulu, page 3410, 2009.
[SJH
+
10] M. Singh, J.-W. Jeong, D.H. Hwang, W. Sungkarat, and P. Gruen. Novel
diffusion tensor imaging methodology to detect and quantify injured
140
regions and affected brain pathways in traumatic brain injury. Magn
Reson Imaging, 28(1):22–40, Jan 2010.
[SKWC04] M. Singh, S. Kim, M. Weiner, and H. Chui. Quantification of fronto-
occipital and thalamo-frontal connectivity in Alzheimer’s disease by DTI-
tractography. In Proceedings 12th Scientific Meeting, International Soci-
ety for Magnetic Resonance in Medicine, Kyoto, page 2731, May 2004.
[SMC
+
06] C.H. Salmond, D.K. Menon, D.A. Chatfield, G.B. Williams, A. Pena, B.J.
Sahakian, and J.D. Pickard. Diffusion tensor imaging in chronic head
injury survivors: correlations with learning and memory indices. Neu-
roimage, 29(1):117–124, Jan 2006.
[SP02] J. Sahuquillo and M.A. Poca. Diffuse axonal injury after head trauma. a
review. Adv Tech Stand Neurosurg, 27:23–86, 2002.
[SRP08] E.V . Sullivan, T. Rohlfing, and A. Pfefferbaum. Quantitative fiber track-
ing of lateral and interhemispheric white matter systems in normal aging:
Relations to timed performance. Neurobiol Aging, 31(3):464–481, May
2008.
[SSP01] J.R. Stone, R.H. Singleton, and J.T. Povlishock. Intra-axonal neurofila-
ment compaction does not evoke local axonal swelling in all traumatically
injured axons. Experimental Neurology, 172(2):320–331, 2001.
[SSR
+
02] S.-K. Song, S.-W. Sun, M.J. Ramsbottom, C. Chang, J. Russell, and A.H.
Cross. Dysmyelination revealed through MRI as increased radial (but
unchanged axial) diffusion of water. Neuroimage, 17(3):1429–1436, Nov
2002.
[SSS
+
07] M. Singh, W. Sungkarat, A. Shetty, A. Rajagopalan, D.H. Hwang,
C. Wong, H. Chui, N. Schuff, and M. Weiner. Toward quantita-
tion of whole-brain tractography group comparisons with application to
Alzheimer disease. In Proceedings 15th Scientific Meeting, International
Society for Magnetic Resonance in Medicine, Berlin, page 343, 2007.
[Ste65] E.O. Stejskal. Use of spin echoes in a pulsed magnetic-field gradient to
study anisotropic, restricted diffusion and flow. The Journal of Chemical
Physics, 43(10):3597–3603, 1965.
[Str61] S.J. Strich. Shearing of nerve fibres as a cause of brain damage due
to head injury: A pathological study of twenty cases. The Lancet,
278(7200):443–448, 1961. ¡ce:title¿Originally published as V olume 2,
Issue 7200¡/ce:title¿.
141
[STvdK
+
10] D.H. Salat, D.S. Tuch, A.J.W. van der Kouwe, D.N. Greve, V . Pappu, S.Y .
Lee, N.D. Hevelone, A.K. Zaleta, J.H. Growdon, S. Corkin, B. Fischl, and
H.D. Rosas. White matter pathology isolates the hippocampal formation
in Alzheimer’s disease. Neurobiology of Aging, 31:244–256, 2010.
[SW06] M. Singh and C.-W. Wong. Recovery of multiple fibers per voxel by ICA
in DTI tractography. Conf Proc IEEE Eng Med Biol Soc, 1:735–738,
2006.
[SW07] M. Singh and C.-W. Wong. Whole-brain tractography incorporating ICA
based crossing-fiber orientations. In Proceedings 15th Scientific Meeting,
International Society for Magnetic Resonance in Medicine, Berlin, page
899, 2007.
[SW09] M. Singh and C.-W. Wong. ICA based multi-fiber tractography. In Pro-
ceedings 17th Scientific Meeting, International Society for Magnetic Res-
onance in Medicine, Honolulu, page 853, 2009.
[SW10] M. Singh and C.-W. Wong. Independent component analysis-based mul-
tifiber streamline tractography of the human brain. Magn Reson Med,
64(6):1676–1684, Dec 2010.
[SWB
+
09] N. Schuff, N. Woerner, L. Boreta, T. Kornfield, L.M. Shaw, J.Q. Tro-
janowski, P.M. Thompson, C.R. Jack, M.W. Weiner, and the Alzheimer’s;
Disease Neuroimaging Initiative. MRI of hippocampal volume loss in
early Alzheimer’s disease in relation to ApoE genotype and biomarkers.
Brain, 132(4):1067–1077, 2009.
[SYZ
+
07a] H. Sun, P.A. Yushkevich, H. Zhang, P.A. Cook, J.T. Duda, T.J. Simon,
and J.C. Gee. Shape-based normalization of the corpus callosum for DTI
connectivity analysis. IEEE Trans Med Imaging, 26(9):1166–1178, Sep
2007.
[SYZ
+
07b] Hui Sun, Paul A Yushkevich, Hui Zhang, Philip A Cook, Jeffrey T Duda,
Tony J Simon, and James C Gee. Evaluation of shape-based normalization
in the corpus callosum for white matter connectivity analysis. Med Image
Comput Comput Assist Interv Int Conf Med Image Comput Comput Assist
Interv, 10(Pt 2):777–784, 2007.
[SZE98] J.G. Sled, A.P. Zijdenbos, and A.C. Evans. A nonparametric method for
automatic correction of intensity nonuniformity in MRI data. IEEE Trans
Med Imaging, 17:87–97, 1998.
142
[Tan78] J.E. Tanner. Transient diffusion in a system partitioned by permeable
barriers. application to NMR measurements with a pulsed field gradient.
J. Chem. Phys., 69(4):1748–1754, 1978.
[TCS08] T.-K. Truong, B. Chen, and A.W. Song. Integrated SENSE DTI with cor-
rection of susceptibility- and eddy current-induced geometric distortions.
Neuroimage, 40(1):53–58, Mar 2008.
[THS10] S. Tsao, D.H. Hwang, and M. Singh. CSF contamination correction
in DTI tractography of the fornix in elderly subjects. In Proceedings
18th Scientific Meeting, International Society for Magnetic Resonance in
Medicine, Stockholm, page 1647, 2010.
[TRB
+
06] J.M. Tyszka, C. Readhead, E.L. Bearer, R.G. Pautler, and R.E. Jacobs.
Statistical diffusion tensor histology reveals regional dysmyelination
effects in the shiverer mouse mutant. Neuroimage, 29(4):1058–1065, Feb
2006.
[TRW
+
02] D.S. Tuch, T.G. Reese, M.R. Wiegell, N. Makris, J.W. Belliveau, and V .J.
Wedeen. High angular resolution diffusion imaging reveals intravoxel
white matter fiber heterogeneity. Magn Reson Med, 48(4):577–582, Oct
2002.
[VDV
+
08] N. Villain, B. Desgranges, F. Viader, V . de la Sayette, F. Mezenge,
B. Landeau, J.-C. Baron, F. Eustache, and G. Chetelat. Relationships
between hippocampal atrophy, white matter disruption, and gray matter
hypometabolism in alzheimer’s disease. J. NeuroSci., 28(24):6174–6181,
June 2008.
[WBD
+
08] J.Y . Wang, K. Bakhadirov, M.D. Devous, Sr, H. Abdi, R. McColl,
C. Moore, C.D. Marquez de la Plata, K. Ding, A. Whittemore, E. Bab-
cock, T. Rickbeil, J. Dobervich, D. Kroll, B. Dao, N. Mohindra, C.J.
Madden, and R.amon Diaz-Arrastia. Diffusion tensor tractography of
traumatic diffuse axonal injury. Arch Neurol, 65(5):619–626, May 2008.
[WBSW07] R. Wang, T. Benner, A.G. Sorensen, and V .J. Wedeen. Diffusion Toolkit:
a software package for diffusion imaging data processing and tractogra-
phy. In Proceedings 15th Scientific Meeting, International Society for
Magnetic Resonance in Medicine, Berlin, page 3720, 2007.
[WCP
+
07] S. Wakana, A. Caprihan, M.M. Panzenboeck, J.H. Fallon, M. Perry, R.L.
Gollub, K. Hua, J. Zhang, H. Jiang, P. Dubey, A. Blitz, P. van Zijl, and
S. Mori. Reproducibility of quantitative tractography methods applied to
cerebral white matter. Neuroimage, 36(3):630–644, Jul 2007.
143
[WJNP
+
04] S. Wakana, H. Jiang, L.M. Nagae-Poetscher, P.C.M. van Zijl, and S. Mori.
Fiber tract-based atlas of human white matter anatomy. Radiology,
230(1):77–87, Jan 2004.
[WKHP
+
02] C.A.M. Wheeler-Kingshott, S.J. Hickman, G.J.M. Parker, O. Ciccarelli,
M.R. Symms, D.H. Miller, and G.J. Barker. Investigating cervical
spinal cord structure using axial diffusion tensor imaging. Neuroimage,
16(1):93–102, May 2002.
[WMH
+
08] E.A. Wilde, S.R. McCauley, J.V . Hunter, E.D. Bigler, Z. Chu, Z.J. Wang,
G.R. Hanten, M. Troyanskaya, R. Yallampalli, X. Li, J. Chia, and H.S.
Levin. Diffusion tensor imaging of acute mild traumatic brain injury in
adolescents. Neurology, 70(12):948–955, Mar 2008.
[WS07] C. Wong and M. Singh. Estimating number of fiber directions per voxel
for multiple fiber DTI tractography. In Proceedings 15th Scientific Meet-
ing, International Society for Magnetic Resonance in Medicine, Berlin,
page 1478, 2007.
[WS10] B. Wilkins and M. Singh. Diffusion histogram as a marker of fiber
crossing within a voxel. In Proceedings 18th Scientific Meeting, Inter-
national Society for Magnetic Resonance in Medicine, Stockholm, page
1699, 2010.
[WW12] R. Wang and V .J. Wedeen. Screenshot of Track Vis. http:
//www.trackvis.org/screenshots/?display=images/
screenshot_xp.png, Accessed on Aug 18, 2012.
[XHB
+
08] D. Xu, X. Hao, R. Bansal, K.J. Plessen, and B.S. Peterson. Seamless
warping of diffusion tensor fields. IEEE Trans Med Imaging, 27(3):285–
299, Mar 2008.
[XMS
+
03] D. Xu, S. Mori, D. Shen, P.C.M. van Zijl, and C. Davatzikos. Spatial
normalization of diffusion tensor fields. Magn Reson Med, 50(1):175–
182, Jul 2003.
[YSD
+
08] B.T.T. Yeo, M.R. Sabuncu, R. Desikan, B. Fischl, and P. Golland. Effects
of registration regularization and atlas sharpness on segmentation accu-
racy. Med Image Anal, 12(5):603–615, Oct 2008.
[ZDL03] S. Zhang, C. Demiralp, and D.H. Laidlaw. Visualizing diffusion ten-
sor MR images using streamtubes and streamsurfaces. Visualization
and Computer Graphics, IEEE Transactions on, 9(4):454–462, Oct–Dec
2003.
144
[ZYRG09] H. Zhang, P. Yushkevich, D. Rueckert, and J. Gee. A computational DTI
template for aging studies. In Proceedings 17th Scientific Meeting, Inter-
national Society for Magnetic Resonance in Medicine, Honolulu, page
3230, 2009.
[ZZL
+
08] W. Zhan, Y . Zhang, P. Lorenzen, S.G. Mueller, N. Schuff, and M.W.
Weiner. Correlations between DTI and FLAIR images reveal the rela-
tionships of microscopic and macroscopic white matter degeneration in
elderly subjects. In Xiaoping P. Hu and Anne V . Clough, editors, Pro-
ceedings of SPIE Medical Imaging: Physiology, Function, and Structure
from Medical Images, volume 6916, page 691609, San Diego, CA, USA,
2008. SPIE.
[ZZZ
+
03] R.L. Zhang, L. Zhang, Z.G. Zhang, D. Morris, Q. Jiang, L. Wang, L.J.
Zhang, and M. Chopp. Migration and differentiation of adult rat sub-
ventricular zone progenitor cells transplanted into the adult rat striatum.
Neuroscience, 116(2):373–382, 2003.
145

Abstract (if available)

Abstract With the advent of diffusion tensor imaging (DTI) came insight into the organization of the most complex organic computer in existence—the human brain. Diffusion Tensor Tractography (DTT) introduced the ability to visualize, in vivo, axonal fiber bundles, the brain’s internal wiring structures. ❧ Rendering tractography in three dimensions aids in the understanding of how the axonal connections of the brain are organized, and is an important tool in illustrating the complex geometry of fiber bundles. To better facilitate the use of 3D visualization for tractography, we wrote flexible custom software targeted at researchers. ❧ The use of tractography need not be limited to visualization

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Diffusion MRI white matter tractography: estimation of multiple fibers per voxel using independent component analysis

PDF

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

PDF

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

PDF

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

PDF

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

PDF

The effects of the vitreous state on the ocular pharmacokinetics following intravitreal drug delivery: an evaluation with quantitative magnetic resonance imaging

PDF

Novel computational techniques for connectome analysis based on diffusion MRI

PDF

Methods for improving reliability and consistency in diffusion MRI analysis

PDF

Engineering scalable two- and three-dimensional striated muscle microtissues for human disease modeling

PDF

Pediatric magnetic resonance image processing: applications to posterior fossa cancer and normal development

PDF

Representation problems in brain imaging

PDF

Dynamic functional magnetic resonance imaging to localize activated neurons

PDF

A wireless implantable MEMS micropump system for site-specific anti-cancer drug delivery

PDF

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

PDF

Modeling anti-tumoral effects of drug-induced activation of the cell-extrinsic apoptotic pathway

PDF

Novel beamforming techniques for robust contrast enhancement in ultrasound imaging

PDF

Characterization of visual cortex function in late-blind individuals with retinitis pigmentosa and Argus II patients

PDF

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

PDF

Applications of graph theory to brain connectivity analysis

PDF

High-frequency ultrasonic transducers for photoacoustic applications

Asset Metadata

Creator Hwang, Darryl Hwa (author)

Core Title Diffusion tensor tractography: visualization and quantitation with applications to Alzheimer disease and traumatic brain injury

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 11/16/2012

Defense Date 10/19/2012

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

Tag diffusion tensor imaging,diffusion tensor tractography,DTI,DTT,MRI,OAI-PMH Harvest,tractography

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor D'Argenio, David Z. (committee chair), Leporé, Natasha (committee member), Wolf, Walter (committee member), Yen, Jesse T. (committee member)

Creator Email darrylhw@gmail.com,darrylhw@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-111591

Unique identifier UC11289535

Identifier usctheses-c3-111591 (legacy record id)

Legacy Identifier etd-HwangDarry-1294.pdf

Dmrecord 111591

Document Type Dissertation

Rights Hwang, Darryl Hwa

Type texts

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

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

Repository Name University of Southern California Digital Library

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