a novel diffusion tensor imaging-based computer...
TRANSCRIPT
1
A NOVEL DIFFUSION TENSOR
IMAGING-BASED COMPUTER-AIDED
DIAGNOSTIC SYSTEM FOR
EARLY DIAGNOSIS OF AUTISM
By
Mahmoud Mostapha
Thesis Defense, Master of Science,
Electrical and Computer Engineering Department,
University of Louisville
Email: [email protected]
University of Louisville
Research Motivation
• Autism spectrum disorders (ASDs) is a group
of lifetime developmental disabilities that are
defined by significant social, communication and
behavioral challenges.
• The Centers for Disease Control and Prevention
(CDC)[1] estimates that one in 68 children has
been diagnosed with ASDs in the United States,
which is approximately 30% greater than previous
estimates reported in 2012 of one in 88 children
• Most children with ASDs are currently diagnosed
after the age of four, despite the fact that ASDs
can be identified as early as age two
2[1] Centers for Disease Control and Prevention (CDC), http://www.cdc.gov/media/releases/2014/p0327-autism-spectrum-disorder.html
Problem and Critical Unmet Need
3
• The American Academy of Pediatrics
(AAP)[1] recommends screening for
developmental disabilities at 9,18 and either
24 or 30 months
• This diagnostic approach is not applicable at
a very early age because it relies on the
observation and assessment of age-
dependent emergent skills in the domains of
communication, social interaction, and play
• There is an urgent need for a non-invasive technology with the capability of
providing new laboratory-based measures that confer an accurate and early
diagnosis of autism
[1] AAP: “The American Academy of Pediatrics . ” http:// http://www.aap.org/
Research Objective
• The ultimate goal of this thesis is to develop a
computer-aided diagnosis (CAD) system for the
accurate and early diagnosis of ASDs using diffusion
tensor imaging (DTI)
4
The Proposed CAD System Framework
5
4D DWI Brain Data
INPUT
Final Assessment
Diagnosis
Control
Autistic
Segmentation
STEP 1
Brain Tissues
Segmentation
Brain Extraction
Classification
STEP 3
CSF
GM
WM
Feature Extraction
STEP 2
Brain cortex
Reconstruction
WM Tracts
Reconstruction
Connectivity
Analysis
Shape
Analysis
Diagnosis and
Brain Mapping
Mismatch
Data Acquisition Challenges
• Challenges of scanning infants:
1. Safety challenges
• RF heating effects
(no sedation and anesthesia)
2. Anatomy challenges
• Smaller structures
(low scan resolution)
3. Behavioral challenges
• Subject cooperation difficult
(short scan time)
• Motion problems
(Bulk motion distortions)
6
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (1)
7
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (1)
8
b0
b1
b2
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (1)
9
Segmentation Challenges (1)
10
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (1)
11
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (1)
12
Diffusion images are sensitive to a number
of artifacts, which includes:
• Noise/SNR issues
• Eddy-current distortions
• Electromagnetic interference
• Venetian blind artifact
• Slice-Wise Inconsistencies
• partial volume effect
Segmentation Challenges (6)
13
Accurate Segmentation of the infant brain tissue is essential for the
early autism diagnosis CAD system. However, the majority of the
existing techniques are developed to work for adult MR brain images
and fail to accurately segment brain tissues from MR infant images
due to the reduced contrast and higher noise
Adult DWI - Axial Adult DTI- Axial Infant DWI - Axial Infant DTI- Axial
Segmentation Challenges (8)
14
Infant brain segmentation in the isointense stage (6-12 months) is
more challenging, in which the signal intensity of the white matter is
increasing during the development due to the myelination and
maturation processes; in this stage, the gray matter has the lowest
signal differentiation with the white matter
Histogram Calculation
Non-Diffusion b0 (T2-weighted) Scan Intensity Histogram of
Different Brain Tissues
DWI Brain Data
15
4D DWI Brain Data
INPUT
Final Assessment
Diagnosis
Control
Autistic
Segmentation
STEP 1
Brain Tissues
Segmentation
Brain Extraction
Classification
STEP 3
CSF
GM
WM
Feature Extraction
STEP 2
Brain cortex
Reconstruction
WM Tracts
Reconstruction
Connectivity
Analysis
Shape
Analysis
Diagnosis and
Brain Mapping
Mismatch
4D DWI Brain Data
b0
b1
b13
b25
Axial CoronalSagittal 3D Volume
Infant Brain Segmentation
17
4D DWI Brain Data
INPUT
Final Assessment
Diagnosis
Control
Autistic
Segmentation
STEP 1
Brain Tissues
Segmentation
Brain Extraction
Classification
STEP 3
CSF
GM
WM
Feature Extraction
STEP 2
Brain cortex
Reconstruction
WM Tracts
Reconstruction
Connectivity
Analysis
Shape
Analysis
Diagnosis and
Brain Mapping
Mismatch
BrainExtraction
1. S. M. Smith, “Fast robust automated brain extraction,” Human Brain Mapping,2002
2. M. Jenkinson et al., “BET2: MR-based estimation of brain, skull and scalpsurfaces,” Human Brain Mapping, 2005
3. D. W. Shattuck and R. M. Leahy, “Brainsuite: an automated cortical surfaceidentification tool,” Medical Image Analysis, 2002
4. F. Shi, et al., “LABEL: Pediatric brain extraction using learning-based meta-algorithm,” NeuroImage, 2012
Brain Segmentation
1. L. Zhang et al., “Spatial–temporal constraint for segmentation of serial infant brain MR images,” MIAR, 2010
2. L. Wang et al., “4Dmulti-modality tissue segmentation of serial infant images,”PLoS One, 2012
3. L. J. Wolff et al., “Differences in White Matter Fiber Tract Development Present From 6 to 24 Months in Infants With Autism,” Psychiatry, 2012
4. L. Wang et al., “Integration of Sparse Multi-modality Representation and Geometrical Constraint for Isointense Infant Brain Segmentation,” MICCAI, 2013
Related Work
18
The Limitations of the Existing Techniques
19
Statistical-based
Fixed Atlas-based
Deformable model-based
•Heavily depend on the atlas selection•And registration accuracy •Time consuming
•Very sensitive to the initialization model •and the guiding forces
•Depend only on pre-defined probability models• Intensity sensitive
• Most of the previously described MR infant segmentation techniques fail in
the case of infants in the isointense stage as they depend on T1 or T2 scans
• Curent DTI-based infant brain segmentation techniques suffer from the
following limitations: (i) atlases constructed from multiple modalities,
(ii) using nonlinear registration, and (iii) rely on longitudinal information
The Proposed Segmentation Framework
20
The Proposed Segmentation Framework
21
DWI Image Acquisition
22[1] IBIS: “Infant brain imaging study. ” http://www.ibisnetwork.org/
b0
b25
• More than 300 Diffusion weighted infant
brain data sets were obtained from the
Infant Brain Imaging Study (IBIS) with 10
data sets manually segmented by an MR
expert
• Diffusion weighted MRI brain scans were
acquired with the following parameters:
• Voxel size: 2mm x 2mm x 2mm
• b values between 0 and 1,000 s/mm2
• Number of gradient directions: 25
• Number of slices: 75–81
• Scan time of 5-6 minutes
The Proposed Segmentation Framework
23
• In order to perform the required DWI quality control (QC),
DTIprep software[1] was used to automatically detects and removes scan
artifacts, and correct motion and eddy current distortions
• Scans with strong remaining artifacts were eliminated, and resultant data
sets with low number of gradient directions, which will produce DTI
estimates with low SNR, were excluded from any further processing
24
Step 1a – DTIprep Preprocessing
[1] Z. Liu, Y. Wang, G. Gerig, S. Gouttard, R. Tao, T. Fletcher, and M. Styner, “Quality control of diffusion weighted images,” in Proceedings of SPIE Medical Imaging 2000: Image Processing (SPIE’10), 2010, pp. 76280J–76280J.
25
Step 1b – DTI Estimation (1)
S0 S1 S3 Sn-1 Sn
zzzyzx
yzyyyx
xzxyxx
DDD
DDD
DDD
D
DWI Data
Diffusion Tensor Ellipsoid Model
• The weighted linear least square (WLLS) method was used to perform
tensor model estimation from the preprocessed DWI data sets, WLLS
method was preferred because of its ability to provide accurate estimates
with low processing times
• DTI to DWI estimation using WLLS method was carried out using
3D Slicer software[1], and to account for bad tensors, related to any
remaining noise or acquisition artifacts, negative eigenvalues (which are
physically meaningless) were shifted
26
Step 1b – DTI Estimation (2)
[1] A. Fedorov, R. Beichel, J. Kalpathy-Cramer, J. Finet, J.-C. Fillion-Robin, S. Pujol, C. Bauer, D. Jennings, F. Fennessy, M. Sonka, etal.,“3D slicer as an image computing platform for the quantitative imaging network,” Multidisciplinary Respiratory Medicine, vol.30, no. 9, pp. 1323–1341, 2012
DWI
DTI
WLLS Estimation
The Proposed Segmentation Framework
27
Step 2a – LCDG Model (1)
• We used the Linear Combination of Discrete Gaussians (LCDG) to
accurately estimate the marginal density of intensity distribution for the
brain and non-brain tissues
);( and 0, )( ,1
)|()|()(
,
11
,
1
,,
1
,
qPcc
qwqcqP
np
np
C
k
n,k
K
k
kp
K
k
knkn
K
k
kpp,k,
c
c Θ
Positive mixture Negative mixture
q Intensity (gray) levelKp Numbers of the positive Gaussian componentsKn Numbers of the negative Gaussian componentsC Weights coefficientsΨ Discrete Gaussian term with mean μ and variance σ2
[1] A. El-Baz, A. Elnakib, F. Khalifa, M. A. El-Ghar, P. McClure, A. Soliman, and G. Gimel’farb, “Precise Segmentation of 3D Magnetic Resonance
Angiography,” IEEE Transaction on Biomedical Engineering, vol. 59, no. 7, pp. 2019–2029, 2012.28
Step 2a – LCDG Model (2)
• At the end of this step, a discriminant threshold τ is calculated in a way
that ensures the best separation between the brain and the non-brain
voxel signals
• This threshold will be used in the next steps to enhance the process of
classifying image voxels into either brain or non-brain
29
• Enhance the spatial inhomegeinty by minimizing the distance between
each voxel and its 26-neighbors using 3D generalized Gauss-Markov
random field (GGMRF)[1] algorithm
• The continuity of 𝑞 values of each brain DWI scan is modified by applying
the gradient descent algorithm to search for the closest minimum of the
following equation:
30
Step 2b – GGMRF Model (1)
[1] C. Bouman and K. Sauer, “A generalized gaussian image model for edge preserving MAP estimation,” IEEE Transactions on
Image Processing, vol. 2, no. 3, pp. 296–310, 1993.
ෝ𝑞𝑠 = 𝑎𝑟𝑔 m𝑖𝑛𝑞𝑠
[ |𝑞𝑠 − 𝑞𝑠| + 𝜌𝛼𝜆𝛽
𝑟∊𝑣𝑠
𝜂𝑠, 𝑟 𝑞𝑠 − 𝑞𝑟𝛽 ]
𝑞𝑠 Original gray level 𝑞𝑠 Expected gray level estimate𝑣𝑠 The 26-neighborhood of 3D location 𝑠 = 𝑥, 𝑦, 𝑧𝜂𝑠, 𝑟 GGMRF potential𝜌, λ Scaling factors𝛼,𝛽 GGMRF controlling parameters
• The voxel signals are nudged additionally towards their most appropriate
grouping through incrementing or decrementing them by a bias value of ϵ.
The latter was chosen experimentally as 0.5% of the maximum gray
value, in accord with the discriminant threshold (τ) determined from the
LCDG model
31
Step 2b – GGMRF Model (2)
Intensity Normalization
3D Gibbs Smoothing
3D generalized Gauss-Markov random field
• After the final modified image is obtained, a 3D region growing is
applied, starting from a seed point at the center of the image volume,
followed by connected component analysis to calculate the final brain
mask, which is used to find the final extracted brain
32
Step 2c – Region Growing
Original Brain
Smoothed Brain
Brain mask
Extracted Brain
The Proposed Segmentation Framework
33
WM
GM
CSF
Others
Va
ria
bil
ity
0%
100%
Step 3 – Initial Segmentation
)(*)|(),( mPmgPmgP sp
NMF-Based Visual Appearance Model
(P(g|m))
Adaptive Shape Model
(Psp(m))
34
• In this thesis, five different anisotropy features were calculated using 3D
Slicer software[1], namely, mean diffusivity (MD), fractional
anisotropy (FA), relative anisotropy (RA), axial diffusivity (λ ), and
radial diffusivity (λ⊥)
35
Step 3a – NMF-Based Model (1)
[1] A. Fedorov, R. Beichel, J. Kalpathy-Cramer, J. Finet, J.-C. Fillion-Robin, S. Pujol, C. Bauer, D. Jennings, F. Fennessy, M. Sonka,et al.,“3D slicer as an image computing platform for the quantitative imaging network,” Multidisciplinary Respiratory Medicine,vol. 30, no. 9, pp. 1323–1341, 2012
=
MD λ λ⊥
FA RA
=
• Nonnegative Matrix Factorization (NMF) is a method for clustering data
by factorizing an input matrix A into a weight matrix W and an output
matrix H such that:
36
Step 3a – NMF-Based Model (2)
I
A
H
N
I J
N
W
J
I The dimensionality of the input data
N The number of input samples
J The dimensionality of the output space
• NMF feature fusion is applied to extract new meaningful features from the
large dimensional DTI feature space (A), which consists of one
appearance feature (b0) and five anisotropy features (FA, RA, MD, λ , λ⊥)
• In the training phase, the weight matrix W and an output matrix H can be
approximated by optimizing:
• The alternating least square (ALS) algorithm was used because of its
high speed and flexibility compared with other competing algorithms
37
Step 3a – NMF-Based Model (3)
=
2
,
2
1minimize WHA
HW
0 subject to W,H
[1] M. W. Berry, M. Browne, A. N. Langville, V. P. Pauca, and R. J. Plemmons, “Algorithms and applications for approximatenonnegative matrix factorization,” Computational Statistics & Data Analysis, 52(1):155–173, 2007.
• Given a weight matrix W learned using NMF, the feature vector of a new
voxel, B, is projected into H-space by using the psuedo-inverse of W
38
Step 3a – NMF-Based Model (4)
• To model the visual appearance of these features, a K-means classifier
was used with the classes J-dimensional centroids 𝐶𝑙; 𝑙 ∊ 𝐿 that were
calculated in the H-space during the training phase
• The NMF-based probabilities for brain label 𝑙 ∊ 𝐿, and voxel (𝑥, 𝑦, 𝑧) ∊ 𝑅 is
defined as:
39
Step 3a – NMF-Based Model (5)
Ll zyxBl
zyxBl
zyx
Hd
HdlmgP
)(
1
)(
1
)|(
,,:
,,:
,,
• Where 𝑑𝑖(𝐻𝐵: 𝑥, 𝑦, 𝑧) is the Euclidian
distance from the vector 𝐻𝐵: 𝑥, 𝑦, 𝑧 to the
centroid of class 𝑙 and 𝐿 is the set of
class labels
• Expected shapes of each brain label are constrained with a adaptive
probabilistic shape prior
• A training set of images, collected from different subjects (10 data sets)
with their new NMF fused features, are co-aligned by 3D affine
transformations with 12 degrees of freedom in a way that maximizes their
Mutual Information (MI)[1]
40
Rzyx
zyxzyxspsp mPmP),,(
,,,,: )()(
Statistical Prior Shapes
WM GMCSF Others
Vari
ab
ilit
y
0%
100%
[1] Viola, P., Wells III, W.M.: “Alignment by Maximization of Mutual Information.” International Journal of Computer Vision, 24(2), 137–154 (1997).
Step 3b – Adaptive Shape Model (1)
Step 3b – Adaptive Shape Model (2)
• Form an initial region map m using the marginal estimated density and
prior shape of each brain label
42
DWI brain images
CSF
GM
WM
)()|(),( mPmgPmgP sp
Step 3c – Initial MAP (1)
• Initial segmentation using the using the constructed adaptive shape
model and the NMF-based visual appearance model
43
Step 3c – Initial MAP (2)
DWI brain images
Initial segmentation
Extracted brain images
Axi
alSa
gitt
alC
oro
nal
The Proposed Segmentation Framework
44
45
3D Spatial Interaction MGRF Model
(Pv(m))
)(*)(*)|(),( mPmPmgPmgP vsp
NMF-Based Visual Appearance Model
(P(g|m))
Adaptive Shape Model
(Psp(m))
Step 4 – Final Segmentation
Improve the region map m using voxel-wise Bayes classifier
WM
GM
CSF
Others
Va
ria
bil
ity
0%
100%
• Markov Gibbs Random Field (MGRF)[1] with nearest 26-neighbors of the
voxels are used to calculate the 3D pair-wise interactions between region
labels and analytic bi-valued Gibbs potentials, that depend only on
whether the nearest pairs of labels are equal or not
Z Normalization factorNa Nearest 26-neighborhoodV Gibbs potential with an analytical estimate [1] :
Step 4 – Final Segmentation
3D Neighborhood system
[1] A. El-Baz, “Novel stochastic models for medical image analysis,” Ph.D. dissertation, University of Louisville,Louisville, KY, USA, 2006. 46
R vm ),,( ),,(,,,, ),( 1
)( zyx szyxzyx mmV
v eZ
P
2
12V eq mf
𝑓𝑒𝑞( 𝑚) The relative frequency of equal
labels in all the neighboring voxel pairs
• Construct the second-order Spatial Interaction (MGRF) Model (Pv(m))
Step 4 – Final Segmentation
47
3D Neighborhood System
GM
Step 4 – Final Segmentation
48
DWI brain images
Initial segmentation
Extracted brain images
Final segmentation
Axi
alSa
gitt
alC
oro
nal
Performance Evaluation Metrics (1)
The performance was evaluated using three metrics[1]:
a) Dice Similarity Coefficient (DSC)[2]
• The DSC characterizes the agreement between the segmented and
ground truth regions:
• TP True Positive
• TN True Negative
• FP False Positive
• FN False Negative
49
FNFPTP
TPDSC
2
2
[1] Babalola, Kolawole Oluwole, et al. "An evaluation of four automatic methods of segmenting the subcorticalstructures in the brain." Neuroimage 47.4 (2009): 1435-1447.
[2] D. Lee R, “Measures of the amount of ecologic association between species,” Ecology, vol. 26, pp. 297–302, 1945.
Performance Evaluation Metrics (2)
b) Modified Hausdorff Distance[1]
• Hausdorff distance (H) from a set S to a set G is defined as the
maximum distance of the set S to the nearest point in the set G:
s points of set S
g points of set G
d(s,g) the Euclidean distance between ‘s’ and ‘g’
• The bidirectional Hausdorff distance, between the segmented
region S and its ground truth G is defined as:
• In this work, the 95-percentile bidirectional Hausdorff distance as a
metric that measures the segmentation accuracy
50
)}},{{min{max),( gsdGSHGgSs
[1] G. Gerig, M. Jomier, and M. Chakos, “Valmet: A new validation tool for assessing and improving 3D object segmentation,” in Medical Image Computing and Computer Assisted Intervention, 2001, pp. 516–523.
S)}H(G,G),max{H(S, G)(S,HBi
Performance Evaluation Metrics (3)
c) Absolute Volume Difference (AVD)
• Defined as the ratio of the absolute difference between the ground
truth volume and the segmented volume, to the ground truth volume
51
AVD =|𝑉𝑠−𝑉𝑔|
𝑉𝑔Vs The segmented volumeVg The ground truth volume
MetricDifferent Brain Structures
CSF GM WM Others Brain
DSC (%) 87.96±3.31 89.92±2.86 95.23±1.18 92.81±5.26 96.64±1.15
H95(mm3) 2.42±0.56 1.98±1.07 1.98±0.01 6.60±2.75 7.17±3.96
AVD (%) 6.10±4.70 9.85±2.34 5.15±2.03 9.74±9.65 2.66±2.80
52
Segmentation Results
Infant Brain Segmentation
Infant Brain Extraction
• 3D visualization for our segmentation results:
CSF GM WM
Feature Extraction
53
4D DWI Brain Data
INPUT
Final Assessment
Diagnosis
Control
Autistic
Segmentation
STEP 1
Brain Tissues
Segmentation
Brain Extraction
Classification
STEP 3
CSF
GM
WM
Feature Extraction
STEP 2
Brain cortex
Reconstruction
WM Tracts
Reconstruction
Connectivity
Analysis
Shape
Analysis
Diagnosis and
Brain Mapping
Mismatch
Study Participants
• This study included data from the Infant Brain Imaging Study (IBIS) with
study participants are 6 moth old infants with high risk of developing
ASDs
• Final assessment was made at age 24 months, and based on an
ASDcutoff threshold, the high-risk infants were divided into two groups:
ASD negative (control) and ASD-positive (autistic)
• From 300 subjects provided, only 28 subjects with available final
diagnosis; however 3 subjects were excluded from further processing
owing to high scan artifacts and motion
• The final group of 25 subjects in this study included 6 infants, which met
the criteria for ASDs (4 males and 2 females), and 19 that did not
(11 males and 8 females)
54[1] IBIS: “Infant brain imaging study. ” http://www.ibisnetwork.org/
Feature Extraction Framework
55
Spherical Harmonics Analysis (1)
• The shape analysis was based on spherical harmonic reconstruction,
which considers 3D surface data (i.e., brain cortex) as a linear
combination of specific basis functions, namely spherical harmonics
(SHs)
• The spherical harmonics shape analysis is performed in five steps:
(i) mesh generation, (ii) mesh smoothing, (iii) unit sphere mapping, (iv)
spherical harmonics reconstruction, and (v) shape metrics calculation
56
Spherical Harmonics Analysis (2)
• To perform a quantitative analysis of the brain shape, two techniques for
measuring the complexity of the cerebral cortex are proposed:
1. SH reconstruction error
• The error between the original cortex mesh nodes and the SH approximated
cortex mesh nodes can be calculated in terms of Euclidean distance
57
Spherical Harmonics Analysis (3)
2. Surface complexity
• A new metric for examining the complexity of the brain using the SH
coefficients is also proposed:
58
𝑆 𝑓 =
𝑁=0
∞
𝑁 𝐵𝑁2
f Unit SphereN Number of HarmonicsB SH Coefficients
Shape Analysis Results
• The performance of the proposed 3D brain cortex shape analysis
methods were evaluated by applying them on both the control and the
autistic groups:
• These preliminary results show that the proposed 3D brain cortex shape
analysis methods are promising features for accurately discriminating
between autistic and control subjects
59
MetricBrain Class
Autistic Control
SH reconstruction error 233.93±28.01 241.96±37.62
Surface complexity 85.61±1.31 86.09±2.12
Tractography Analysis (1)
• A deterministic tractography approach built in 3D Slicer software[1] was
used to generate the required white matter fiber tracts
• The tractography connectivity analysis is performed in three steps:
(i) fiber orientation extraction, (ii) pathway propagation, (iii) propagation
termination
60
[1] A. Fedorov, R. Beichel, J. Kalpathy-Cramer, J. Finet, J.-C. Fillion-Robin, S. Pujol, C. Bauer, D. Jennings, F. Fennessy, M. Sonka,et al.,“3D slicer as an image computing platform for the quantitative imaging network,” Multidisciplinary Respiratory Medicine,vol. 30, no. 9, pp. 1323–1341, 2012
Tractography Analysis (2)
• After white matter fiber tracts extraction, three DTI measurements were
generated and mapped to for each fiber tract, namely, fractional
anisotropy (FA), axial diffusivity (λ∥), and radial diffusivity (λ⊥)
61
Anisotropic Diffusion in
white matter Fibers
• FA values were generated for each fiber
tract to measure the degree of anisotropy of
local diffusivity:
𝐹𝐴 =3[ 𝜆1− ҧ𝜆 2+ 𝜆2− ҧ𝜆 2+ 𝜆3− ҧ𝜆 2
2 𝜆12+ 𝜆2
2+ 𝜆32
• λ∥ and λ⊥ values, which represent diffusion parallel and
transverse to axonal directions, were also produced:
λ∥= 𝜆1 λ⊥=𝜆2+ 𝜆3
2
Connectivity Analysis Results
• The performance of the proposed connectivity analysis methods were
evaluated by applying them on both the control and the autistic groups
• These preliminary results show relatively higher values of all the
generated DTI measurements in the autistic brains, when compared to
the control brains. These initial results indicates that the extracted DTI
features are promising in differentiating between autistic and control
subjects
62
MetricBrain Class
Autistic Control
Mean Fractional Anisotropy 0.690±0.025 0.672±0.077
Mean Axial Diffusivity 5.290±3.076 4.810±3.957
Mean Radial Diffusivity 1.549±1.074 1.177±0.561
Infant Brain Classification
63
4D DWI Brain Data
INPUT
Final Assessment
Diagnosis
Control
Autistic
Segmentation
STEP 1
Brain Tissues
Segmentation
Brain Extraction
Classification
STEP 3
CSF
GM
WM
Feature Extraction
STEP 2
Brain cortex
Reconstruction
WM Tracts
Reconstruction
Connectivity
Analysis
Shape
Analysis
Diagnosis and
Brain Mapping
Mismatch
Classification (1)
• Based on the five features extracted from the infant DTI data, the
potential of each feature is tested to identify autistic and control subjects
• To perform this task, five k-means classifiers were built, using each
feature, to evaluate the diagnostic capability of each feature
64
Feature Control Autistic Overall
SH Reconstruction Error57.895%
(11 out of 19)50%
(3 out of 6)56%
(14 out of 25)
Surface Complexity42.105%
(8 out of 19)50%
(3 out of 6)44%
(11 out of 25)
Mean Fractional Anisotropy57.895%
(11 out of 19)66.667%
(4 out of 6)60%
(15 out of 25)
Mean Axial Diffusivity84.211%
(16 out of 19)33.333%
(2 out of 6)72%
(18 out of 25)
Mean Radial Diffusivity68.421%
(13 out of 19)33.333%
(2 out of 6)60%
(15 out of 25)
Classification (2)
• The preliminary results based on the available limited data sets
(25 subjects: 6 autistic and 19 control) show that some features are
outstanding candidates to distinguish between the autistic and control
groups
• The mean FA shows a potential to identify the autistic subjects with an
accuracy of 67%, where as the mean axial diffusivity shows a potential
to identify the autistic subjects with an accuracy of 84%
• To build more powerful classifier, our future work includes collecting
additional data and to fuse between the extracted features based on the
developed genetic algorithm proposed by Khalifa et al. [1]
65[1] F. Khalifa, G. Beache, M. Abou El-Ghar, T. El-Diasty, G. Gimelfarb, M. Kong, and A. El-Baz, “Dynamic contrast-enhanced MRI-based early detection of acute renal transplant rejection,” IEEE, 2013.
Thesis Conclusions
66
•This thesis proposes a novel CADsystem for the early diagnosis ofASDs using shape and connectivityfeatures, extracted from DTI images
•The accuracy of the proposed CADsystem has been validated on 25infants with a high risk of developingASDs
•The preliminary diagnostic resultsare promising in identifying autisticfrom control patients at an earlystage
Thesis Main Contributions
67
•A novel infant brain extractionapproach to automatically remove anynon-brain from the input DWI infantbrain images
•A novel atlas-based brain segmentationframework for the automatedsegmentation of different brainstructures from DTI infant brain images
•A novel NMF-based visual appearancemodel that has the ability to model alarge dimensional feature space
•A novel framework for the accurate andearly diagnosis of ASDs using DTI infantbrain images
Thesis Future Work (1)
68
•The proposed brain segmentation frameworkwill be extended to include advanced DTIfeatures with adaptive optimal level of NMFdimensionality reduction
•The proposed approach will be used to segmentother brain structures (e.g. Cerebellum,Hippocampus, Corpus Callosum,…)
• In the future, we plan to investigate theutilization of multiple classification featuresderived from spherical harmonics and brainconnectivity to be integrated in the proposedCAD system
•The proposed brain classification framework willbe extended to involve advanced identificationof brain regions that have significant differencesbetween autistic and control subjects usingconstructed brain maps
Thesis Future Work (2)
69
•The proposed CAD system for earlydiagnosis of ASDs will be tested on largerdata sets with known ground truths
•The proposed CAD system will be used toexplore longitudinal scans to trackchanges in the brain that are attributed toASDs
•Also will extend the CAD system tocharacterize other brain disorders (e.g.Alzheimer’s, dyslexia,…)
•Another future direction is applying thedeveloped models in other clinicalapplications (e.g., acute renal rejection,lung cancer detection, cancerous cellsdetection in the prostate……)
• At the end, I would like to thank my committee members:
Dr. El-Baz, Dr. Inanc, and Dr. Guru
• Also, I would like to thank all the bioimaging lab members
for their support during my work in this thesis
Acknowledgment
70
71
Thanks!!!Questions…