Presentation delivered by Pradeep Reddy Raamana at 2016 international workshop on Pattern Recognition in Neuroimaging on the topic of histogram-weighted cortical thickness networks for the detection of Alzheimer's disease.
Histogram-weighted cortical thickness networks for the detection of Alzheimer's disease
1. Histogram-weighted cortical
thickness networks for the
detection of Alzheimer's disease
Pradeep Reddy Raamana, PhD
Postdoctoral fellow
Research interests:
Machine learning
Medical image analysis
5. Limited Power of Thickness
• Cortical thickness is a great imaging biomarker, but its prognostic
power is limited.
• According to an extensive comparison study Cuingnet et al. (2011):
Example Study
on CN vs. MCIc
Type of
Cortical Thickness
Sensitivity Specificity Satisfactory?
Klöppel et al. (2008) Direct & Raw 54% 96% No
Desikan et al. (2009) Summarized 57% 93% No
Marcus et al. (2007) ROI-based 65% 94% No
5MCI: Mild cognitive impairment, MCIc: MCI converters.
10. Multiple Kernel Learning
• Typically, many features
combined into a single bag
to train SVM
• Combining the ThickNet
features to maximize
collective predictive power
• Using Variational Bayes
probabilistic MKL (VBpML)1
10
Composite
Classifier
Optimized
Kernel 1
ThickNet
Feature 1
Optimized
Kernel 2
ThickNet
Feature 2
Optimized
Kernel 3
ThickNet
Feature 3
1 Damoulas, T., & Girolami, M. A. (2008). Bioinformatics, 24(10), 1264-1270.
11. • Class-imbalance is not uncommon.
• Classifiers can be sensitive to class-imbalance
• The most popular classifier SVM is.
• This can result in biased estimates, making the
classifier either too sensitive, or too specific.
Repeated Holdout, Stratified
Training Set ( RHsT )
Controls (n=159) MCIc (n=56)
Training (MCIc)Training (CN) Test Set (CN)
11
Tes
14. Alzheimer’s Dataset
• Evaluated on ADNI-1 to
enable comparison to
published literature.
• Exact subset as
published in Cuingnet et
al., Neuroimage, 2011.
• except for exclusions
from quality-control.
Diagnostic Group #Subjects
Healthy controls (CN) 159
Alzheimer’s disease (AD) 136
MCI converters (MCIc) 56
MCI non-converters (MCInc) 130
Total 481
14ADNI: Alzheimer Disease Neuroimaging Initiative
15. Best Performance
15Results in Raamana et al. 2014, Neurobiol. Aging.
0.5
0.6
0.7
0.8
0.9
1
CN vs. AD CN vs. MCIc MCIc vs. MCInc
0.64
0.76
0.9
0.65
0.74
0.8
0.64
0.76
0.89
0.68
0.83
0.92
AUC Accuracy Sensitivity Specificity
16. Improvement over Thickness
Full AUC
0.75
0.813
0.875
0.938
1
CN vs. AD CN vs. MCIc
0.832
0.924
0.807
0.916 Mean Thickness
Thickness
16AUC: Area under ROC; Partial AUC is bounded by specificity > 85%
17. Partial AUC
0.05
0.063
0.075
0.088
0.1
CN vs. AD CN vs. MCIc
0.068
0.097
0.057
0.09
Mean Thickness
ThickNet
Improvement over Thickness
17AUC: Area under ROC; Partial AUC is bounded by specificity > 85%
18. Summary
• A predictive model with attractive properties:
• individual feature significance
• most discriminative regions
• improved classification power
• intuitive interpretation.
18
19. Few applications
1. Early detection of
Alzheimer Disease
2. Amnestic MCI
sub-classification
3. Differential diagnosis
of AD and
Frontotemporal
Disease (FTD)
19
overlap
Normal Aging
ADFTD
Others
(VaD etc.)
Healthy
Prodromal
dementia
Dementia
Mild cognitive
impairment (MCI)
Single
domain
Multi-
aMCI
MCI
AD: Alzheimer disease, MCI: Mild cognitive impairment, VaD: vascular dementia
22. Edge definitions
Type of base
distribution
Type of
Edge Metric
Acronym Definition
Summarized
Similarity (diff. in medians) MD
exp(similarity) EMD
raw distribution Wilcoxon ranksum statistic RS ranksum statistic
normalized
histogram
histogram correlation HCOR
𝝌2 statistic CHI2
histogram intersection HINT
spatial scale: m= 400, 1000, 2000, 3000, 5000 and 10000 vertices per patch.
718, 273, 136, 97, 74 and 68 patches per brain.
24. Alzheimer’s Dataset
• Evaluated on ADNI-1 to enable comparison to published
literature.
• Only CN and AD are chosen to focus the comparison on
edge metrics alone.
Diagnostic Group N #Females Age MMSE
Healthy controls (CN) 224 109 75.79 ( 4.99) 29.11 ( 1.01)
Alzheimer’s disease (AD) 188 89 75.22 ( 7.49) 23.29 ( 2.04)
Total 412 198 only MMSE differed significantly
24ADNI: Alzheimer Disease Neuroimaging Initiative
31. Summary
• Simpler methods (MD) are just as predictive.
• Impact of spatial scale m on predictive
performance seems to be not significant.
• If I may overstate it, the most popular way of
computing edge weights in group-wise analysis i.e.
correlation, seems to be the least-predictive of
disease-status.
31
32. Future work
• Further validate it on
• another separability, such as MCI vs. CN.
• another dataset, such as AIBL.
• another disease, such as FTD.
• another parcellation scheme!
• Compare individual graph properties