1. 3D-Patch Based Machine Learning Systems
Thesis Defense
3D-Patch Based Machine Learning Systems
An analysis on FDG-PET data for the clinical determination of Alzheimer’s
Anant Srivastava
Committee:
Dr. Yalin Wang
Dr. Ajay Bansal
Dr. Jianming Liang
Geometry Systems Laboratory (GSL) *

2. Table of contents Introduction Problem Background Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Experiments
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

3. Introduction Alzheimer’s Disease (AD)
“A progressive disease that destroys memory and other important mental functions. No cure exists, but medications and management strategies may temporarily improve symptoms.”
Mayo Clinic
Geometry Systems Laboratory (GSL) *

4. Problem Background
Alzheimer’s Disease is a growing concern as it affects 1 in 3 elders.
Recent years have shown immense propensity towards finding neuroimaging techniques and different biological markers
It is required that AD can be classified with high accuracy as non-AD dementias would not benefit from AD specific treatment.
Also AD progression should be tracked and distinguished from lower level of dementias.
Geometry Systems Laboratory (GSL) *

5. Alzheimer’s Disease
Is a very common disease more than 3 million U.S. cases/year.
Starts off 1-2 decades prior to the first symptoms.
Growing urgency in the early diagnosis.
Amyloid plaques build improperly and cause neuron death, The cause of which is still unknown.
Can these early developmental changes be assessed?
Geometry Systems Laboratory (GSL) *

6. Research
“The institute hosts the largest collection of brain data in the world, housing more than 4,800 terabytes of information” - USC News
ADNI PPMI multi site projects (59 accusation sites ADNI). 12 mill in 2015 for the year in a 5 year plan
Geometry Systems Laboratory (GSL) *

7. { A.D.N.I. Alzheimer’s Disease Neuroimaging Initiative
Multi Study with 59 accusation sites
Goal : Track the progression of the disease using biomarkers.
ADNIs three phases : {
ADNI 1
ADNI GO
ADNI 2
ADNI is a global research effort that actively supports the investigation and development of treatments that slow or stop the progression of AD. Biomarkers are tracers which help measurable some phenomenon such as disease, infection, or environmental exposure.
ADNI1: 5 years from October 2004
ADNIGO: 2 years from September 2009
ADNI2: 5 years from September 2011
ADNI GO
ADNI 1
ADNI 2
Geometry Systems Laboratory (GSL) *

8. Clinical Stages. CN : Cognitively Normal
EMCI : Early Mild Cognitive Impairment
MCI : Mild Cognitive Impairment
LMCI : Impairment
AD : Alzheimer’s disease
Geometry Systems Laboratory (GSL) *

9. } } Biomarkers in Alzheimer’s
The most commonly used biomarker for Alzheimer’s Disease (AD) are :
Amyloid beta imaging modality (CSF and Amyloid PET)
Neurodegeneration (FDG-PET)
Brain atrophy and neuron loss (MRI, most notably hippocampus)
Memory loss (Assessment)
General cognitive decline (Assessment) }
Pre Diagnosis }
Post Diagnosis
Geometry Systems Laboratory (GSL) *

10. MRI PET Functional (neuronal injury) Structural Still in exploration
Highly developed toolset
PET
Functional (neuronal injury)
Still in exploration
Cellular changes are important.
Geometry Systems Laboratory (GSL) *

11. Table of contents Introduction History of PET and Machine Learning
Problem Background
Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

12. History of PET and Machine Learning
FDG-PET
Statistical Voxel-Based Analysis Approach
Machine Learning
Independent component analysis (ICA), Principal component analysis (PCA) (Feature extraction)
(support vector machine )SVM, AdaBoost, Neural Nets (Classification, regression)
Sparse Coding
Geometry Systems Laboratory (GSL) *

13. History ... Patch based analysis
Brain extraction is important in analysis of brain image.
A patch is a region of voxels.
2-D patches along with sparse coding has proven to be efficient in classifying Alzheimer's Applying sparse coding to surface multivariate tensor-based morphometry to predict future cognitive decline
Geometry Systems Laboratory (GSL) *

14. Table of contents Introduction History of PET and Machine Learning
Problem Background
Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

15. Data Gathering
Geometry Systems Laboratory (GSL) *

16. Preprocessing SPM (Statistical Parametric Mapping)
Segement
Align
Original scans
Aligned & Normalized
Segmented
Geometry Systems Laboratory (GSL) *

17. Demographics
Geometry Systems Laboratory (GSL) *

18. System Design.
We design two systems for the classification of AD vs lower level dementia.
Patch Based Sparse Coding (PSC)
Patch Generator
Down Sampler
Stochastic Coordinate Coding
Adaboost classifier
Patch Based Feature Extraction (PFE)
Pooler
Feature Extraction (PCA)
Geometry Systems Laboratory (GSL) *

19. Table of contents Introduction History of PET and Machine Learning
Problem Background
Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

20. Pipeline 1 - Patch Based Sparse Coding (PSC)
Geometry Systems Laboratory (GSL) *

21. Architecture Geometry Systems Laboratory (GSL)
Input : Segmented FDG-PET scans
Geometry Systems Laboratory (GSL) *

22. Components 1. Patch Generation 10 x 10 x 10 patches
generate random overlapping patches
ensure overlap by creating many patches.
1. Patch Generation
Geometry Systems Laboratory (GSL) *

23. Components 2. Downsample Done to reduce computational time and space.
332,000,000 value matrix
Reduces the 10 x 10 x 10 (1000 feature vector) matrix to 125 vector by using a window of 2 x 2 x 2.
41,500,000
Objective matrix for Sparse Coding.
2. Downsample
Geometry Systems Laboratory (GSL) *

24. Components 3. Stochastic Coordinate Coding
Used in audio and image processing domain
Model the data vectors as sparse linear combination of basis elements.
Over complete dictionaries may represent data more flexibly
3. Stochastic Coordinate Coding
Geometry Systems Laboratory (GSL) *

25. Components 3. Stochastic Coordinate Coding
Patches of 1st sample
Patches of 2nd sample
Patches of nth
sample
Input Matrix
Dictionary learning
Geometry Systems Laboratory (GSL) *

26. Components 3. Stochastic Coordinate Coding
Take an image patch xi
Perform few steps of coordinate descent to find the support (non zero entries) of the sparse code.
Update the support of the dictionary by second order stochastic gradient descent to obtain a new dictionary
10 epochs
Geometry Systems Laboratory (GSL) *

27. Components 4. AdaBoost Geometry Systems Laboratory (GSL)
Strong classifier based on weak learners
Use weights to force weak learners to classify generally misclassified examples.
Figure on the right -
Round 1 : First weak classifier (left vertical line)
Round 2 : Second weak classifier (right vertical line)
Round 3 : Third weak classifier (horizontal line)
Final Adaboost classifier.
Geometry Systems Laboratory (GSL) *

28. Pipeline 2 - Patch Based Feature Extraction (PFE)
Geometry Systems Laboratory (GSL) *

29. Pipeline - P.F.E. Geometry Systems Laboratory (GSL)
Input : Segmented FDG-PET scans
Geometry Systems Laboratory (GSL) *

30. Components 1. Patch Generation 10 x 10 x 10 patches
Generate structured uniform overlapping patches
1. Patch Generation
Geometry Systems Laboratory (GSL) *

31. Components 2. Pooling Feature Selection Technique (R80x95x80 R1x4050)
(reduce noise, highlight relevant feature in a patch, translation invariant)
2. Pooling
3-D
Max-pooling
2-D
Geometry Systems Laboratory (GSL) *

32. Components 3. Feature Extraction Probabilistic PCA (M.E. Tipping)
To further reduce dimensionality
probabilistic version of PCA
deals with missing data
finds MLE (maximum likelihood estimate) ~ Principal components
3. Feature Extraction
Geometry Systems Laboratory (GSL) *

33. Components 4. AdaBoost As seen previously
Geometry Systems Laboratory (GSL) *

34. Table of contents Introduction History of PET and Machine Learning
Problem Background
Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

35. Experiments.
Geometry Systems Laboratory (GSL) *

36. Results Run the experiments over the two proposed pipelines PSC PFE
Comparing sets of features
Meta analysis
PSC
comparing classifier
no. of patches involved in training
PFE
comparing classifiers
comparing feature extraction techniques. *

37. Comparing sets of features
AF = Apoe + FAQ
AFAG = Apoe + FAQ + Age + Gender
AD vs. CU
stages with greatest class separation
voxels + AFAG ~ 96%
AD vs. EMCI
stages with class separation of 2
voxels + AFAG ~ 83%
Geometry Systems Laboratory (GSL) *

38. Comparing sets of features
Geometry Systems Laboratory (GSL) *

39. PSC -classifier comparison
NN : nearest Neighbour, SVM: support vector machine
AdaBoost & SVM gives near 97% F1 measure, 2000 patches
Geometry Systems Laboratory (GSL) *

40. PSC -number of patches patches of 500, 1000, 2000. F1 Measure
In exp. with high class separation f1 increases with number of patches
Decode Time is proportional to the number of patches.
Geometry Systems Laboratory (GSL) *

41. PFC -classifier comparison
NN : nearest Neighbour, SVM: support vector machine
(Gaussian performs the best) ,PCA , Adaboost more stable
Geometry Systems Laboratory (GSL) *

42. PFC -feature extraction comparison
PCA : principle component analysis, SVD: singular valued Decomposition
PCA’s performs vary when compared to other feature extraction methods
Geometry Systems Laboratory (GSL) *

43. AUC Curves
Geometry Systems Laboratory (GSL) *

44. Table of contents Introduction History of PET and Machine Learning
Problem Background
Alzheimer’s Disease
History of PET and Machine Learning
Method
Data Gathering
Preprocessing
System Design
Pipeline - P.S.C.
Components
Pipeline - P.F.E.
Result
Conclusion
Future Work
Questions ?
Geometry Systems Laboratory (GSL) *

45. Conclusion
Comprehensive experimentation shows the effectiveness of 3D patch based methods.
We examine other rich features like the APOE gene information, FAQ scores Age and Gender.
huge boost when other features are involved . 90% to 96%
With experiments with lower group separability the methods struggled.
May indicate lower cellular metabolism difference. *

46. Future Work This method can be used to classify DTI images
PET imaging is effective in detecting Parkinson's and so these methods can be applied to detect Parkinson's
Use of autoencoders similar to Sparse coding in the pipeline might give interesting results.
Geometry Systems Laboratory (GSL) *

47. Publications
MICCAI 2017 (under review) paper titled, “3D-Patch Based Sparse Coding System for Alzheimer’s Disease Diagnosis via FDG-PET Analysis”
AAC May 2017 (poster publication) Arizona Alzheimer’s Consortium Scientific Conference. Abstract titled “3D-Patch Analysis Based Sparse Coding System for Alzheimer’s Clinical Group Classification ” *

48. Questions?
Geometry Systems Laboratory (GSL) *

49. Appendix

50. Appendix

51. Appendix

52. Appendix

53. Appendix