Date: 21 Apr 2016 Local: University College London Event: Machine Learning for Medical Imaging Reading Group

1. Introduction to Deep Learning
and some Neuroimaging
Applications
Walter Hugo Lopez Pinaya
Universidade Federal do ABC (UFABC), São Paulo – Brazil
Supervisor: João Ricardo Sato
PhD student in Neuroscience and Cognition
walter.lopez@kcl.ac.uk

2. Agenda
What is Machine Learning?
What is Deep Learning?
How Deep Learning works?
Deep learning for neuroimaging
2

3. Machine Learning
3

4. Write programs that
solve the problem
4

5. Write programs that
learn to
solve the problem
from examples
5

6. What is Machine Learning?
•Algorithms that can learn from and make predictions on
data
•Computational statistics and mathematical optimization to
discover trends and patterns
6

7. Supervised Learning
Classification and Regression 7
Outcome
Responses to treatment
Diagnosis
Machine Learning Model
Voxels
Pixels
Clinical data
Features…

8. Traditional Machine Learning
8

9. Traditional Machine Learning
9

10. Audio Features
10

11. ComputerVision Features
11

12. HandTuned Features
Careful engineering and considerable domain expertise
Substantial cost in terms of knowledge and computer
time
12

13. Good Features
are essential for a
successful Machine Learning
80% ~ 90% of effort
13

14. Good Representations?
14

15. 15

16. Deep Learning
16

17. 17
Let’s learn rather than
manually design our
features

18. Multiple levels of representation of increasing complexity
18

19. What is Deep Learning?
Algorithms that exploit the unknown input data structure to extract
multiple levels representations
Higher-level learned features as a non-linear composition of lower-level
concepts
High-level concepts are more invariant to most of the variations that
are frequently present in the input data
19

20. Applications
• Speech recognition
• Computer Vision
• Object recognition
• Object detection
• Natural language processing
20

21. Applications
• Medical Research
21
Detecting Mitosis in
Breast Cancer Cells
-IDSIA
Predicting the Toxicity
of new drugs
-Johannes Kepler University
Understanding Gene Mutation
to prevent Disease
- University of Toronto

22. How it works?
22

23. Artificial Neural Network
23

24. 24
Repeat
Forward Propagation
Backward Propagation
“Turtle”
Compute weight update to nudge
from “turtle” towards “dog”
Inference
Trained Neural
Net Model
“Cat”
Training

25. The Problem with Large Networks
Optimization is hard (Underfitting)
• Vanishing gradient problem
• Backpropagation becomes less useful in
passing information to the lower layers
25

26. The Problem with Large Networks
Overfitting
• We are exploring a space of complex functions
• Deep nets usually have lots of parameters
• Fitting the training data too closely
26

27. Breakthrough in 2006
Use of layer-wise unsupervised features learning
27

28. 28

29. Deep Supervised Neural Nets (~2013)
Now we can train them without unsupervised
pre-training:
• Better initialization
• Regularization
• Nonlinearities
Unsupervised pre-trained:
• Rare classes, smaller labelled sets, or as extra
regularization
29

30. Examples of networks
30

31. Convolutional Neural Network
• Inspired by human visual cortex
• Local pixel level features are scale and translation invariant
31

32. Winner ILSVRC 2014
GoogleLeNet: 22 Layers, intermediate targets
32

33. Combinations of layers
33

34. Revolution of depth
Winner ILSVRC 2015
• GoogleLeNet Error rate = 6.7
• ResNet Error rate = 3.57
34

35. Recurrent Neural Networks
Sequential data
• Videos
• Connected handwriting
Natural language processing
• Language modelling
• Machine translation
35

36. Combination of networks
LECUN,Yann; BENGIO,Yoshua; HINTON, Geoffrey. Nature, 2015. 36

37. Deep Learning
• Deep Boltzmann Machine (DBM);
• Generative Stochastic Networks (GSN);
• Neural Autoregressive Distribution Estimator (NADE)
• Stacked Denoising Autoencoders (SdA);
• Gated Recurrent Units (GRU);
• Generative Adversarial Networks (GAN);
• Ladder Networks;
• ….
37

38. Deep Learning DeploymentWorkflow
38

39. Deep Learning in a nutshell
• Allows computer systems to improve with experience and data
• Great power and flexibility by learning to represent the data as a nested
hierarchy of concepts
• Each concept defined in relation to simpler concepts, and more abstract
representations computed in terms of less abstract ones
• No free lunch: Tendency for overfitting
39

40. 40

41. 41

42. Effect of the depth of a DBN on sMRI
Schizophrenia patients
• Structural MRI (1.5T)
• 389 subjects (198 Control/191 Schizophrenia)
• 60465 voxel gray matter images
PLIS, Sergey M., et al. Frontiers in neuroscience, 2014. 42

43. PLIS, Sergey M., et al. Frontiers in neuroscience, 2014. 43

44. 2D Maps
PLIS, Sergey M., et al. Frontiers in neuroscience, 2014. 44

45. Large-Scale Huntington Disease Data
• Dataset of 3500 structural MRI scans
• 2641 were from patients and 859 from healthy
controls
• Model architecture  DBN (50-50-100)
PLIS, Sergey M., et al. Frontiers in neuroscience, 2014. 45

46. 46

47. Hierarchical feature representation and
multimodal fusion
Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset
• MRI and 18-Fluoro-DeoxyGlucose PET (FDG-PET) data
• 93 AD subjects
• 204 MCI subjects including MCI converters and MCI non-converters
• 101 NC subject
• Downsampled GM density maps and PET images to 64×64×64 voxels
SUK, Heung-Il, et al. NeuroImage, 2014. 47

48. Multimodal fusion
SUK, Heung-Il, et al. NeuroImage, 2014. 48

49. Hierarchical feature representation
SUK, Heung-Il, et al. NeuroImage, 2014. 49
MRI
PET

50. Comparison of classification accuracy with state-
of-the-art methods
50

51. Hierarchical feature representation and
multimodal fusion
SUK, Heung-Il, et al. NeuroImage, 2014. 51

52. Deep Belief Networks modeling schizophrenic patients
using neuromorphometric features
Walter H. L. Pinaya a; Ary Gadelha b; Orla M. Doylec, Cristiano Noto b; André
Zugman c; Quirino Cordeiro b, d; Andrea P. Jackowski b; Rodrigo A. Bressan b; João
R. Sato a, b
52

53. Objectives
• Use of the DBN to explore and extract latent features from brain morphometry
data of healthy controls subjects and Schizophrenia patients
53

54. Data
Structural MRI (1.5 T)
• 146 chronic schizophrenia patients
• 83 healthy controls
• Cortical thickness and volume of anatomical structures (113 variables)
54

55. Results
55
From error rate=29.8% to 25.6%

56. Conclusion
•Learning methods have recently made notable advances in the
tasks of classification and representation learning.
•These tasks are important for brain imaging and neuroscience
discovery, making the methods attractive for porting to a
neuroimager’s toolbox.
56

57. Acknowledgments
Joao Ricardo Sato (UFABC, Brazil) Supervisor
Joana Balardin (UFABC, Brazil)
Ary Gadelha (UNIFESP, Brazil)
Rodrigo A. Bressan (UNIFESP, Brazil)
Andrea P. Jackowski (UNIFESP, Brazil)
Quirino Cordeiro(UNIFESP, Brazil)
Orla Doyle (KCL, UK) Supervisor
Steven Williams (KCL, UK)
57

58. Acknowledgments
58

59. Thank you
59