Deep Learning for Neuroimaging of Disease
Sergey Plis
- introduction
- deep learning
- training data generation
- ongoing work
introduction
brain imaging must be used to inform brain disorder understanding
lets collect more data at the finest resolution
Problem: collected data is high dimensional and a disorder often impacts the brain in many ways
Solution: methods that automatically extract meaning from data
Deep learning
understanding large scale data
deep models
stacking game
stacking game (cont.)
Tensorflow playground example
Convolutional Neural Network
convolution
Convolutional Neural Network
architecture
higher level concepts
harder to interpret for MRI
MNIST dataset
Van der Maaten, L. et al., Journal of
Machine Learning Research 9, 85 (2008)
Van der Maaten, L. et al., Journal of
Machine Learning Research 9, 85 (2008)
400 brains
schizophrenia vs. healthy controls
depth effect (dynamic)
depth effect (static)
Plis, S. et al., Frontiers in neuroscience 8 (2014)
3500 brains
Huntington's disease
severity within prodromal HD
CAG-length Age Product (CAP) score- Low
score
very similar to healthy controls in cognitive and motor performance - Medium
score
increasing deficit - High
score
largest deficit in cognitive and motor performance though not yet clinically diagnosable
gray matter concentration differences by group
- Here are the
groups
What are the brain differences? - Here are the
brains
What are the differences?
Plis, S. et al., Frontiers in neuroscience 8 (2014)
Various DL Projects
simulating data
synthetic-data pre-training
multimodal deep learning
multimodal deep learning
multimodal deep learning: SNP + MRI
"translation" of structure to function
segmentation: dilated convolution
segmentation: architecture
segmentation: results
segmentation: multimodal
Open Questions
- Can we simulate data using GAN or AAE?
- Can MeshNet directly generate surfaces?
- How to visualize what feature caused networks decision?
Capture SNPs
