Deep Graph Models of Functional Brain Networks
Abstract
Modeling activity in the human brain as a graph offers many potential uses such as the identification of bio-markers, the effects of pathologies and modeling cognitive networks. However, existing methods of creating graphs for medical imaging greatly simplifies the complex behavior in brains to the most simplest of graphs and then typically simplifies/analyzes these graphs in a single step. The end result is several significant compromises most notably the representation (the graph) is learnt independent of the analysis/computation. The deep learning approach of simultaneously learning a useful representation for a given problem will be applied to graph analysis of medical imaging. A novelty of our work is that unlike previous discriminative deep learning models for classification our models are generative learners. Our goal is directly learning a graph from the raw time series whilst also simultaneously learning a calculation/computation on the learnt graph. Most importantly our method does not learn a single graph graph structure, but rather a multitude of graph structures of different granularity. We explore three deep learning models: an auto-encoder that learns a variety of graphs of various granularity, a Siamese network to learn the graphs properties that differentiates two cohorts and finally a deep network to perform block modeling of information flow where the number of blocks can vary. Importantly we explore interpreting and explaining the results of these methods to a domain scientist using visualization.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 26, 2018
- Source ID
- N000141812485
Entities
People
- Ian Davidson
Organizations
- Office of Naval Research
- United States Navy
- University of California, Davis