A connectome is a graph: brain regions are nodes, connectivity strengths are edges. Standard deep learning architectures require flattening that graph into a vector, which throws away the relational structure and forces the model to relearn it from data it does not have enough of. Graph neural networks operate on the graph directly.
The practical payoff is parameter efficiency. Because a GNN shares weights across nodes rather than learning a separate parameter per connectome entry, it needs far fewer parameters to reach the same accuracy — which matters a great deal when your clinical cohort has a few hundred subjects.
Why it appears in this program¶
Both projects using GNNs are prediction problems on brain networks where interpretability is as important as accuracy. Methods such as GNNExplainer can probe a trained model to identify which specific edges or regions drove a prediction, which is what makes the output clinically meaningful rather than just a score.
CNS Projects¶
Josh Neudorf — GNN deep learning to predict progression from mild cognitive impairment to dementia, integrating structural and functional connectivity in one model along with regional attributes (graph theory measures, neurotransmitter expression, cortical thickness) and individual-level covariates
Sima Abbasi Habashi — graph learning as part of a multi-modal pipeline linking genetic variation to white matter hyperintensity burden
Getting started¶
PyTorch Geometric — the most developed technical document in this book, with a worked neuroimaging example
Deep learning — general context