New GNN Improves Alzheimer's Classification with Adaptive Brain Modules

Functional magnetic resonance imaging (fMRI) is a crucial tool for studying the brain, and recent advancements using graph neural networks (GNNs) have shown promise in classifying brain disorders like Alzheimer's disease (AD). However, existing GNN methods often assume a fixed number of functional modules across all subjects, overlooking the significant variability between individuals. Additionally, the discovered modules are rarely used to directly inform the learned connectivity patterns within the network. To address these limitations, researchers proposed the Meta Probabilistic Pooling GNN (MPP-GNN). This model frames its task as a coupled, bilevel optimization process. It adaptively partitions the brain graph hierarchically to identify subject-specific functional modules. Crucially, these discovered brain modules then serve as an explicit prior, guiding the refinement of edge connectivity and the learning of representations within the GNN. MPP-GNN was validated on two public datasets for AD classification, achieving superior AUC scores compared to established baseline methods. Further analysis revealed that MPP-GNN's discovered modules align significantly with the canonical functional-network organization defined by the Yeo brain atlas, and it effectively uncovers a network-level dedifferentiation pattern associated with Alzheimer's disease.