Depression Classification from rs-fMRI: A Local-to-Global Graph Neural Network Approach

Graph Convolutional Networks (GCNs) have revolutionized the analysis of complex diseases by mapping the intricate brain connectivity data inherent in brain disorders. Building upon these foundations, this research explores the novel application of the Local-to-Global Graph Neural Network (LG-GNN) framework to Major Depressive Disorder (MDD), leveraging resting-state functional magnetic resonance imaging (rsfMRI) data. This adaptation refines a proven model to uncover both localized and systemic disruptions characteristic of MDD. The approach integrates a local region of interest (ROI) based GCN, which identifies critical biomarkers and extracts pivotal features from distinct brain regions, with a global GCN that synthesizes these data points with supplementary non-imaging information to articulate a comprehensive view of network pathologies. This refined application enhances MDD classification by capitalizing on both the relational and attribute-based information of brain networks. By extending the use of sophisticated feature learning and comprehensive network analysis to MDD, this research not only broadens the model’s applicative spectrum but also significantly enhances the classification accuracy of MDD. This improvement underscores the model's capability to delineate depression's underlying mechanisms more clearly, potentially facilitating earlier and more precise diagnostic and therapeutic interventions.