IMG-GCN: Interpretable Modularity-Guided Structure-Function Interactions Learning for Brain Cognition and Disorder Analysis

Brain structure-function interaction is crucial for cognition and brain disorder analysis, and it is inherently more complex than a simple region-to-region coupling. It exhibits homogeneity at the modular level, with regions of interest (ROIs) within the same module showing more similar neural mechanisms than those across modules. Leveraging modular-level guidance to capture complex structure-function interactions is essential, but such studies are still scarce. Therefore, we propose an interpretable modularity-guided graph convolution network (IMG-GCN) to extract the structure-function interactions across ROIs and highlight the most discriminative interactions relevant to fluid cognition and Parkinson's disease (PD). Specifically, we design a modularity-guided interactive network that defines modularity-specific convolution operation to learn interactions between structural and functional ROIs according to modular homogeneity. Then, an MLP-based attention model is introduced to identify the most contributed interactions. The interactions are inserted as edges linking structural and functional ROIs to construct a unified combined graph, and GCN is applied for final tasks. Experiments on HCP and PPMI datasets indicate that our proposed method outperforms state-of-the-art multi-model methods in fluid cognition prediction and PD classification. The attention maps reveal that the frontoparietal and default mode structures interacting with visual function are discriminative for fluid cognition, while the subcortical structures interacting with widespread functional modules are associated with PD.