mBMI: a multimodal brain multiplex interaction network for multimodal brain network analysis

Multi-modal brain network analysis (MBA) is an important research area in neuroscience. Complex interactions between brain regions are the key to brain activity. Existing methods focus on network fusion, ignore the interaction of correlated information in brain function and structure, and are unable to suppress the multi-modal redundant information and extract effective features, which limits the effectiveness of existing analysis models. Therefore, in this paper, we propose a multi-modal Brain Multiplex Interaction network (mBMI) for multi-modal brain network analysis. The Fine-Grained Relevancy Unit (FGR-Unit) uses association encoders to explore the correlation of structural and functional connectivity, enabling brain region-level interactions of complementary information from the brain's structural and functional networks. Aiming at the sparse topological characteristics of brain data, Global Interaction Network (GI-Net) combines positional encoding to capture long-distance relationships in brain regions, and constructs cross-modal interaction attention to learn multi-modal interactions at the whole brain-level and reduce modal redundancy information. Further, the extensive experiments on the Human Connectome Project dataset (HCP) and the Major Depressive Disorder (MDD) Disease dataset demonstrate the effectiveness of the proposed framework.