Sub-Connection Learning for fMRI-Based Brain Functional Network

Brain functional network analysis models the brain as a graph of regions of interest (ROIs) and quantifies the correlations across different regions derived from functional magnetic resonance imaging (fMRI). Recently, artificial intelligence-based brain functional network analysis methods have demonstrated exceptional performance in diagnosing related neurological disorders. These approaches primarily focus on extracting relevant information from global connectivity patterns to analyze functional brain networks. However, medical research indicates that the impact of brain disorders predominantly manifests in localized functional connections among disease-relevant regions. Treating all connections equally risks introducing interference from irrelevant brain regions, thereby compromising diagnostic accuracy. To address this issue, we propose a novel sub-connection learning method that effectively identifies diagnostically specific connections while suppressing ineffective redundant connections. Specifically, we begin by employing a dynamic functional connectivity construction strategy to generate a functional connectivity matrix encapsulating dynamic features. Subsequently, we design a sub-connection Mask Learning strategy, which employs a multi-head self-attention mechanism to adaptively learn connection masks from functional connectivity matrices, enabling the capture of disease-specific connections and the suppression of noise connections. Additionally, we introduce a Multi-mask Fusion strategy and a Mask Iterative Optimization strategy to further enhance mask quality. Experimental results demonstrate that our model outperforms state-of-the-art methods on the ABIDE I and ADNI datasets, achieving accuracies (ACC) of 72.30% and 80.99%.