Cross-Attentioned Dynamic Hierarchical Representation Learning Utilizing Mamba Fusion for Brain Network Analysis

Functional brain networks (FBNs) derived from resting-state functional Magnetic Resonance Imaging (rs-fMRI) have become pivotal tools for the auxiliary diagnosis of brain disorders, including Alzheimer's Disease (AD) and Major Depressive Disorder (MDD). However, most existing FBN analysis methods struggle to effectively capture both the temporal dynamics of rs-fMRI data and the hierarchical topological structures within FBNs, due to the inherent spatiotemporal complexity of brain activities. To address these challenges, we propose a novel framework for brain network analysis, called Crossattentioned Dynamic Hierarchical representation learning with Mamba fusion (CDHM). CDHM includes four principal phases: (1) dynamic FBNs construction via overlapping sliding windows, (2) cross-attentioned spatial encoding to capture local-to-global spatial interactions, (3) hierarchical graph pooling to distill multiscale FBN organisation, and (4) Mamba-based fusion modelling long-range temporal dependency and adaptive information flow regulation. Our framework pioneers synergistic integration of spatial cross-attention mechanisms with Mamba-based temporal fusion, enabling joint learning of transient neural dynamics and multi-scale hierarchical brain representations. We validate our approach through comprehensive experiments on two publicly available datasets, demonstrating superior performance compared to existing methods.