AdapHBNA: Adaptive hierarchical spatio-temporal brain network analysis for brain disease detection.

Abstract

Brain Network Analysis (BNA) from resting-state functional MRIs (rs-fMRIs) has been widely applied to the prediction and understanding of brain disorders, by modeling connectivities among brain regions of interest (ROIs) to identify potential biomarkers. However, the majority of existing studies construct static brain networks utilizing a single predefined spatial scale (i.e., the number of brain ROIs), neglecting the inherently hierarchical nature of brain networks across various temporal and spatial scales. To address these limitations, we propose AdapHBNA, an Adaptive Hierarchical spatio-temporal Brain Network Analysis framework for brain disorder diagnosis. Specifically, we incorporate feature-channel-guided temporal hierarchical learning and Modular Brain Clustering (MBC)-driven spatial hierarchical learning strategies into the spatio-temporal encoding process of brain network, utilizing Mamba and Graph Neural Networks. It seamlessly merges the multi-scale learning, hierarchical brain representation learning and automatic spatio-temporal fusion into a unified end-to-end framework, which adaptively adjusts the temporal and spatial scales to capture hierarchical complementary brain representations across a spectrum of fine-to-coarse granularities. Extensive validation on the ABIDE, ADNI and REST_MDD datasets for Autism Spectrum Disorder, Early Mild Cognitive Impairment and Major Depressive Disorder demonstrates that AdapHBNA outperforms state-of-the-art methods by leveraging complementary diagnostic insights across multiple scales.