HieMaGT: Hierarchical Multi-Scale Graph Transformer for Brain Disorder Diagnosis

Brain disorder diagnosis using functional magnetic resonance imaging (fMRI) is crucial but challenging, largely due to the brain's complex, multi-scale, and dynamic nature. Existing models often fall short by focusing on single-scale or pairwise connections, or by requiring predefined higher-order interactions. To address these limitations, we propose HieMaGT (Hierarchical Multi-scale Graph Transformer), a novel end-to-end framework designed to adaptively learn dynamic, higher-order, and multi-scale functional connectivity directly from fMRI time series. HieMaGT integrates parallel Multi-scale Graph Transformer layers to capture interactions across various granularities, a Hierarchical Pooling module for progressive feature abstraction, and a Robustness Enhancer based on contrastive learning to ensure stable and generalizable disease biomarkers. Comprehensive experiments on three real-world fMRI datasets for conditions like schizophrenia, Alzheimer's Disease, and various brain states demonstrate that HieMaGT consistently achieves superior diagnostic performance. HieMaGT significantly outperforms state-of-the-art methods, showing substantial improvements across all datasets. These results highlight HieMaGT's advanced capability in leveraging complex brain functional interactions for accurate and robust brain disorder diagnosis.