Dynamic Graph Transformer for Brain Disorder Diagnosis

Abstract

Dynamic brain networks are crucial for diagnosing brain disorders, as they reveal changes in brain activity and connectivity over time. Previous methods exploit the sliding window approach on fMRI data to construct these networks. However, this approach encounters two major issues: fixed temporal length, which inadequately captures the temporal dynamics of brain activity, and global spatial scope, which introduces noise and reduces sensitivity to localized dysfunctions when applied across the entire brain. These issues can lead to inaccurate brain network representations, potentially resulting in misdiagnosis. To overcome these challenges, we propose BrainDGT, a dynamic Graph Transformer model that adaptively captures and analyzes modular brain activities for improved diagnosis of brain disorders. BrainDGT addresses the fixed temporal length issue by estimating adaptive brain states through deconvolution of the Hemodynamic Response Function (HRF), avoiding the constraints of fixed-size windows. It also addresses the global spatial scope issue by segmenting fMRI scans into functional modules based on established brain networks for detailed, module-specific analysis. The model employs a dual attention mechanism: graph-attention extracts structural features from dynamic brain network snapshots, while self-attention identifies significant temporal dependencies. These spatio-temporal features are adaptively fused into a unified representation for disorder classification. BrainDGT’s effectiveness is validated through classification experiments on three real fMRI datasets ADNI, PPMI, and ABIDE demonstrating superior performance compared to state-of-the-art methods. BrainDGT improves brain disorder diagnosis by offering adaptive, localized analysis of dynamic brain networks, advancing neuroimaging and enabling more precise treatments in biomedical research.