BrainGT: Multifunctional Brain Graph Transformer for Brain Disorder Diagnosis

Brain networks play a crucial role in the diagnosis of brain disorders by enabling the identification of abnormal patterns and connections in brain activities. Previous studies exploit the Pearson’s correlation coefficient to construct functional brain networks from fMRI data and use graph learning to diagnose brain diseases. However, correlation-based brain networks are overly dense (often fully connected), which obscures meaningful connections and complicates subsequent analyses. This dense connectivity poses substantial performance challenges to traditional graph transformers, which are primarily designed for sparse graphs. Consequently, this results in a notable reduction in diagnostic accuracy. To address this challenging issue, we propose a multifunctional brain graph transformer model for brain disorders diagnosis, namely BrainGT, which is capable of constructing multifunctional brain networks rather than a dense brain network from fMRI data. It utilizes the fusion of self-attention and cross-attention mechanisms to learn important features within and across multiple functional brain networks. Classification (diagnosis) experiments conducted on three real fMRI datasets (i.e., ADNI, PPMI, and ABIDE) demonstrate the superiority of the proposed BrainGT over state-of-the-art methods.