MSPPGT: A Multi-Scale Pyramid Pooling-Based Graph Transformer Framework for Neuropsychiatric Disorder Diagnosis.

Abstract

Graph Neural Networks (GNNs) model functional connectivity patterns between brain regions via neighborhood information aggregation. However, most GNN based methods primarily focus on short-range dependencies between brain regions, making it difficult to mine the full-scale functional connectivity patterns of the entire brain. While Transformer captures long-range dependencies via self-attention mechanism, they suffer from unnecessary computational overhead for local features. To address this, we propose a Multi-Scale Pyramid Pooling-based Graph Transformer (MSPPGT) framework for neuropsychiatric dis order diagnosis. Specifically, it extracts graph features at multiple scales and simultaneously extracts multilevel inter actions between local and global brain regions. The Multi Scale Pyramid Pooling (MSPP) module integrates features from multiple scales into the Multi-Head Self-Attention (MHSA) module, enhancing the perception of multi-scale in formation while reducing computational complexity. The MHSA module also quantifies the interaction strength between brain regions to identify discriminative brain regions and improving interpretability. On public datasets ABIDE and REST-meta-MDD, the proposed framework achieved state-of the-art accuracies of 72.2% for Autism Spectrum Disorder (ASD) and 65.5% for Major Depressive Disorder (MDD), respectively. These results validate the effectiveness of MSPPGT and demonstrate that it outperforms prior methods in neuropsychiatric disorder diagnosis.