Integrating multi-atlas neuroimaging data for robust biomarker identification in neuropsychiatric disorders.

Introduction: Accurate diagnosis of neuropsychiatric disorders like autism spectrum disorder (ASD) and post-traumatic stress disorder (PTSD) is challenging due to complex disruptions in brain functional connectivity. Existing graph neural network (GNN) methods are often limited by reliance on single-atlas representations. Methods: We propose MSAT-LAFFNet, a multi-atlas GNN framework integrating a Structure-Aware Graph Transformer (SAT) for enhanced structural encoding and a Lightweight Attentional Feature Fusion network (LAFFNet) for adaptive cross-atlas fusion. The model was validated on the public ABIDE-I dataset and a private PTSD dataset (n=138). Results: Our model achieved an AUC of 82.9%/accuracy of 81.59% on ABIDE-I (ASD) and an AUC of 89.96%/accuracy of 89.45% on the PTSD dataset, outperforming competing methods. Interpretability analysis identified disorder-specific overlapping abnormal regions, including the right middle frontal gyrus in ASD and the left amygdala in PTSD, linking them to known network dysfunctions. Conclusion: MSAT-LAFFNet demonstrates superior classification performance and provides neurobiologically interpretable findings, showing potential as an effective tool for the auxiliary diagnosis of disorders characterized by disrupted functional networks.