Dynamic Cross-Atlas Interaction Framework for Robust Autism Spectrum Disorder Classification

Autism Spectrum Disorder (ASD) is a prevalent psychiatric condition where timely diagnosis and intervention are critical for patient outcomes. In this study, we propose a novel dynamic cross-atlas interaction Transformer framework (DCAIFormer). The method is used to resolve the existing problem that synergistic information loss caused by independent processing of multi-atlas features. In the DCAI-Former, a feature calibration encoder is implemented to eliminates redundant information. Then, an innovative cross-atlas attention mechanism facilitates dynamic inter-atlas correlations among AAL, CC200, and DOS160 atlases. Finally, The framework incorporates a dual-stage convolutional pyramid for capturing multi-scale functional connectivity patterns, augmented with a gated adaptive module enabling dynamic weight fusion of heterogeneous features. Tenfold cross-validation trials reveal our model yields 85.8% classification accuracy and 0.85 ROC-AUC in ASD recognition, demonstrating significant advancements over existing approaches.