Abstract
Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder, and the existing clinical diagnosis mainly relies on subjective behavioral assessment and lacks objective biomarkers. This paper proposes a hierarchical deep learning architecture, VaeTF, incorporating community-aware mechanisms based on resting-state functional magnetic resonance imaging (rs-fMRI) data. VaeTF introduces a priori knowledge of the functional community, extracts localized features through a variational auto-encoder (VAE), captures global dependencies across brain regions using the Transformer module, and incorporates an improved pooling mechanism to enhance the expressive power and model generalization performance. Experimental results on the ABIDE database show that VaeTF achieves 71.4% accuracy in ASD and typically performs well in group classification tasks. Further feature weighting analysis reveals that VaeTF is capable of identifying local functional abnormalities and cross-network functional synergistic dysfunctions closely related to ASD, thereby uncovering the underlying neurobiological mechanisms. VaeTF not only improves the classification performance of ASD but also provides a new method and theoretical support for objective assessment and early diagnosis based on fMRI.