Cross-attention-guided subject-adaptive graph learning for multimodal autism classification: integrating structural and functional MRI data
TL;DR
This work proposes a cross-attention-guided subject-adaptive graph network (CAS-GNN) model that integrates structural MRI and resting-state functional connectivity data, effectively fusing complementary multimodal information and provides a promising tool for interpretable and robust ASD diagnosis, accelerating biomarker discovery and development.
Abstract
Autism spectrum disorder (ASD) is a complex neurodevelopmental condition marked by structural atypicality and abnormal functional connectivity. It remains challenging to accurately delineate an ASD-associated neural marker due to individual heterogeneity and multi-site data variability. To address these issues, we propose a cross-attention-guided subject-adaptive graph network (CAS-GNN) model that integrates structural MRI and resting-state functional connectivity data, effectively fusing complementary multimodal information. By modeling individualized brain network topologies and incorporating a site-invariant learning strategy, our approach enhances discriminability and cross-site generalization. On the ABIDE-I dataset, CAS-GNN significantly outperformed machine learning baselines and achieved an accuracy of 79.25% ± 4.71% on independent test data and an average accuracy of 78.75% ± 1.56% on five-fold cross-validation. Exploratory analyses identified key ASD-related brain regions and connections, revealing a notable right-hemisphere dominance consistent with atypical asymmetry in ASD. Our framework offers valuable neurobiological insights and provides a promising tool for interpretable and robust ASD diagnosis, accelerating biomarker discovery and development.