Adversarial-Contrastive Domain Generalization for Cross-Site Autism Spectrum Disorder Classification from Resting-State fMRI

Originally accepted as a full paper at the MWM Workshop, MICCAI 2026 (Strasbourg, France). Cross-site generalization remains the primary barrier to deploying resting-state fMRI (rs-fMRI)-based Autism Spectrum Disorder (ASD) classifiers in clinical settings. Graph transformer models trained on source acquisition sites achieve intra-site AUC above 0.82 on ABIDE-I but drop to 0.73 on fully unseen target sites, a gap driven by scanner-induced distribution shifts in functional connectivity matrices. We propose DANN-ALTER-Con, combining a Domain Adversarial Neural Network, the Long-range Brain Graph Transformer (ALTER) backbone, and supervised contrastive learning, as a dual joint-learning framework for cross-site ASD classification. Evaluated on ABIDE-I under a strict site-disjoint protocol with 7 source sites for training and 11 fully unseen sites for testing, DANN-ALTER-Con achieves AUC 0.75 ± 0.08, outperforming the ALTER source-only baseline by 1.80% and reducing cross-site AUC standard deviation from 0.09 to 0.08. Code: https://github.com/fezz-bazay/DANN-ALTER-Con.