Multi-site diagnostic classification of Autism spectrum disorder using adversarial deep learning on resting-state fMRI

•The sliding window strategy was adopted to preserve both the spatial and temporal information of rs-fMRI.•A two-stage adversarial learning model was proposed to fully extract spatial–temporal features and address the heterogeneity problem among multi-site study.•The proposed method achieved satisfactory classification performance (0.80 mean accuracy, 0.81 mean sensitivity and 0.80 mean specificity) in the ten-fold cross validation.•The proposed method presents a new way worthy of exploring for ASD diagnosis.