Self-Supervised Temporal-Connectivity Contrastive Learning for fMRI Analysis on Autism Spectrum Disorder Classification

Diagnosing Autism Spectrum Disorder (ASD) using neuroimaging data is challenging due to the reliance on large labeled datasets, which are scarce in clinical settings. Existing deep learning methods typically focus on either time series or functional connectivity of fMRI data, missing the advantages of combining both. Here, we propose a novel self-supervised framework named TCCL-AE (Temporal-Connectivity Contrastive Learning with AutoEncoder) for ASD diagnosis based on resting-state fMRI. Our framework introduces a contrastive learning model that integrates both time series and functional connectivity through two parallel branches, capturing latent dynamic and connective brain activity patterns. We further enhance the representational capacity of the temporal-connectivity encoder by incorporating a complementary learning strategy with an autoencoder. Finally, we introduce a fine-tuning method for the downstream ASD classification task, which includes an attention mechanism to improve both the accuracy and interpretability of the model. Extensive experiments demonstrate that our TCCL-AE model outperforms state-of-the-art methods in ASD classification. Moreover, the ASD-related biomarkers identified by our model are consistent with findings from neuroscience research.