We propose NeuroCLR, a self-supervised contrastive learning framework for learning robust and generalizable neural representations directly from raw resting-state functional magnetic resonance imaging (rs-fMRI) data. Our approach leverages contrastive objectives, anatomically consistent sampling, and augmented views of unlabeled fMRI time series to extract invariant representations that are consistent across subjects, imaging sites, and diagnostic categories. Unlike supervised and disorder-specific SSL approaches, NeuroCLR is pre-trained in a disorder-agnostic manner, enabling effective transfer to downstream classification tasks with limited labeled data. We pre-train NeuroCLR on large-scale multisite neuroimaging data comprising more than 3,600 participants from 44 imaging centers and over 720,000 region-specific fMRI time series. The resulting pre-trained model is fine-tuned for multiple disorder-specific classification tasks and consistently outperforms both supervised deep learning models and SSL methods trained on single disorders. Extensive experiments demonstrate robust generalizability across sites, highlighting NeuroCLR’s ability to learn biologically meaningful and transferable representations from unlabeled fMRI data. These findings establish NeuroCLR as a scalable and reproducible self-supervised framework for multisite neuroimaging analysis and cross-disorder clinical modeling.
Status
The method development for this work is complete