Neuroimaging analysis in brain disorders faces a persistent challenge: brain signals are complex and high-dimensional, while high-quality labeled datasets remain limited. This review article systematically examines how self-supervised learning can help address that gap by learning meaningful representations directly from unlabeled neuroimaging data. It covers major methodological families, including contrastive, generative, and hybrid generative-contrastive approaches, and discusses their applications in functional MRI, EEG, and multimodal brain network analysis.
The review argues that self-supervised learning offers more than annotation efficiency. It may enable more transferable and clinically useful representations for disease screening, diagnosis, and prognosis across heterogeneous datasets and disorders. At the same time, interpretability, data heterogeneity, missing modalities, and clinical validation remain major barriers. Future work will likely focus on stronger multimodal fusion, better cross-site generalization, and more clinically adaptable model design.
