Self-Supervised Representation Learning for Dynamic Functional Connectivity With Subjectwise and Temporal Contrasts

Dynamic functional connectivity (dFC) captures temporal dynamics in functional magnetic resonance imaging (fMRI) to better characterize brain activity, supporting multiple downstream analyses. However, effective modeling strategies for extracting representative signals from dFC remain underexplored. We proposed a novel Subject-wise and Temporal Contrastive Transformer (STCT) framework that advances dynamic connectivity modeling through a dual-constraint contrastive learning strategy. Our STCT framework uniquely integrates two objectives: 1) subject-wise contrast to preserve inter-individual specificity, and 2) temporal contrast to capture dynamic dependencies, which is a critical dimension overlooked in prior methods. Comprehensive evaluations demonstrated superior performance of STCT over both supervised and self-supervised approaches based only on subject-wise contrast in various predictive domains spanning demographics, cognition, and mental disorder diagnosis. Interpretability analysis further revealed lateralized connectivity patterns in Autism Spectrum Disorder (ASD) classification, suggesting the potential of STCT to highlight clinically relevant connectivity patterns.