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

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Dynamic functional connectivity (dFC) analysis of functional magnetic resonance imaging (fMRI) is crucial for understanding brain activity dynamics.
• Existing methods for modeling dFC signals are underexplored, limiting downstream applications.
• There is a need for advanced modeling strategies to effectively extract representative signals from dFC data.

Purpose of the Study:

• To introduce a novel Subject-wise and Temporal Contrastive Transformer (STCT) framework for advanced dFC modeling.
• To enhance the characterization of brain activity by integrating inter-individual specificity and temporal dynamics.
• To improve the performance of predictive models in various domains using dFC data.

Main Methods:

• Developed a novel STCT framework utilizing a dual-constraint contrastive learning strategy.
• Integrated subject-wise contrastive learning to preserve inter-individual specificity.
• Incorporated temporal contrastive learning to capture dynamic dependencies in fMRI data.

Main Results:

• STCT demonstrated superior performance compared to existing supervised and self-supervised methods.
• The framework excelled in predictive tasks including demographics, cognition, and mental disorder diagnosis.
• Interpretability analysis revealed lateralized connectivity patterns relevant to Autism Spectrum Disorder (ASD) classification.

Conclusions:

• The STCT framework offers a significant advancement in modeling dynamic functional connectivity from fMRI data.
• STCT effectively captures both inter-individual differences and temporal dynamics, outperforming previous approaches.
• The model shows potential for clinical applications, such as identifying biomarkers for neurological and psychiatric disorders like ASD.