Self-supervised representational learning for automated risk assessment in longitudinal imaging

Motivation: Techniques that allow automated evaluation of the evolution of disease risk over time can be of great value for active surveillance and other imaging-based monitoring. Goal(s): We introduce a novel self-supervised framework to learn representations that can identify increases in risk over time. Approach: We propose a contrastive learning model to first learn subject-specific representations from low-slice-resolution images followed by learning a risk axis in the representational space to provide information on global changes in risk over time. Results: The developed framework was used to assess risk of new metastases in a cohort of subjects from the NYU-Mets longitudinal imaging dataset. Impact: A key question when moving to lower field strengths in MRI is if we can get comparable information from lower-quality images as we can from the current standard of high-quality, high-resolution images. Self-supervised contrastive learning approaches can hold the key.