A Generalized Cascaded Self-Supervised Registration Pipeline with Physics-Informed Learning for Enhanced Quantitative Cardiac MRI

Motivation: Multi-parametric mapping has become a viable tool for myocardial tissue characterization. However, misregistration between multi-parametric maps makes pixel-wise analysis challenging. Goal(s): To develop a registration pipeline for accurate registration of quantitative maps acquired at different scans. Approach: A cascaded self-supervised registration pipeline was developed, integrating a generalizable physics-informed module and a two-level contrast-agnostic groupwise registration module to optimize model efficiency and minimize bias from physical constraints. Results: The model demonstrates high registration quality for T1 mapping of post-Gd administration kinetics, enabling reliable registration of dynamic mapping with varying T1 contrast. Impact: The cascaded self-supervised pipeline with a physics-informed module offers a scalable framework to facilitate accurate and efficient image registration for image contrast modulation following multiple physical and physiological models.