Foundation model‑enhanced deformable registration for 4D cardiac MRI

We propose a progressive three-stage self-supervised framework for 3D cardiac MRI registration across temporal sequences. The framework introduces a hierarchical learning strategy that advances from global anatomical understanding to fine-grained temporal correspondence, enabling the model to capture both structural features of the heart and dynamic motion patterns characteristic of the cardiac cycle. We begin with masked image modeling (MIM) pretraining on 4,647 3D cardiac volumes to establish robust anatomical representations, followed by registration-aware pretraining that incorporates synthetic deformation prediction and local alignment tasks designed to mimic cardiac motion. Finally, the pretrained encoder is transferred to state-of-the-art registration architectures, including VoxelMorph and TransMorph, for unsupervised fine-tuning on paired cardiac phases. Our evaluation on the ACDC dataset demonstrates substantial improvements over baseline methods, with VoxelMorph achieving a Dice score of 84.3% (compared to 74.6% baseline) and TransMorph reaching 76.7% (compared to 72.9% baseline). The proposed framework shows particular strength in registering the right ventricle, traditionally the most challenging cardiac structure. Our findings suggest that foundation model principles can effectively enhance cardiac image registration, paving the way for improved clinical applications in cardiac function assessment, motion tracking, and treatment planning.