Registration-Guided Cardiac Functional Assessment from Limited Annotations in a Single Breath-hold Cine

Motivation: CMR is the gold standard for evaluating cardiac function, but its lengthy acquisition and dependence on expert readers limit its efficiency. Goal(s): We aim to develop an automated accelerated cardiac function assessment method that requires minimal data acquisition and human intervention. Approach: We propose an efficient framework for cardiac function assessment, trained on sparsely labeled, accelerated images. Our multi-task method employs a synergistic loop of registration, motion-compensated reconstruction, and segmentation, enabling mutual refinement. Results: We demonstrate reliable ventricular function analysis from accelerated MRI data, acquired even within single breath-hold Cine, and achieved a 13%-25% improvement in Dice similarity over other deep learning-based methods. Impact: Our framework enables automated cardiac function assessment, even for highly accelerated single breath-hold scans. We improve CMR accessibility for studies with limited subjects and sparse manual annotations. Results indicate reliable motion estimation, ventricular function measures and myocardial strain analysis.