Self-supervised motion-compensated reconstruction for cardiac Cine MRI

Motivation: To enable single breath-hold cardiac Cine MRI with efficient deep learning reconstruction that utilizes only undersampled training data. Current existing self-supervised approaches mainly focus on static images. Goal(s): We aim to develop a self-supervised method for dynamic MRI reconstruction without needing fully-sampled data while striving to achieve high-quality reconstruction under high acceleration rates. Approach: We propose SSL-MoCo, which combines self-supervised learning with motion-compensated reconstruction for a joint image registration and reconstruction network. Results: SSL-MoCo achieves high-quality reconstruction performance and improved artifact removal compared to other self-supervised methods in highly accelerated cardiac Cine MRI, even outperforming supervised learning while not requiring fully-sampled data. Impact: We propose a fully self-supervised framework that enables high-quality reconstruction of cardiac Cine MRI acquired in a single breath-hold. The adaptability of this framework opens new research avenues for leveraging undersampled data and extends to other dynamic MRI modalities.