Motivation: Long scan times can limit the spatial and temporal resolution of cardiac cine, especially at low field. Goal(s): To propose a novel self-supervised motion estimation and motion corrected deep learning reconstruction to single heartbeat TR-resolved cardiac cine at 1.5T and 0.55T. Approach: We propose MC-DIP, a novel self-supervised approach that combines non-rigid motion estimation using a neural field physics-informed network with deep image prior motion-corrected reconstruction from undersampled radial data to enable TR-resolved dynamic cardiac imaging. Results: MC-DIP obtains high quality single heartbeat TR-resolved cines at both 1.5T and 0.55T, surpassing state of the art methods. Impact: The proposed approach enables TR-resolved single heartbeat cardiac cine at both 1.5T and 0.55T. This acceleration can be leveraged to reduce scan time, for example allowing more slides in the same breath-hold.
Catalán et al. (Tue,) studied this question.