Motivation: Real-time free-breathing cardiac MRI requires high acceleration rates and dedicated reconstruction techniques. Deep Learning (DL) methods are often infeasible due to the lack of training data. Goal(s): Develop an unsupervised reconstruction method for dynamic cardiac MRI that can model content variation as well as in-plane and through-plane motion. Approach: Our method is based on the Deep Image Prior (DIP) and combines a low-rank approach with the generation of motion fields. Results: M-DIP outperformed state-of-the-art DIP methods in a phantom and two in-vivo studies. It further achieved similar image quality as a supervised DL method in real-time cine, without requiring a large training dataset. Impact: Our method enables real-time free-breathing cine and free-breathing LGE imaging with high resolution and motion fidelity. It requires no training data and can be extended to other types of dynamic MRI acquisitions.
Vornehm et al. (Tue,) studied this question.