Motivation: Self-supervised neural network reconstruction improves multi-shot diffusion MRI (dMRI), yet suffers from prohibitively long computation times. Goal(s): To develop a zero-shot self-supervised learning method for fast multi-shot dMRI reconstruction. Approach: We propose a physics-guided neural network that operates in both k- and image-spaces to combine information from different EPI shots. We show that reconstruction quality can be improved with a novel sampling mask strategy, and that faster training is possible with a new training strategy. Finally, we extend our results to SMS acquisitions. Results: Our results show that the proposed method provides improved and fast reconstructions compared to 2-shot LORAKS and 2-shot ZS-SSL. Impact: The proposed physics-guided self-supervised learning method provides fast and high-quality reconstruction of multi-shot diffusion MRI volumes, while also eliminating the need for external training datasets.
Support the authors with ResearchCoin