Retrospective k-space Subsampling schemes For Deep MRI Reconstruction
In this paper, we investigate and compare various retrospective k-space subsampling patterns and their effect on the quality of DL-based reconstructions. We retrospectively generated subsampling masks on the Cartesian grid for different types of subsampling patterns, such as rectilinear, non-rectilinear, and non-Cartesian patterns. We performed experiments using our state-of-the-art DL-based accelerated MRI reconstruction method, the Recurrent Variational Network, under two setups: scheme-specific and multi-scheme. Our results demonstrated that non-rectilinear and non-Cartesian patterns produced higher-quality reconstructions with fewer artifacts, particularly for higher acceleration factors. Additionally, our multi-scheme setup showed improvements in reconstruction performance for rectilinear subsampled measurements. Overall, our findings suggest that non-rectilinear and non-Cartesian subsampling patterns may be more suitable for DL-based reconstructions.