Escape the Corrupted Lines by a Learning Sampling Pattern for MRI Reconstruction

Magnetic Resonance Imaging (MRI) is a powerful medical imaging technique in clinical practice. Unfortunately, the long scan times of itself and the patient's uncontrollable motion result in image artifacts, and increase the operational costs. In this work, we propose an end-to-end joint learning framework of sampling and reconstruction, which can select the un-corrupted lines as many as possible to improve the robustness of the reconstruction module to motion by a learnable Gaussian mixture module. Experimental evaluations on the fastMRI knee dataset with simulated motion demonstrate that our method is efficient in achieving high-quality reconstruction and correcting artifacts.