Self-supervised Learning for MRI Reconstruction with a Parallel Network Training Framework

Image reconstruction from undersampled k-space data plays an important role in accelerating the acquisition of MR data, and a lot deep learning-based methods have been exploited recently. Despite achieved inspiring results, optimization these commonly relies on fully-sampled reference which are time-consuming difficult to collect. To address this issue, we propose novel self-supervised learning method. Specifically, during model optimization, two subsets constructed by randomly selecting part then fed into parallel networks perform information recovery. Two losses defined all scanned points enhance network’s capability recovering frequency information. Meanwhile, constrain learned unscanned network, difference loss is designed enforce consistency between networks. In way, can be properly trained with only data. During evaluation, treated as inputs either expected reconstruct high-quality results. The proposed method flexible employed any existing effectiveness evaluated open brain MRI dataset. Experimental results demonstrate that achieve competitive performance compared corresponding supervised at high acceleration rates (4 8). code publicly available https://github.com/chenhu96/Self-Supervised-MRI-Reconstruction.