Self-supervised learning for MRI reconstruction through mapping resampled k-space data to resampled k-space data

In recent years, significant advancements have been achieved in applying deep learning (DL) to magnetic resonance imaging (MRI) reconstruction, which traditionally relies on fully sampled data. However, real-world clinical scenarios often demonstrate that the fully sampled data can be challenging or impossible to obtain due to physiological constraints, such as organ motion, and physical constraints, such as signal decay. In this paper, we introduce a self-supervised DL approach, termed randomly self-supervised learning via data undersampling (abbreviated as RSSDU), which is proficient in efficiently and accurately reconstructing images from undersampled MRI data without requiring fully sampled datasets as references. The proposed method involves resampling the acquired k-space data twice to generate two subsets using the same undersampling pattern as the original acquisitions, albeit with different acceleration factors. Subsequently, a network is trained to learn to map from one of the sets to the other in a supervised manner. Extensive experiments demonstrate that the RSSDU method outperforms several well-known self-supervised methods, including SSDU and K-band, regarding peak signal-to-noise ratio and structural similarity index measurement.