Self-Supervised MR Image Reconstruction From Single Measurement

Recently, deep learning (DL)-based methods have gained popularity in accelerating magnetic resonance imaging (MRI). However, DL-MRI training demands a substantial amount of paired data, which is often challenging to obtain in practice. This paper aims to establish a self-supervised deep learning MRI reconstruction method that doesn't rely on any external training data. Inspired by Self2Self, we propose a single-image reconstruction approach that includes Bernoulli sampling applied to the input image, a drop strategy during training to eliminate artifacts on undersampled images, and the incorporation of the physical processes of MRI. Experimental results demonstrate that the method performs well.