Fig. 2

Workflow of this study. This study was conducted in five steps. Firstly, pairs of ground truth images and noise images were collected for a total of 1,104 patients (1,341 pairs of MR images). Subsequently, the ground truth and noise images were subjected to preprocessing, including N4 bias field correction, normalization and image registration. Thirdly, the de-artifact model was trained based on the pre-processed images of 90% of the patients. The network included in this study comprised DDPM (Denoising Diffusion Probabilistic Models), EDSR (Enhanced Deep Residual Networks for Single Image Super-Resolution) and U-Net. Fourthly, the performance of models were evaluated in the remaining 10% of patients and independent testing cohort by comparing the difference between the generated image and the pre-processed ground truth images. Furthermore, subgroup analyses were conducted to assess the performance of images with varying axes.