Learning domain transfer for unsupervised magnetic resonance imaging restoration and edge enhancement

Abstract

AbstractRecently, many effective medical image restoration methods based on deep learning have been proposed. Most of the methods are used to solve single image processing tasks, such as image noise removal, image deblurring, and image super‐resolution. However, real medical images often suffer from multiple degradation factors, such as signal interference in the process of shooting or the relative movement of patients during the process of image acquisition. The image restoration methods only considering a single image degradation factor often fail to yield satisfactory results for the restoration purpose of practical medical images. It is difficult to obtain paired medical images to incorporate real image processing tasks into the framework of supervised learning. For the restoration of real medical images with multiple degradation factors, an unsupervised learning method based on domain transfer to resolve the MRI restoration problem is proposed. Specifically, learning invariant representations from a degraded image and a transferred image via adversarial domain adaption and then extra self‐supervised modules to further improve the robustness of representations are used. In addition, the Laplacian operator is embedded in the encoder to extract additional edges, which can form additional constraints on edge recovery. Experimental results show that the proposed method is effective in both image restoration and edge preservation.