Learning the degradation distribution for medical image superresolution via sparse swin transformer

Abstract

High-resolution (HR) medical images can provide rich details, which are important for discovering subtle lesions to make diagnoses. Convolutional neural networks (CNNs) are widely used in this field, but struggle to model long-range dependencies. Although transformer-based methods have improved in this respect, this method requires large quantities of data. Unfortunately, large quantities of low-resolution (LR) and HR medical image pairs may not always be available. In addition, most medical image superresolution (SR) methods are deterministic, while the degradation in real scenarios is stochastic. To address these problems, we introduce a probabilistic degradation model that combines natural and medical images for training. This design alleviates the problem of insufficient medical image pairs and learns the degradation process of the natural scene. In addition, we propose a new medical image SR model that consists of CNNs and the Swin Transformer structure to excavate both local and global semantic features. Moreover, to reduce computational stress, the spherical locality-sensitive hashing (SLSH) module is employed in the nonlocal attention (NLA) mechanism to form the ENLA module. This design enables the proposed Sparse Swin Transformer (SSFormer) model to generate HR medical images without extensive training images. Experiments on diverse datasets (natural images and medical images) demonstrate that the proposed method is robust and effective, qualitatively and quantitatively outperforming other medical image SR methods. Code is available at https://github.com/codehxj/SSFormer.