Research on low-dose CT image enhancement algorithm based on CycleGAN improvement

Low-dose CT imaging has a wide range of applications in the field of medical imaging. However, problems such as noise and low contrast in low-dose CT images pose challenges to the diagnosis and analysis of the images. Therefore, researchers have been working on developing and improving low-dose CT image enhancement methods to improve image quality and feasibility of clinical applications. Aiming at the problems of insufficient image feature extraction and large number of network parameters in CT images by CycleGAN, this thesis proposes an improved low-dose CT image enhancement network based on CycleGAN. A Shallow Feature pre-Extraction Module, which can extract rich semantic information, is designed to enhance the CycleGAN generator’s ability to extract image features in CT images; and some ordinary convolutions in the generator are replaced by Depthwise Separable Convolution with a smaller number of parameters to reduce the network parameters. The experimental results show that the improved CycleGAN exhibits better performance in terms of structural detail preservation, noise and artifacts suppression compared to both the traditional algorithm and other unsupervised algorithms.