Semi-supervised dual generative adversarial network for low-dose CT artifact suppression

Supervised learning-based deep learning methods have enabled performance improvements in denoising low-dose computed tomography (LDCT). However, there are practical difficulties in data acquisition, and the use of paired CT training samples is not sufficient. Furthermore, it is difficult to directly apply supervised learning approaches to unpaired CT image data. To address these issues, we designed a semi-supervised dual generative adversarial network, which not only leverages paired data for the supervised training of denoising networks but also uses a large amount of unpaired data for training to improve the networks’ performance. Through alternating iterations of supervised and unsupervised training, and utilizing the combined training of paired and unpaired data, we achieved effective suppression of artifacts and noise in LDCT. In the design of the network structure, the task of LDCT image generation and noise reduction is performed through the interactive confrontation training of two generative adversarial networks. In addition, a multi-description loss function corresponding to the network training mode is designed to improve the stability of the semi-supervised network training and the algorithm performance. The experimental results show that, compared with existing supervised and unsupervised noise reduction methods, the proposed network approach can achieve better noise reduction while effectively preserving the original structure of the CT images.