Study of Low-Dose CT Image Enhancement Network Based on Improved Cycle GAN Method

Cycle-consistent Adversarial Network (Cycle GAN) is an image generation model based on a Generative Adversarial Network (GAN), which has been widely used. In this paper, an improved image enhancement network based on Cycle GAN is proposed for low-dose CT image enhancement. By introducing a new loss function and network structure, the network improves the effect of image enhancement. In order to reduce the damage of X-rays to the human body, Low-dose CT (LDCT) is widely used in clinical diagnosis. However, quantum noise will be introduced into the projection data while the radiation dose is reduced. The quality of reconstructed CT images is reduced. Aiming at the problem of image noise in low-dose CT, the traditional image noise reduction methods cannot meet the practical needs due to the complicated process of noise reduction or the defects of the algorithm itself. A network model based on the residuals module is established, three identical residuals blocks are added to the original network, and the residual function after adding residuals blocks is optimized. In the part of generating network, deconvolution is used instead of reshaping the enhanced image, and the improved network is more stable in performance. The validity of the network model is verified by the test data.