Improved CyeleGAN for MR to CT synthesis
Guogang Cao, Liu Shunkun, Hongdong Mao, Shu Zhang · 2021
Radiotherapy treatment planning requires CT images to accurately calculate the dose distribution, but sometimes only MR images can be obtained, therefore it is necessary to generate CT images from its corresponding MR images. In order to synthesize quality pseudo CT images, an improved cycle generative adversarial network (CycleGAN) with residual network (ResNet) and U-Net was proposed. The improved CycleGAN was trained by unpaired data to get better result, and approved by the cycle inconsistency of CycleGAN, the slow-down gradient disappearance of ResNet and the multi-scale fusion of U-Nets. Avoiding the disappearance of input information and the vanishing of gradient information, the improved network synthesized more quality CT images. Compared with the original CycleGAN method, the MAE of the proposed method was reduced by 5.12%, the SSIM increased by 0.7% and the PSNR increased by 1.29%, which was trained and tested on a dataset of 18 patients. Additional, compared with the DCNN method, the atlas-based method, and the pix2pix method, the relative error of MAE was reduced by 1.73%, 2.215% and 0.228% respectively. The proposed method synthesizes more vivid CT images owing to the advantages of deep learning model, which better meets the requirements of clinical analysis.