An unsupervised deep learning network model for artifact correction of cone-beam computed tomography images

Unsupervised deep learning network model cycle-consistent generative adversarial network (CycleGAN) is increasingly applied for artifact correction of cone-beam computed tomography (CBCT) images owing to the registration-free advantage of dataset. However, synthetic Planning CT images (sPCT) based on the model lose the anatomical details of the original CBCT images. Therefore, to improve the accuracy of adaptive radiation therapy (ART), it is necessary to maintain the anatomical structures between the sPCT and original CBCT images, while improving CBCT image quality. An improved CycleGAN model was designed based on an attention module and a structural consistency loss function. The improved CycleGAN model was trained using CBCT and Planning CT (PCT) images of 43 patients to generate sPCT images from CBCT images. Images of nine other patients were used to verify the effectiveness of the improved CycleGAN model. As compared to the original CycleGAN model, the sPCT images generated by the improved CycleGAN model increased by 2.87%, 9.64%, and 7.91%, respectively, in the image quality evaluation indicators PSNR, MAE, and RMSE, while increased by 2.43% and 32.03%, respectively, in the structural consistency evaluation indicators SSIM and MIND. The improved CycleGAN model generated high quality sPCT images and accurately preserved the anatomical details of the original CBCT images, thereby demonstrating great potential for clinical applications of ART.