Figure 2.

CycleGAN architecture for bidirection synthesis of MR and CT images. The forward cycle generated synthetic CT from input MR by G while F translate the synthetic CT back to the MR image domain. D_CT discriminate whether the generated images is real or fake CT. The backward cycle generated synthetic MR from input CT by F while G translate the synthetic MR back to the CT image domain. D_MR discriminate whether the generated images is real or fake MR. Two cycle-consistency loss was introduced to capture the intuition that the synthetic image should be translated back to the original image modality.