서강대학교

초록

In this study, we proposed a generative adversarial network (GAN) model, called CypixGAN, to synthesize brain FLAIR MR images from PET and/or CT images. CypixGAN architecture was designed by combining CycleGAN framework with L1 loss function of pix2pix. It was a multi-channel GAN model using various data types as an input, such as PET and CT images, PET-only images, or CT-only images. Three loss functions were mixed in training process. One was the adversarial loss calculated when the discriminator tried to distinguish synthesized FLAIR MR images from label FLAIR MR images. Another was the cycle loss calculated when L1 distance was measured between the synthesized PET and/or CT images through the generator and the label PET and/or CT images. The other was the L1 loss, which represented the L1 distance between synthesized FLAIR MR images and label FLAIR MR images. Ninety-eight images were used for training and 45 images were used for testing the trained network. The differences between label FLAIR MR images and synthesized FLAIR MR images generated by CypixGAN, pix2pix, and CycleGAN with three types of inputs were quantitatively analyzed by calculating PSNRs and GMSDs. The synthesized FLAIR MR images were successfully generated by CypixGAN proposed in this study and its quality was better than those by pix2pix and CycleGAN. The quality of synthesized FLAIR MR images generated using PET and CT images was superior to those of PET-only images or CT-only images regardless of the frameworks. Best PSNRs and GMSDs were achieved by CypixGAN with PET and CT images. The quantitative results of CypixGAN, pix2pix, and CycleGAN with PET and CT images were 20.23±1.97, 19.35±1.66, and 19.41±1.81 in PSNR, and 0.18±0.02, 0.21±0.02, and 0.20±0.02 in GMSD. In this study, we demonstrated that high quality FLAIR MR images from the PET and CT images could be synthesized by the CypixGAN. This proved that CypixGAN helps to improve the synthesis performance of the network by learning more feature distributions.