Unpaired medical image colorization using generative adversarial network

Abstract We consider medical image transformation problems where a grayscale is transformed into color image. The colorized should have the same features as input because extra synthesized can increase possibility of diagnostic errors. In this paper, to secure images and improve quality images, well leverage unpaired training data, colorization network proposed based on cycle generative adversarial (CycleGAN) model, combining perceptual loss function total variation (TV) function. Visual comparisons experimental indicators from NRMSE, PSNR, SSIM metrics are used evaluate performance method. results show that GAN-based style conversion be applied images. As well, introduction TV produced result better than generated by only using CycleGAN model.