MMI-GAN: multi medical imaging translation using generative adversarial network

Abstract

Medical image translation is considered a new frontier in the field of medical image analysis, with great potential for application. However, existing approaches have limited scalability and robustness in handling more than two image domains, since different models must be created independently for each pair of domains. To address these limitations, we developed MMI-GAN, a new approach for translation between multiple image domains, capable of translating intermodal (CT and RM) and intramodal (PD, T1 and T2) images using only a single generator and a discriminator, trained with image data from all domains. We propose a GAN architecture that can be easily extended to other translation tasks for the benefit of the medical imaging community. MMI-GAN is based on recent advances in the area of GANs (Generative Adversarial Network), using an adversary structure with a new combination of non-adversarial losses, which allows the simultaneous training of several data sets with different domains in the same network, as well as the innovative capacity to translate with flexibility between and inter/intra modalities. The images translated by MMI-GAN managed to obtain MAE of 5.792, PSNR of 27.398, MI of 1.430 and SSIM of 0.900. Its results were shown, often statically comparable or superior to Pix2pix and in almost all translations it was superior to Cyclegan.