CT-to-MRI mapping using forward diffusion-based conditional GAN

Abstract Cross-modality medical image translation is a vital research topic in the domain of medical imaging, which allows medical experts to obtain multi-modality medical images and thus to make more wisdom decisions. On the other hand, the emerging denoising diffusion model has got more and more attention since it can synthesize higher-quality and more diverse samples in comparison to other generative models such as VAE and GAN. However, its main drawback is that it has a very slow inference. To take advantage of the denoising diffusion model to synthesize much higher-quality cross-modality medical images while having a fast inference, in the paper we propose a forward diffusion-based conditional GAN model, and apply it to translate volumetric CT-to-MRI. This model consists of three main components: the forward diffusion is used to iteratively adds noise to the clear data to get noisy data. The generator is used to directly generate clear data from pure random noise. The discriminator is used to distinguish the fake clear data synthesized by the generator from the real ones. To show the effectiveness and superiority of our method, we apply two clinical datasets to it, and qualitatively and quantitatively conduct comparative experiment and ablation study. The experimental results show that compared to the state-of-the-art translation methods, our method can perform a more accurate volumetric CT-to-MRI synthesis while having a fast inference. The experimental results also show that our method compromises contrast and details, but with better intensities and brightness in comparison with the diffusion model.