SSDiffusion: State-Space Diffusion Model for Medical Image Synthesis

Medical image synthesis enables imputation of missing slices from acquired modalities, thereby reducing the need for repetitive and prolonged procedures in diagnostics. However, this is a non-trivial task due to the ill-posed and nonlinear characteristics of the problem. Generative adversarial networks (GANs) have shown promise in addressing these challenges. However, GANs can suffer from instability and mode collapse. Denoising diffusion models (DDMs) are recently demonstrated to be superior to GANs in terms of training stability and high fidelity. Yet, existing diffusion-based techniques rely on U-Net backbones, which are suboptimal for capturing contextual relationships between different tissue parts. Although transformers are adept at extracting the global context, their quadratic complexity makes them impractical for processing individual pixels rather than image patches. State-space models (SSMs) provide an efficient alternative with lower complexity, enabling the handling of extremely long sequences while effectively extracting contextual relationships. In this paper, we introduce SSDiffusion, a novel state-space diffusion model designed for multimodal medical image translation. By leveraging the strengths of state-space representations, our approach effectively captures contextual dependencies improving the translation performance. We evaluate SSDiffusion on a variety of medical image translation tasks. Experiments indicate that SSDiffusion outperforms state-of-the-art GAN and diffusion methods.