Diffusion model for medical image denoising, reconstruction and translation

Diffusion models, as a class of generative models, have demonstrated significant performance in image generation since their inception. The fundamental principle behind diffusion models is the definition of a forward process and a reverse process. The input data is progressively perturbed by adding random noise during the forward process, and the expected noise distribution is learned. In the reverse process, noise is gradually reduced from a Gaussian distribution to generate the image. Recently, diffusion models have been widely adopted in various image processing tasks, including text-to-image synthesis, denoising, segmentation, and object detection. In medical image analysis, diffusion models have shown considerable potential for improving diagnostic accuracy and image quality. This article provides a comprehensive overview of diffusion models, particularly their applications in medical image denoising, reconstruction, and translation. Specifically, we categorize diffusion models into two types: denoising diffusion probabilistic models and score-based models and introduce the solid theoretical foundations and fundamental concepts underlying these models. Additionally, we introduce publicly available datasets and evaluation metrics relevant to these methods. Most importantly, we provide detailed introductions to several representative articles, summarize current applications of diffusion models in these domains, and discuss potential future directions for development and challenges.