Diffusion Models for Medical Image Computing: A Survey

Abstract

Diffusion models are a type of generative deep learning model that can process medical images more efficiently than traditional generative models. They have been applied to several medical image computing tasks. This paper aims to help researchers understand the advancements of diffusion models in medical image computing. It begins by describing the fundamental principles, sampling methods, and architecture of diffusion models. Subsequently, it discusses the application of diffusion models in five medical image computing tasks: image generation, modality conversion, image segmentation, image denoising, and anomaly detection. Additionally, this paper conducts fine-tuning of a large model for image generation tasks and comparative experiments between diffusion models and traditional generative models across these five tasks. The evaluation of the fine-tuned large model shows its potential for clinical applications. Comparative experiments demonstrate that diffusion models have a distinct advantage in tasks related to image generation, modality conversion, and image denoising. However, they require further optimization in image segmentation and anomaly detection tasks to match the efficacy of traditional models. Our codes are publicly available at: https://github.com/hiahub/CodeForDiffusion.