Diffusion models in medical imaging: A comprehensive survey.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Denoising diffusion models are generative models that excel in data generation quality and coverage.
• These models are gaining traction in medical imaging due to advances in computer vision.
• Despite computational costs, their potential in medical applications is significant.

Purpose of the Study:

• To provide a comprehensive overview of diffusion models in medical imaging.
• To systematically categorize diffusion models based on applications, modalities, organs, and algorithms.
• To guide researchers through the growing body of work in this domain.

Main Methods:

• Introduction to the theoretical foundations of diffusion models and their three main frameworks: diffusion probabilistic models, noise-conditioned score networks, and stochastic differential equations.
• Systematic taxonomy and multi-perspective categorization of diffusion models in medicine.
• Review of extensive applications including image-to-image translation, reconstruction, registration, classification, segmentation, denoising, generation, and anomaly detection.

Main Results:

• Diffusion models demonstrate broad applicability across diverse medical imaging tasks.
• The survey categorizes existing research, highlighting key algorithms and their use in various medical contexts.
• Practical use cases and limitations are discussed, alongside future research directions.

Conclusions:

• Diffusion models offer a promising avenue for advancing medical imaging analysis and generation.
• Further research is needed to address limitations and fully realize their potential in clinical settings.
• An open-source repository of reviewed studies is provided to facilitate ongoing research and development.