Physic-informed diffusion models for medical imaging

This workshop will introduce the UCLA research community to the application of diffusion probabilistic models for medical imaging problems, with a focus on MRI. Participants will learn both the fundamentals of diffusion models (Gaussian vs. cold diffusion, measurement-conditioned models) and how they can be adapted to physics-constrained scenarios, such as k-space undersampling in MRI. Demonstrations will be given using publicly available datasets (e.g., CMRxRecon, OCMR) and open-source tools (PyTorch, MONAI framework) with supporting notebooks.

Target Audience:  Graduate students, postdocs, and faculty in computational sciences, biomedical physics, computer science, and engineering. Imaging scientists and clinicians interested in machine learning for medical image reconstruction. Any researchers in other fields (astronomy, microscopy, geoscience) where inverse problems and undersampled acquisitions are common.

Learning Outcomes:

• Understand the fundamentals of diffusion probabilistic models and the difference between classical Gaussian noise training and physics-informed degradations (cold diffusion).
• Understand the basics of MRI reconstruction and how it is treated as an inverse problem.
• Gain insight into how k-space undersampling can be formulated as a “forward process” for cold diffusion.
• Learn about tools for implementing custom forward processes in PyTorch.
• Explore how measurement conditioning integrates physical constraints with learned priors.
• Discuss broader applications of physics-informed diffusion models across other scientific imaging domains.