Meeting Banner
Abstract #4007

TGD-BO: Task-specific Guidance Design with Bayesian Optimization using unconditional diffusion models for image restoration problems

Naoto Fujita1 and Yasuhiko Terada1
1Graduate School of Science and Technology, University of Tsukuba, Tsukuba, Japan

Synopsis

Keywords: AI Diffusion Models, Machine Learning/Artificial Intelligence, Foundamental model, Bayesian Optimization

Motivation: Deep learning models are excellent in image restoration tasks, but often suffer from domain shift and require task-specific learning, which limits their practical use.

Goal(s): To develop Task-specific Guidance Design with Bayesian Optimization (TGD-BO) a method that optimizes diffusion guidance parameters for various image restoration tasks without task-specific training.

Approach: Conditional generation using data-consistency and gradient guidance, with Bayesian Optimization to efficiently tune guidance parameters through B-spline curve optimization.

Results: TGD-BO achieved comparable or superior performance to supervised models across various tasks (compressed sensing, super-resolution, denoising), particularly excelling in joint tasks like simultaneous noise removal and reconstruction.

Impact: We proposed a guidance design method adaptable to any image restoration problems and verified its effectiveness. This versatile framework enables diffusion model to handle multiple MRI restoration tasks without task-specific training, potentially serving as a foundation model.

How to access this content:

For one year after publication, abstracts and videos are only open to registrants of this annual meeting. Registrants should use their existing login information. Non-registrant access can be purchased via the ISMRM E-Library.

After one year, current ISMRM & ISMRT members get free access to both the abstracts and videos. Non-members and non-registrants must purchase access via the ISMRM E-Library.

After two years, the meeting proceedings (abstracts) are opened to the public and require no login information. Videos remain behind password for access by members, registrants and E-Library customers.

Click here for more information on becoming a member.

Keywords