Automated Tuning for Diffusion Inverse Problem Solvers without Generative Prior Retraining.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Diffusion/score-based models are powerful generative priors for inverse problems like accelerated MRI reconstruction.
• Model performance relies on data fidelity weights, especially with fast sampling and few denoising steps.
• Current methods use fixed weights or heuristics, limiting generalization across varied measurement conditions and schedules.

Purpose of the Study:

• To introduce Zero-shot Adaptive Diffusion Sampling (ZADS), a novel test-time optimization method.
• To adaptively tune fidelity weights for diffusion models in MRI reconstruction without retraining.
• To improve MRI reconstruction performance across arbitrary noise schedules and acquisition settings.

Main Methods:

• ZADS optimizes fidelity weights in a self-supervised manner using only undersampled measurements.
• The diffusion denoising process is treated as a fixed unrolled sampler.
• No retraining of the diffusion prior is required.

Main Results:

• ZADS consistently outperforms traditional compressed sensing and existing diffusion-based methods.
• High-fidelity reconstructions are achieved across varying noise schedules and acquisition settings.
• Demonstrated effectiveness on the fastMRI knee dataset.

Conclusions:

• ZADS offers a robust solution for adaptive fidelity weight tuning in diffusion-based MRI reconstruction.
• The method enhances generalization capabilities for accelerated MRI.
• ZADS represents a significant advancement in self-supervised learning for medical image reconstruction.