Diffusion Posterior Sampling for Tomographic Reconstruction with Mixed Resolution Priors.

Area of Science:

• Medical imaging
• Deep learning
• Computational imaging

Background:

• Deep learning excels at image reconstruction, but training data limitations hinder fidelity.
• Diffusion posterior sampling (DPS) integrates generative priors with measurement models.
• Current DPS methods are constrained by the quality and scope of training data.

Purpose of the Study:

• To develop a novel DPS framework for enhanced image reconstruction.
• To improve regional spatial resolution while preserving global image consistency.
• To address limitations of standard clinical data in training deep learning priors.

Main Methods:

• Integrated global and regional diffusion models for mixed prior reconstruction.
• Employed frequency-domain methods to combine low-frequency global and high-frequency regional priors.
• Utilized a shifted patch division mechanism to mitigate boundary discontinuities.
• Implemented a regional sampling scheme with a binary mask for targeted prior application.

Main Results:

• The proposed framework achieved superior reconstruction quality.
• Preserved fine-grained details in regions of interest.
• Maintained global structural coherence without sacrificing resolution.
• Demonstrated high-fidelity and efficient posterior sampling in inverse imaging problems.

Conclusions:

• Combining multi-scale diffusion priors offers a powerful approach for high-fidelity image reconstruction.
• The novel DPS framework effectively enhances regional details and global stability.
• This method has significant potential for advanced medical imaging applications.