Latent Diffusion Model With Estimation Posterior Sampling: A Unified Framework for General Medical Image Restoration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Accelerated or low-dose clinical imaging protocols often result in degraded image quality, impacting diagnostic accuracy.
• Variations in degradation types and severities across different imaging modalities hinder the development of universal restoration solutions.

Purpose of the Study:

• To develop a unified framework for medical image restoration using self-supervised Latent Diffusion Models (LDMs).
• To enhance image fidelity and anatomical detail retention through an Estimation Posterior Sampling (EPS) strategy.

Main Methods:

• Formulating medical image restoration as posterior sampling from LDMs pretrained on multi-modal high-quality images.
• Implementing an EPS strategy with estimated diffusion initialization and gradient-balanced optimization.
• Enabling Plug-and-Play (PnP) deployment for diverse degradations without retraining.

Main Results:

• Demonstrated superior quantitative and qualitative performance compared to supervised baselines and state-of-the-art posterior sampling methods.
• Achieved significant PSNR improvements: up to +2.9 dB for MRI, +1.1 dB for CT, and +0.9 dB for PET in PnP mode.
• Validated performance across deterministic (under-sampled MRI, sparse-view CT) and blind (low-dose PET) degradations.

Conclusions:

• The proposed unified framework offers a robust and broadly applicable solution for medical image restoration.
• The EPS strategy effectively enhances both data fidelity and anatomical detail preservation.
• The PnP capability allows for versatile deployment across various clinical imaging scenarios without task-specific retraining.