Diffusion model with Rician-Gaussian priors for robust MR image synthesis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic Resonance Imaging (MRI) is crucial for noninvasive diagnosis but relies heavily on expert interpretation.
• AI enhances MRI diagnosis, yet data scarcity due to privacy concerns limits training high-quality models.
• Existing data augmentation methods fail to accurately replicate Rician noise in MRI, reducing synthetic image fidelity.

Purpose of the Study:

• To develop a novel generative framework for synthesizing realistic MRI images.
• To address the limitations of conventional methods in reproducing Rician noise patterns.
• To improve the quality and utility of synthetic MRI data for AI-driven diagnostic applications.

Main Methods:

• Introduction of a Rician-Gaussian denoising diffusion model (RGDM).
• Integration of Rician fidelity and Gaussian stability for image synthesis.
• Compensation for Rician noise's noncentral bias to enhance training stability and reduce signal distortion.

Main Results:

• RGDM successfully synthesizes high-fidelity MR images with fine structural detail and natural signal variation.
• The model demonstrates superior performance in downstream tasks like denoising, anomaly map generation, and tumor segmentation.
• Synthetic images generated by RGDM led to improved diagnostic accuracy.

Conclusions:

• RGDM offers a robust solution for generating high-quality synthetic MRI data, overcoming limitations of existing methods.
• The framework enhances diagnostic accuracy in various MRI applications by providing realistic training data.
• This advancement holds significant potential for AI development in medical imaging, addressing data scarcity challenges.