A Self-Supervised Diffusion Model With Edge Prior for Unpaired LDCT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) minimizes radiation exposure but introduces image noise and artifacts.
• Deep learning methods for LDCT denoising face challenges like over-smoothing and training instability.
• Diffusion models offer improvements but require extensive sampling and paired training data, limiting practical use.

Purpose of the Study:

• To develop a self-supervised diffusion model for effective LDCT denoising.
• To address the limitations of existing diffusion models, including inference time and data requirements.
• To enhance denoised image clarity and reduce computational complexity in LDCT imaging.

Main Methods:

• Introduced a self-supervised diffusion model incorporating an edge prior for unpaired LDCT denoising.
• Utilized denoising within a lower-dimensional space to reduce computational load.
• Employed a noise-conditioned encoding strategy for self-supervised training on unpaired CT data.
• Integrated compressed LDCT encoding as intermediate sampling results to accelerate inference.

Main Results:

• Achieved competitive performance against state-of-the-art methods in PSNR, SSIM, and LPIPS.
• Demonstrated practical applicability with significantly reduced inference time, enabling real-time processing.
• Successfully denoised LDCT images using unpaired data, overcoming a key limitation of previous diffusion models.

Conclusions:

• The proposed self-supervised diffusion model offers an efficient and effective solution for LDCT image denoising.
• The method enhances image quality while maintaining practical inference speeds, making it suitable for clinical applications.
• This approach advances the field of medical image processing by enabling robust denoising with reduced data and computational demands.