Accelerating Diffusion: Task-Optimized latent diffusion models for rapid CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Computed tomography (CT) is vital for diagnostics but involves radiation risks.
• Low-dose CT (LDCT) reduces radiation but introduces noise, impacting diagnostic accuracy.
• Deep learning, including CNNs, GANs, and DDPMs, has been used for LDCT denoising, but faces challenges like detail preservation and computational cost.

Purpose of the Study:

• To develop a novel and efficient framework for low-dose CT denoising.
• To address the limitations of existing deep learning methods, particularly DDPMs, in terms of computational cost and sampling speed.
• To enhance diagnostic accuracy in LDCT by effectively removing noise and artifacts.

Main Methods:

• Integration of the Latent Diffusion Model (LDM) with the Cold Diffusion Process for LDCT denoising.
• Utilizing LDM to perform the diffusion process in a low-dimensional latent space, reducing computational demands.
• Employing a CT denoising task-specific degradation approach within the Cold Diffusion Process, replacing traditional Gaussian noise for improved efficiency.

Main Results:

• The proposed LDM-Cold Diffusion framework demonstrated superior performance compared to DDPM in key metrics (PSNR, SSIM, RMSE).
• Achieved up to 2x faster training times and 14x faster sampling speeds.
• Successfully preserved fine image details while effectively reducing noise and artifacts.

Conclusions:

• The proposed framework offers a practical and effective solution for LDCT denoising, overcoming the computational and efficiency limitations of previous methods.
• This advancement holds significant potential for improving the clinical applicability of LDCT by enhancing image quality and reducing scan times.
• The integration of LDM and Cold Diffusion presents a promising direction for future research in medical image reconstruction and analysis.