CoreDiff: Contextual Error-Modulated Generalized Diffusion Model for Low-Dose CT Denoising and Generalization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) images are prone to noise and artifacts.
• Existing deep learning denoising models face challenges like over-smoothing and training instability.
• Conventional diffusion models have long inference times due to extensive sampling steps.

Purpose of the Study:

• To develop an efficient and effective denoising model for LDCT images.
• To address the limitations of existing diffusion models in terms of inference speed and generalization.
• To introduce a novel approach for improving the quality of LDCT scans.

Main Methods:

• A novel COntextual eRror-modulated gEneralized Diffusion model (CoreDiff) inspired by cold diffusion.
• Utilizes LDCT images as the starting point for noise displacement, employing a mean-preserving degradation operator to reduce sampling steps.
• Introduces a ContextuaL Error-modulAted Restoration Network (CLEAR-Net) to mitigate error accumulation and structural distortion.
• Implements a one-shot learning framework for rapid generalization to unseen dose levels.

Main Results:

• CoreDiff significantly reduces sampling steps compared to traditional diffusion models.
• The CLEAR-Net effectively constrains the sampling process and modulates features for improved restoration.
• The one-shot learning framework enables fast and robust generalization to new dose levels.
• Extensive experiments show CoreDiff outperforms competing methods in denoising and generalization performance.

Conclusions:

• CoreDiff offers a promising solution for LDCT denoising, balancing image quality and computational efficiency.
• The proposed model achieves clinically acceptable inference times, making it practical for clinical application.
• CoreDiff demonstrates superior denoising and generalization capabilities compared to existing methods.