[Paper Review] CoreDiff: Contextual Error-Modulated Generalized Diffusion Model for Low-Dose CT Denoising and Generalization

Qi Gao, Zilong Li+3 more|arXiv (Cornell University)|Apr 4, 2023

Medical Imaging Techniques and ApplicationsMedicine60 references3 citations

TL;DR

CoreDiff proposes a contextual error-modulated generalized diffusion model for low-dose CT denoising that accelerates inference by using LDCT images as the starting point of sampling, employs a mean-preserving degradation operator to model physical CT noise, and introduces CLEAR-Net to reduce error accumulation. It achieves state-of-the-art denoising performance with only 10 sampling steps, enabling clinically viable inference time.

ABSTRACT

Low-dose computed tomography (CT) images suffer from noise and artifacts due to photon starvation and electronic noise. Recently, some works have attempted to use diffusion models to address the over-smoothness and training instability encountered by previous deep-learning-based denoising models. However, diffusion models suffer from long inference times due to the large number of sampling steps involved. Very recently, cold diffusion model generalizes classical diffusion models and has greater flexibility. Inspired by the cold diffusion, this paper presents a novel COntextual eRror-modulated gEneralized Diffusion model for low-dose CT (LDCT) denoising, termed CoreDiff. First, CoreDiff utilizes LDCT images to displace the random Gaussian noise and employs a novel mean-preserving degradation operator to mimic the physical process of CT degradation, significantly reducing sampling steps thanks to the informative LDCT images as the starting point of the sampling process. Second, to alleviate the error accumulation problem caused by the imperfect restoration operator in the sampling process, we propose a novel ContextuaL Error-modulAted Restoration Network (CLEAR-Net), which can leverage contextual information to constrain the sampling process from structural distortion and modulate time step embedding features for better alignment with the input at the next time step. Third, to rapidly generalize to a new, unseen dose level with as few resources as possible, we devise a one-shot learning framework to make CoreDiff generalize faster and better using only a single LDCT image (un)paired with NDCT. Extensive experimental results on two datasets demonstrate that our CoreDiff outperforms competing methods in denoising and generalization performance, with a clinically acceptable inference time. Source code is made available at https://github.com/qgao21/CoreDiff.

Motivation & Objective

• Address the over-smoothing and training instability of prior deep learning-based LDCT denoising methods.
• Reduce the long inference time typical of diffusion models in medical imaging.
• Improve generalization to unseen dose levels without retraining or extensive hyperparameter tuning.
• Mitigate error accumulation and misalignment during the sampling process in diffusion-based denoising.
• Enable rapid adaptation to ultra-low-dose CT scenarios using only one paired LDCT/NDCT image.

Proposed method

• Utilizes low-dose CT (LDCT) images as the initial noisy input to the diffusion process, replacing random Gaussian noise, thereby reducing sampling steps.
• Introduces a novel mean-preserving degradation operator to simulate the physical CT image degradation process, improving sampling efficiency and fidelity.
• Proposes CLEAR-Net, a Contextual Error-modulated Restoration Network that leverages spatial context to constrain structural distortion and modulates time-step embeddings for better alignment across steps.
• Designs a one-shot learning framework that fine-tunes CoreDiff using only one (un)paired LDCT and NDCT image to generalize to new, unseen dose levels.
• Employs a generalized diffusion framework inspired by cold diffusion, enabling flexible and robust sampling from LDCT images.
• Integrates the degradation operator and CLEAR-Net into a unified diffusion sampling pipeline to enhance both speed and image quality.

Experimental results

Research questions

• RQ1 Can using LDCT images as the starting point of the diffusion process significantly reduce sampling steps while maintaining high denoising quality?
• RQ2 How effectively can a mean-preserving degradation operator model the physical noise characteristics of low-dose CT images compared to standard Gaussian noise?
• RQ3 Can CLEAR-Net reduce error accumulation and misalignment during diffusion sampling by leveraging contextual information and adaptive time-step modulation?
• RQ4 To what extent can CoreDiff generalize to new, unseen dose levels using only one paired LDCT/NDCT image in a one-shot learning setup?
• RQ5 Does CoreDiff achieve clinically acceptable inference speed while outperforming existing diffusion and non-diffusion models in ultra-low-dose scenarios?

Key findings

• CoreDiff achieves state-of-the-art denoising performance on four public LDCT datasets, outperforming competing methods in both quantitative metrics and visual quality.
• The model reduces inference time to just 10 sampling steps—10× faster than standard diffusion models—making it clinically viable despite the inherent cost of diffusion sampling.
• The one-shot learning framework enables rapid generalization to new, unseen dose levels using only a single LDCT/NDCT pair, eliminating the need for hyperparameter tuning on test data.
• CoreDiff demonstrates strong robustness in ultra-low-dose scenarios (e.g., 5% dose), where it maintains high image fidelity and diagnostic relevance.
• CLEAR-Net significantly reduces structural distortion and misalignment during sampling, as evidenced by improved PSNR and SSIM scores compared to baseline diffusion models.
• The mean-preserving degradation operator enhances sampling efficiency and fidelity by better approximating the actual physical degradation process in CT imaging.