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

Qi Gao; Zilong Li; Junping Zhang; Yi Zhang; Hongming Shan

Source-linked AI summary

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

Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, Hongming Shan

arXiv:2304.01814v2Apr 4, 2023eess.IVcs.CVcs.LGphysics.med-ph

TL;DR

LDCT denoising requires handling severe noise and artifacts while avoiding the long inference times of conventional diffusion models. CoreDiff uses LDCT-informed generalized diffusion, CLEAR-Net, and one-shot adaptation, and reports strong denoising and generalization performance with 0.12-second-per-slice inference. Its operator is not strictly mean-preserving in practice, and inference remains slower than RED-CNN- and GAN-based models.

Problem
LDCT images contain noise and artifacts, while diffusion models can require many iterative sampling steps and CT noise is not simply Gaussian.
Method
CoreDiff combines LDCT-informed generalized diffusion with a mean-preserving degradation operator, CLEAR-Net error correction, and one-shot learning for unseen dose levels.
Results
0.12 seconds per slice: CoreDiff reports superior quantitative, qualitative, and task-specific performance, while using 10 sampling steps and generalizing through paired or unpaired one-shot training.
Takeaways & Limitations
CoreDiff provides a clinically oriented LDCT denoising framework that supports rapid sampling and adaptation to new dose levels using a single LDCT image.
Takeaways & Limitations
The degradation operator may not be strictly mean-preserving, and CoreDiff inference remains 10 times slower than RED-CNN- and GAN-based models.

Abstract

from arXiv · show

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.

I. INTRODUCTION

LDCT denoising must suppress severe noise and artifacts without sacrificing detail, while diffusion models face costly iterative sampling. CoreDiff addresses these challenges with generalized diffusion, CLEAR-Net, and one-shot adaptation to unseen dose levels.

Motivation: LDCT images contain severe noise and artifacts that can compromise radiologists’ diagnosis, while raw-data reconstruction methods may blur edges or incur high computational cost.Image post-processing methods avoid requiring inaccessible raw data.
Motivation: Conventional deep-learning denoisers can over-smooth images, whereas GANs preserve textures but are difficult to train and optimize.These limitations motivate diffusion-based alternatives.
Diffusion-model limitations: 1,000 sampling steps: DDPM illustrates the expensive inference caused by progressively generating images from Gaussian noise.This computational burden limits diffusion-model deployment in practice.
CoreDiff: CoreDiff starts reverse sampling from informative LDCT images and uses a mean-preserving degradation operator to mimic CT degradation, reducing required sampling steps.The operator is designed to preserve the image expectation while representing CT-specific degradation.
CoreDiff: CLEAR-Net uses contextual information and error-modulated time-step features to reduce structural distortion, misalignment, and accumulated restoration error.Cold-diffusion restoration errors can accumulate because predictions become misaligned with subsequent time steps.
Generalization: CoreDiff includes one-shot learning that adapts to an unseen dose level using one LDCT image, optionally paired with NDCT.The framework targets rapid adaptation with few resources.

B. The Proposed CoreDiff Model

The proposed CoreDiff architecture combines LDCT-conditioned generalized diffusion with a degradation operator designed for CT physics. Its overview identifies the generalized diffusion process and the CLEAR-Net restoration network as core components.

Generalized diffusion: CoreDiff uses LDCT images as the endpoint of generalized diffusion and applies a new degradation operator to mimic CT image degradation.The architecture is designed around an LDCT-informed diffusion process rather than Gaussian-noise initialization alone.
Motivation: Prior diffusion-based LDCT methods model degradation by adding Gaussian noise, although CT noise has complex statistics that cannot be simply represented by a Gaussian distribution.This motivates a CT-specific degradation formulation.

1) Generalized diffusion model for low-dose CT:

CoreDiff replaces random Gaussian-noise initialization with LDCT-based degradation and a mean-preserving operator that better reflects CT dose reduction. This warm-state formulation retains LDCT noise statistics and enables sampling with fewer steps.

Mean-preserving degradation: The conventional degradation operator causes intermediate-image expectations to drift from x0 because √αt + (1 − αt) is not consistently 1.This produces CT-number shifts and departs from the physical process of dose reduction.
Generalized diffusion process: CoreDiff uses LDCT images xT rather than random Gaussian noise as the diffusion endpoint for LDCT denoising.The LDCT endpoint contains semantic information from the corresponding NDCT image, allowing sampling to begin from a warm state.
Mean-preserving degradation: The proposed operator D(x0, xT, t) = αtx0 + (1 − αt)xT preserves the expected clean image while retaining LDCT-specific noise statistics.The paper notes that practical artifacts and complicated noise may prevent strict mean preservation.
Sampling efficiency: Because sampling starts from an informative LDCT warm state instead of random noise, CoreDiff can use a smaller number of sampling steps.The LDCT image is treated as an intermediate state between the clean and random-noise states.

2) Contextual Error-modulated Restoration Network (CLEAR-Net):

CLEAR-Net addresses error accumulation and time-step misalignment during CoreDiff sampling. It uses adjacent-slice context to constrain structural continuity and an error-modulated module to adapt time-step features to the latest prediction.

CLEAR-Net design: CLEAR-Net splits each training time step into restoration and redegradation stages to mitigate accumulated errors from imperfect restoration.Stage I estimates the clean image from xt, while Stage II redegrades the latest prediction before another restoration step.
Contextual information: Adjacent starting-point slices are concatenated with xt to provide contextual input that constrains the network to produce continuous z-axis structures.The adjacent slices remain unchanged throughout sampling.
Error modulation: The error-modulated module calibrates misalignment between the next-step input and its time-step embedding features.It modulates the embedding features using factors estimated from the latest prediction and the initial LDCT image.
Error modulation: The modulated features are used after each upsampling and downsampling operation in CLEAR-Net during Stage II.This feature modulation is part of the network’s second-stage processing.

C. One-shot Learning for Rapid Generalization

The one-shot learning framework adapts trained CoreDiff to an unseen dose level using a single LDCT image, with or without a paired NDCT image. It freezes CLEAR-Net and learns only step-wise combination weights for denoised outputs.

Motivation: LDCT images vary across equipment and protocols, making rapid adaptation to unseen dose levels with few resources an important clinical question.The framework targets this adaptation setting directly.
One-shot adaptation: The one-shot framework learns from one single LDCT image by combining denoised images generated across CoreDiff time steps.The mean-preserving operator yields outputs with the same mean and varying noise levels.
Optimization: During adaptation, CLEAR-Net parameters are frozen and only the T step-wise weights are learned to form the optimal denoised image.A perceptual loss guides these weights to improve visual perception without over-smoothing.
Unpaired adaptation: The framework does not introduce structural distortions when the new LDCT and NDCT images are unpaired because all intermediate outputs correspond to the same NDCT image.

III. EXPERIMENTS AND RESULTS

The experiments use simulated dose levels, a Mayo Clinic dataset with multiple vendors and dose settings, and external mixed-dose testing. The evaluation covers denoising and generalization across doses and datasets.

Experimental setup: The experiments use four datasets spanning different dose levels, centers, and imaging objects.
Mayo 2016 Dataset: The NIH-AAPM-Mayo dataset contains 5,936 1 mm-thickness NDCT slices from 10 patients, with nine patients for training and one for testing.Low-dose projections are generated with a Poisson+Gaussian noise model.
Dose settings: The simulated data include 50%, 25%, 10%, and 5% doses, with 5% representing the ultra-low-dose setting.Deep learning methods are trained and tested at either 25% or 5% dose, while 50% and 25% doses assess generalization.
External testing: The Mayo 2020 external dataset contains 299 scans from two vendors, including 25% head and abdomen data and 10% chest data.The mixed-dose test set uses five chest and five abdomen scans containing 800 images.

2) Mayo 2020 Dataset:

The supplied passages identify datasets, comparison methods, implementation settings, and image-quality metrics used in the study’s evaluation.

Test datasets: The evaluation uses real piglet data spanning 50%, 25%, 10%, and 5% dose scans, with 25% and 10% selected for testing.The dataset contains 850 CT images acquired with a GE Discovery CT750 HD scanner.
Test datasets: The phantom evaluation uses 271mAs and 108mAs scans, corresponding to 54.31% and 21.64% dose levels.Slices 10 to 21 were selected to preserve visibility of cylindrical implants.
Comparison methods: The study compares PWLS, RED-CNN, PDF-RED-CNN, GAN-based methods, and diffusion-based methods including DDM2 and IDDPM.CoreDiff and related variants are evaluated alongside these baselines.
Implementation details: CoreDiff uses T = 10 for both training and inference, whereas IDDPM is compared using 1000, 50, and 10 sampling steps.The resulting IDDPM models are named IDDPM-1000, IDDPM-50, and IDDPM-10.
Evaluation metrics: Evaluation includes PSNR, SSIM, RMSE, FSIM, VIF, and NQM as objective image-quality metrics.The additional metrics are intended to align more closely with radiologists’ subjective assessments.

C. Performance Comparison on Mayo 2016 Dataset

On Mayo 2016 data, CoreDiff is evaluated at 25% and 5% dose levels using quantitative metrics, qualitative images, lesion CNR, and inference-time comparisons.

25% dose evaluation: At 25% dose, CoreDiff preserves texture and detail better than competing methods in representative qualitative comparisons.RED-CNN-based methods blur details, while several GAN and diffusion baselines exhibit artifacts or structural issues.
25% dose evaluation: CoreDiff outperforms all compared methods on 25% dose data and exceeds PDF-RED-CNN by a large margin across all metrics.The comparison includes deep-learning methods and iterative reconstruction.
5% dose evaluation: At 5% dose, CoreDiff shows the best qualitative denoising performance among the compared methods in residual maps and zoomed-in ROIs.The ultra-low-dose setting produces severe noise and streak artifacts in FBP, while several baselines become over-smoothed or artifact-prone.
5% dose evaluation: +1.46 dB PSNR, +1.39% SSIM, and -15.45% RMSE are the average improvements of CoreDiff over PDF-RED-CNN on 5% dose test data.The same evaluation reports that CoreDiff’s inference speed is faster than diffusion-based models and clinically acceptable.
Lesion detectability: CNR evaluation compares lesion and background ROIs to assess the detectability of low-contrast lesions.Higher CNR corresponds to a higher probability of detecting low-contrast lesions.

D. Ablation Study

Ablation studies examine sampling-step choices, model components, and one-shot generalization across dose levels, centers, species, and phantom data.

Ablation on different T settings: T = 10 is selected as a suitable CoreDiff setting by balancing image sharpness, quantitative performance, and inference time.T = 1 produces blurred edges, while T ≥ 50 increases accumulated errors and reduces quantitative performance.
Ablation of model components: +3.08 dB PSNR, +2.23% SSIM, and -29.85% RMSE are achieved by the full model over the original cold-diffusion baseline.The ablation attributes contributions to the mean-preserving generalized diffusion process, contextual information, and EMM in CLEAR-Net.
One-shot generalization: The one-shot learning framework evaluates generalization to new doses, collections from the same center, different species and centers, and phantom data.The experiments cover Mayo 2016, Mayo 2020, piglet, and phantom datasets.

1) Generalization to new dose levels on the Mayo 2016 dataset:

CoreDiff was evaluated for generalization to unseen dose levels on Mayo 2016, Mayo 2020, and piglet data. One-shot learning improved texture similarity and quantitative performance, although comparisons involving additional training data are not fully fair.

Mayo 2016 dataset: CoreDiff was trained on 5% dose data and evaluated on 50% and 25% dose Mayo 2016 test data.Comparisons involving PDF-RED-CNN∗, CoreDiff+OSLp, and CoreDiff+OSLu used additional 50% and 25% dose training data.
One-shot learning: The generalization experiments combined one-shot models trained from single dose-specific slices with unpaired or paired normal-dose slices.Separate one-shot models were trained for Mayo 2020 chest and abdomen slices before testing on mixed dose levels.
Mayo 2020 dataset: On Mayo 2020 chest CT, diffusion-based methods provided a favorable noise-suppression and image-fidelity trade-off, while OSL produced textures closer to ground truth.The comparison included 10% dose images and multiple CNN, GAN, and diffusion baselines.
Mayo 2016 dataset: CoreDiff outperformed PDF-RED-CNN∗ and IDDPM quantitatively even without one-shot learning, while OSL further improved PSNR and SSIM.The paper attributes this to progressive denoising, contextual error modulation, and learning an optimal denoised image.
Piglet dataset: On piglet data, CoreDiff and PDF-RED-CNN∗ preserved fine details at 10% dose, while CoreDiff+OSLp and CoreDiff+OSLu produced textures closest to NDCT.At 25% dose, most methods except the two OSL variants had quantitative metrics worse than FBP.

4) Generalization to the phantom dataset:

CoreDiff was tested on phantom data for low-contrast signal detection and qualitative denoising. It achieved the strongest task-specific performance, with one-shot learning providing further enhancement.

Qualitative results: CoreDiff preserved edge details better than competing methods in the 21.64% dose scenario, while RED-CNN and CNCL-U-Net blurred edges and WGAN-VGG and IDDPM introduced artifacts.PDF-RED-CNN∗ outperformed other compared methods but still fell short of CoreDiff on edge preservation at 21.64% dose.
Task-specific evaluation: The phantom evaluation used CHO to quantify low-contrast signal detection from denoised images.Representative images covered 54.31% and 21.64% dose levels.
Quantitative results: CoreDiff achieved the best task-specific performance, and integrating the OSL framework further enhanced performance beyond NDCT.The paper attributes poorer competing-model performance partly to distribution shift between phantom test data and patient training data.

IV. DISCUSSION

The discussion presents CoreDiff as a generalized diffusion approach that improves LDCT denoising, generalization, and sampling efficiency. It also identifies physical-model mismatch, residual inference cost, and limited reader-study evidence as boundaries.

Benefits of CoreDiff: CoreDiff outperformed competing models in quantitative, qualitative, and task-specific performance, including against a 1,000-step IDDPM variant.The discussion attributes advantages to deterministic sampling, the degradation operator, and CLEAR-Net.
Benefits of CoreDiff: The mean-preserving degradation operator introduces LDCT-specific noise and artifacts while preserving each intermediate pixel’s mean CT number relative to NDCT.This design is intended to mimic physical CT degradation during the diffusion process.
Benefits of CoreDiff: CLEAR-Net uses contextual information to constrain sampling against structural distortion and modulates time-step features to address accumulated restoration error.The discussion frames error accumulation as especially important for pixel-accurate denoising.
One-shot learning framework: One-shot learning generalized CoreDiff across four datasets with few resources, but further work is needed for high-dose-trained models tested on lower-dose data.When test dose was lower than training dose, CoreDiff without OSL after 10 steps was already the best obtainable result reported by the authors.
Limitations: The degradation operator is not strictly mean-preserving when LDCT noise has nonzero mean, and its physical approximation requires improvement.The authors cite scattering, beam hardening, patient motion, and metal implants as contributing factors to complex LDCT noise.
Limitations: CoreDiff inference remains 10 times slower than RED-CNN-based and GAN-based models, while fewer sampling steps may trade generalization performance for speed.The paper also notes that a reader study with only two lesions lacks statistical significance and plans a larger MRMC study.
Conclusion: The conclusion reports that CoreDiff uses 10 sampling steps and is much faster than classical diffusion models for clinical use.The method combines LDCT endpoints, mean-preserving degradation, CLEAR-Net, and one-shot learning.