ProDM: Synthetic Reality-driven Property-aware Progressive Diffusion Model for Coronary Calcium Motion Correction in Non-gated Chest CT

TL;DR

Coronary artery calcium (CAC) scoring from non-gated chest CTs is prone to motion artifacts, limiting reliable risk stratification. ProDM introduces a progressive diffusion-based motion correction framework guided by a differentiable calcium consistency loss and trained on realistically simulated non-gated data derived from gated calcium CTs, enabling robust restoration of motion-free CAC lesions. The approach uses a 2.5D diffusion backbone, a differentiable volume-based surrogate for CAC burden, and a motion-simulation data engine to produce paired training data, achieving superior CAC score preservation, lesion fidelity, and risk-category accuracy, with favorable radiologist assessments on real non-gated scans. These results suggest that property-aware, progressive diffusion methods can enable opportunistic CAC quantification from widely available non-gated CTs, potentially improving preventive cardiovascular care.

Abstract

Coronary artery calcium (CAC) scoring from chest CT is a well-established tool to stratify and refine clinical cardiovascular disease risk estimation. CAC quantification relies on the accurate delineation of calcified lesions, but is oftentimes affected by artifacts introduced by cardiac and respiratory motion. ECG-gated cardiac CTs substantially reduce motion artifacts, but their use in population screening and routine imaging remains limited due to gating requirements and lack of insurance coverage. Although identification of incidental CAC from non-gated chest CT is increasingly considered for it offers an accessible and widely available alternative, this modality is limited by more severe motion artifacts. We present ProDM (Property-aware Progressive Correction Diffusion Model), a generative diffusion framework that restores motion-free calcified lesions from non-gated CTs. ProDM introduces three key components: (1) a CAC motion simulation data engine that synthesizes realistic non-gated acquisitions with diverse motion trajectories directly from cardiac-gated CTs, enabling supervised training without paired data; (2) a property-aware learning strategy incorporating calcium-specific priors through a differentiable calcium consistency loss to preserve lesion integrity; and (3) a progressive correction scheme that reduces artifacts gradually across diffusion steps to enhance stability and calcium fidelity. Experiments on real patient datasets show that ProDM significantly improves CAC scoring accuracy, spatial lesion fidelity, and risk stratification performance compared with several baselines. A reader study on real non-gated scans further confirms that ProDM suppresses motion artifacts and improves clinical usability. These findings highlight the potential of progressive, property-aware frameworks for reliable CAC quantification from routine chest CT imaging.

Figures (8)

• Figure 1: Overview of CAC Motion Data Engine for simulating non-gated CT from gated calcium CT. The displacement trajectory of the calcified lesion is computed based on the selected CAC motion profile and projection parameters with added stochasticity, which is applied to the motion-free gated calcium CT introducing motion perturbation to the calcified lesion at each partial angle of the forward projection. The resulting sinogram then undergoes inverse Radon transform to obtain the simulated non-gated CT. The bottom-right panel demonstrates that the simulated CAC motion profiles cover a superset of real motion patterns, including simulated scans with motion profiles similar to real CAC motion (orange) as well as additional, less realistic scans (blue) introduced to increase diversity.
• Figure 2: Overview of the ProDM framework demonstrating the training and inference processes for CAC motion correction in non-gated chest CTs. The top figure shows training pipeline, where motion-corrupted non-gated CT volumes (orange) and corresponding motion-free gated calcium CT volumes (green) are used to construct supervised training pairs. Training is set up through a precomputed forward diffusion process where noise is progressively added to obtain intermediate states $x(t)$, a UNet denoiser learns to predict this noise at randomly sampled timesteps, guided by reconstruction loss and calcium consistency loss. Bottom figure shows the sliding-window inference pipeline used at test time, where all motion-corrupted patches within an ROI undergo progressive motion correction. The central slices of neighboring context windows are stacked together to form the final motion-corrected reconstruction of the ROI.
• Figure 3: Qualitative comparison of motion correction performance of different models on simulated motion-corrupted samples. SSIM with the ground truth is shown for reference.
• Figure 4: Radiologist assessment results from the reader study. (a) Violin plots show the distribution of radiologists' ratings for motion suppression, clinical usability, and vessel fidelity across methods. (b) Box-and-whisker plots summarize the mean ratings and variability for the original (non-corrected), UNet baseline, and ProDM, with corresponding $p$-values indicating statistical significance.
• Figure 5: Percentage-based confusion matrices for Agatston score–based cardiovascular risk stratification using three methods: no correction (motion-corrupted), UNet baseline, and ProDM. Each cell represents the proportion of subjects within a given ground-truth risk category (rows) assigned to a predicted category (columns). Subjects are grouped into five Agatston-based risk levels: 0 (no CAC), 1–10 (minimal CAC), 11–100 (mild CAC), 101–400 (moderate CAC), and $>$400 (severe CAC).