Introduction
Coronary artery calcium (CAC) scoring is a critical clinical metric for predicting cardiovascular disease (CVD) risk. While ECG-gated CT scans are the established gold standard for CAC scoring due to their ability to synchronize imaging with the cardiac cycle, they are costly and not routinely used for general screening. In contrast, non-gated chest CT scans are performed significantly more frequently for various indications but suffer from severe motion artifacts. These artifacts "smear" calcium deposits, leading to inaccurate Agatston scores and unreliable clinical assessments.
The Challenge of Motion in Non-gated CT
The primary obstacle in correcting motion artifacts in non-gated CT is twofold. First, there is a lack of large-scale, paired "motion-corrupted" and "motion-free" datasets required for supervised learning. Second, standard image restoration models, such as U-Nets, often prioritize global image smoothness, which can inadvertently "wash out" small, high-intensity calcium lesions—the very features essential for accurate scoring.
Methodology: The ProDM Framework
To address these challenges, the authors propose the Property-aware Progressive Diffusion Model (ProDM). This framework integrates a physics-informed data generation engine with a property-aware diffusion process.
1. CAC Motion Simulation Data Engine
To overcome the data scarcity issue, ProDM utilizes a physics-based simulation engine. Starting with clean, gated CT scans, the engine applies 21 diverse motion profiles—including jitters, drifts, and oscillations—to simulate realistic cardiac movement. These motion-perturbed states are then processed through Radon transforms to generate synthetic sinograms, which are reconstructed into motion-corrupted CT images. This approach ensures that the synthetic data captures the complex artifacts inherent in the CT reconstruction process.
2. Property-Aware Learning and Calcium Consistency
A key innovation of ProDM is the inclusion of a Property-aware loss function. Since the clinical Agatston score is non-differentiable, the researchers developed a differentiable Volume Score surrogate. This loss function forces the model to preserve the specific physical properties of calcium lesions, such as intensity and spatial distribution, during the restoration process. This prevents the over-smoothing common in traditional deep learning models.
3. Progressive Diffusion Backbone
ProDM employs a diffusion-based architecture that progressively refines a motion-corrupted image into a motion-free state. The model uses a 2.5D strategy, processing 3-slice patches to maintain essential 3D anatomical context while remaining computationally efficient. During training, a U-Net denoiser learns to predict the noise added to the motion-free patches. During inference, the model iteratively removes artifacts to reconstruct a high-fidelity, motion-corrected image.
Experimental Results and Clinical Validation
The model was trained and evaluated using the Stanford COCA dataset and a large-scale synthetic benchmark of over 9,000 scans.
- Quantitative Performance: ProDM significantly outperformed baseline models. It achieved a Mean Absolute Error (MAE) in Agatston scores of 9.39, a stark improvement over the 37.88 MAE of uncorrected scans. Notably, standard U-Nets performed worse than uncorrected scans (MAE 45.16) due to their tendency to over-smooth small calcifications.
- Risk Stratification: The model demonstrated a 95.3% accuracy in classifying patients into correct clinical risk categories (e.g., Minimal, Mild, Moderate, Severe), which is vital for clinical decision-making.
- Radiologist Evaluation: In a blinded study involving 30 real-world non-gated scans, clinical readers rated ProDM-corrected images significantly higher than original scans in terms of motion suppression (2.97 vs. 1.42) and overall clinical usability (2.78 vs. 1.44) on a 5-point Likert scale.
Conclusion and Future Directions
ProDM represents a significant step toward making routine, non-gated chest CTs a viable tool for accurate CAC scoring. By bridging the gap between low-cost routine imaging and gold-standard gated scans, this technology could enable widespread, opportunistic screening for cardiovascular disease. Future work will focus on optimizing inference speeds and expanding the motion simulation engine to encompass even more complex patient-specific movement patterns.
Summary updated