Introduction
Coronary artery disease (CAD) remains a primary driver of global mortality. Clinical assessment often relies on the Agatston score, a metric derived from non-contrast CT scans that quantifies calcified lesions in the coronary arteries. While highly predictive of cardiovascular events, manual calculation of the Agatston score is labor-intensive and subject to significant inter-observer variability. This paper presents a fully automated, 3D deep learning approach designed to calculate Coronary Artery Calcium (CAC) scores with high precision and speed, potentially streamlining routine clinical workflows.
Methodology
The proposed system utilizes an ensemble of five 3D U-Net architectures. Unlike traditional 2D models that process individual image slices, the 3D approach leverages volumetric spatial context. This is crucial for distinguishing coronary calcium from anatomically adjacent structures, such as the mitral valve or aortic wall, which often lead to false positives in simpler models.
Key technical features include:
- Architecture: Four-level deep U-Nets with residual connections and Exponential Linear Unit (ELU) activations.
- Pre-processing: Volumes are downsampled to a standard isotropic resolution. To prevent the loss of low-density calcifications during this process, an intensity-enhancement step is applied to voxels exceeding 130 Hounsfield Units (HU).
- Ensemble Strategy: The final segmentation is determined by majority voting across the five models. This strategy reduces the magnitude of errors and improves overall robustness compared to individual models.
Datasets and Distribution
The model was trained and validated using a diverse set of data to ensure cross-vendor and cross-site generalizability:
- SFR Data Challenge 2020: 520 CT examinations from multiple clinical sites and four different vendors, divided into subsets J1, J2, and J3.
- Internal AD Dataset: 361 annotated CT acquisitions used to expand the training pool and improve the Concordance Index.
- orCaScore Challenge: A public benchmark of 40 test volumes used for external validation and head-to-head comparison.
As shown in the figure above, the datasets encompass a range of risk categories (A through E), representing varying degrees of calcium buildup. The model's ability to generalize across these distributions is key to its clinical utility.
Results and Discussion
The automated system demonstrates state-of-the-art performance across multiple metrics:
- Accuracy: On the J3 test set, the final model achieved a Concordance Index (C-index) of 0.951 and a category accuracy of 85.7%.
- Error Profile: Most misclassifications were "off-by-one" category shifts (e.g., predicting category B instead of A), which are clinically less severe than major misses.
- Benchmark Performance: The model ranked first on the orCaScore leaderboard at the time of submission, with an F1-score of 0.974.
- Efficiency: The entire process, from raw CT to final Agatston score, takes less than two seconds per patient, a significant improvement over manual segmentation.
While currently validated on dedicated ECG-gated CT scans, the next frontier is "opportunistic screening." Applying this model to routine, non-gated scans (e.g., lung cancer screenings) could enable incidental cardiovascular risk assessment for a much broader patient population without additional radiation exposure or cost.
Summary updated