Machine-Learning Based Detection of Coronary Artery Calcification Using Synthetic Chest X-Rays

TL;DR

This paper tackles the gap between costly CT-based CAC scoring and scalable CXRs by evaluating digitally reconstructed radiographs (DRRs) as a surrogate training domain. It presents a full pipeline using the COCA CT dataset to generate DRRs, apply pre-projection super-resolution and post-projection preprocessing, and train lightweight CNNs with various fusion and training strategies, achieving a mean AUC up to $0.754$ on DRRs. Key contributions include systematic analysis of fidelity restoration, signal enhancement, and multi-view fusion, showing that DRRs can provide a label-rich, scalable pretraining domain suitable for future domain adaptation to real CXRs. The work highlights the potential for DRR-based pretraining to support population-scale CAC screening with existing radiography infrastructure, while acknowledging the need for domain adaptation and real-CXR validation.

Abstract

Coronary artery calcification (CAC) is a strong predictor of cardiovascular events, with CT-based Agatston scoring widely regarded as the clinical gold standard. However, CT is costly and impractical for large-scale screening, while chest X-rays (CXRs) are inexpensive but lack reliable ground truth labels, constraining deep learning development. Digitally reconstructed radiographs (DRRs) offer a scalable alternative by projecting CT volumes into CXR-like images while inheriting precise labels. In this work, we provide the first systematic evaluation of DRRs as a surrogate training domain for CAC detection. Using 667 CT scans from the COCA dataset, we generate synthetic DRRs and assess model capacity, super-resolution fidelity enhancement, preprocessing, and training strategies. Lightweight CNNs trained from scratch outperform large pretrained networks; pairing super-resolution with contrast enhancement yields significant gains; and curriculum learning stabilises training under weak supervision. Our best configuration achieves a mean AUC of 0.754, comparable to or exceeding prior CXR-based studies. These results establish DRRs as a scalable, label-rich foundation for CAC detection, while laying the foundation for future transfer learning and domain adaptation to real CXRs.

Figures (3)

• Figure 1: A CT volume (left) is projected into a DRR using Siddon’s algorithm (DiffDRR) and fed to a classifier to predict a binary CAC label (Agatston $>100$).
• Figure 2: Preprocessing for Native DRR (Top) and Super-resolved DRR (bottom)
• Figure 3: Architecture of the custom CNN5_GAP model, consisting of five convolutional blocks and a 2-layer classifier.