Predicting categories of coronary artery calcium scores from chest X-ray images using deep learning

Abstract

Background: The coronary artery calcium (CAC) score (CACS) is recommended in clinical guidelines for coronary artery disease evaluation. However, it is being replaced by coronary computed tomography angiography as the primary diagnostic tool for patients with stable chest pain. This study aimed to develop and validate a deep learning model for predicting the CACS categories from chest X-ray radiographs (CXRs).

Methods: We included 10,230 patients with available CXRs and CACSs obtained within six months. Three models were trained based on the CACS thresholds (0, 100, and 400) to distinguish zero from non-zero CACSs, CACSs of <100 and ​≥ ​100, and CACS of <400 and ​≥ ​400. The final CXR integration models incorporating clinical factors, including age, sex, and body mass index, were also trained. All models were evaluated using 10-fold cross-validation. External validation was also performed. We experimentally demonstrated the prognostic value of the predicted CACS for major adverse cardiovascular events, comparing it to the actual CACS classification.

Results: The CACS classification performance of the deep learning model was promising, with areas under the curve (AUCs) of 0.74 (zero vs non-zero), 0.75 (<100 vs. ≥100), and 0.79 (<400 vs. ≥400). The accuracy of the model further improved upon the integration of clinical factors; the AUCs reached 0.77, 0.79, and 0.82, respectively, for the same CACS categories. The external validation results were consistent (AUCs of 0.78, 0.79, and 0.81, respectively).

Conclusions: The deep learning model effectively classified the CACS from CXRs, especially for cases of severe calcification. This approach can cost-effectively improve coronary artery disease risk assessment and support clinical decision-making while minimizing radiation exposure.