Automatic coronary artery calcium scoring on routine chest computed tomography (CT): comparison of a deep learning algorithm and a dedicated calcium scoring CT.

Background: The aim of this study was to investigate the reliability and accuracy of automatic coronary artery calcium (CAC) scoring and risk classification in non-gated, non-contrast chest computed tomography (CT) of different slice thicknesses using a deep learning algorithm. Methods: This retrospective study was performed at 2 tertiary hospitals. Paired, dedicated calcium-scoring CT scans and non-gated, non-contrast chest CT scans taken within a month from the same patients were included. Chest CT images were grouped according to the slice thickness (group A: 1 mm; group B: 3 mm). For internal scans, the CAC score manually measured on dedicated calcium scoring CT images was used as the gold standard. The deep learning algorithm for group A was trained using 150 chest CT scans and tested using 144 scans, and that for group B was trained using 170 chest CT scans and tested using 144 scans. The intraclass correlation coefficient (ICC) was used to evaluate the correlation between the algorithm and the gold standard. Agreement between the deep learning algorithm, the manual results on chest CT, and the gold standard was determined by Bland-Altman analysis. Cardiac risk categories were compared. External validation was performed on 334 paired scans from a different organization. Results: A total of 608 internal paired scans (1 mm: 294; 3 mm: 314) of 406 individuals and 334 external paired scans (1 mm: 117; 3 mm: 117) of 117 individuals were included in the analysis. The ICCs between the deep learning algorithm and the gold standard were excellent in both group A (0.90; 95% CI: 0.85-0.93) and group B (0.94; 95% CI: 0.92-0.96). The Bland-Altman plots showed good agreement in both groups. For the cardiovascular risk category, the deep learning algorithm accurately classified 71% of cases in group A and 81% of cases in group B. The Kappa values for risk classification were 0.72 in group A and 0.82 in group B. External validation yielded equally good results. Conclusions: The automatic calculation of CAC score and cardiovascular risk stratification on non-gated chest CT using a deep learning algorithm was reliable and accurate on both 1 and 3 mm scans. Chest CT with a slice thickness of 3 mm was slightly more accurate in CAC detection and risk classification.