Automated total and vessel-specific coronary artery calcium (CAC) quantification on chest CT: direct comparison with CAC scoring on non-contrast cardiac CT.

BACKGROUND

This study aimed to evaluate the artificial intelligence (AI)-based coronary artery calcium (CAC) quantification and regional distribution of CAC on non-gated chest CT, using standard electrocardiograph (ECG)-gated CAC scoring as the reference.

METHODS

In this retrospective study, a total of 405 patients underwent non-gated chest CT and standard ECG-gated cardiac CT. An AI-based algorithm was used for automated CAC scoring on chest CT, and Agatston score on cardiac CT was manually quantified. Bland-Altman plots were used to evaluate the agreement of absolute Agatston score between the two scans at the patient and vessel levels. Linearly weighted kappa (κ) was calculated to assess the reliability of AI-based CAC risk categorization and the number of involved vessels on chest CT.

RESULTS

The AI-based algorithm showed moderate reliability for the number of involved vessels in comparison to measures on cardiac CT (κ = 0.75, 95% CI 0.70-0.79, P < 0.001) and an assignment agreement of 76%. Considerable coronary arteries with CAC were not identified with a per-vessel false-negative rate of 59.3%, 17.8%, 34.9%, and 34.7% for LM, LAD, CX, and RCA on chest CT. The leading causes for false negatives of LM were motion artifact (56.3%, 18/32) and segmentation error (43.8%, 14/32). The motion artifact was almost the only cause for false negatives of LAD (96.6%, 28/29), CX (96.7%, 29/30), and RCA (100%, 34/34). Absolute Agatston scores on chest CT were underestimated either for the patient and individual vessels except for LAD (median difference: - 12.5, - 11.3, - 5.6, - 18.6 for total, LM, CX, and RCA, all P < 0.01; - 2.5 for LAD, P = 0.18). AI-based total Agatston score yielded good reliability for risk categorization (weighted κ 0.86, P < 0.001) and an assignment agreement of 86.7% on chest CT, with a per-patient false-negative rate of 15.2% (28/184) and false-positive rate of 0.5% (1/221) respectively.

CONCLUSIONS

AI-based per-patient CAC quantification on non-gated chest CT achieved a good agreement with dedicated ECG-gated CAC scoring overall and highly reliable CVD risk categorization, despite a slight but significant underestimation. However, it is challenging to evaluate the regional distribution of CAC without ECG-synchronization.