Performance of artificial intelligence-based coronary artery calcium scoring in non-gated chest CT. Issue 145 (December 2021)

• Record Type: Journal Article
• Authors: Xu, Jie; Liu, Jia; Guo, Ning; Chen, Linli; Song, Weixiang; Guo, Dajing; Zhang, Yu; Fang, Zheng
• Is Part Of: European journal of radiology. Issue 145(2021)
• Publication Date: 2021-12
• Subjects: Artificial intelligence -- Coronary artery -- Calcium; ADMIRE Advanced modelled iterative reconstruction algorithm -- AI Artificial intelligence -- ASIR Adaptive statistical iterative reconstruction -- AUC Area under the curve -- CABG Coronary artery bypass grafting -- CACS Coronary artery calcium score -- CAD Coronary artery disease -- CCTA Coronary computed tomography angiography -- CT Computed tomography -- DSCT Dual-source computed tomography -- ECG Electrocardiogram -- FBP Filtered back projection -- HIS Hospital information system -- PCI Percutaneous transluminal coronary intervention -- ROC Receiver operating characteristic; Medical radiology -- Periodicals; Radiology -- Periodicals; Radiologie médicale -- Périodiques; Medical radiology; Periodicals; 616.075705
• DOI: 10.1016/j.ejrad.2021.110034
• Languages: English
• ISSNs: 0720-048X

Abstract

Highlights: AI-CACS can provide CACS information based on non-gated chest CT. AI-CACS can serve for risk classification of coronary artery disease. AI-CACS has stable performance over different vendors and scanner types. Abstract: Objectives: To evaluate the risk category performance of artificial intelligence-based coronary artery calcium score (AI-CACS) software used in non-gated chest computed tomography (CT) on three types of CT machines, considering the manual method as the standard. Methods: A total of 901 patients who underwent both chest CT and electrocardiogram (ECG)-gated non-contrast-enhanced cardiac CT with the same equipment within a 3-month period were enrolled in the study. AI-CACS software was based on a deep learning algorithm and was trained on multi-vendor, multi-scanner, and multi-hospital anonymized data from the chest CT database. The AI-CACS was automatically obtained from chest CT data by the AI-CACS software, while the manual CACS was obtained from cardiac CT data by the manual method. The correlation of the AI-CACS and manual CACS, concordance rate and kappa value of the risk categories determined by the two methods were calculated. The chi-square test was used to evaluate the differences in risk categories among the three types of CT machines from different manufacturers. The risk category performance of the AI-CACS for dichotomous risk categories bounded by 0, 100 and 400 was assessed. Results: The correlation of the AI-CACS with the manual CACS was ρ = 0.893 (p < 0.001). The Bland-Altman plot (AI-CACS minus manual CACS) showed a mean difference of −27.2 and 95% limits of agreement of −290.0 to 235.6. The agreement of risk categories for the CACS was kappa (κ) = 0.679 (p < 0.001), and the concordance rate was 80.6%. The risk categories determined by the AI-CACS software on three types of CT machines were not significantly different (p = 0.7543). As dichotomous risk categories bounded by 0, 100 and 400, the accuracy, kappa value, and area under the curve of the AI-CACS were 88.6% vs. 92.9% vs. 97.9%, 0.77 vs. 0.77 vs. 0.83, and 0.885 vs. 0.964 vs. 0.981, respectively. Conclusions: There was good correlation and agreement between the AI-CACS and manual CACS in terms of the risk category. It is feasible to obtain the CACS using AI software based on non-gated chest CT data in a short time without increasing the radiation dose or economic burden. The AI-CACS software algorithm has good clinical universality and can be applied to CT machines from different manufacturers.