Performance assessment of an artificial intelligence-based coronary artery calcium scoring algorithm in non-gated chest CT scans of different slice thickness
Kejie Yin,Wenping Chen,Guochu Qin,Jing Liang,Xue Bao,Hongming Yu,Hui Li,Jianhua Xu,Xingbiao Chen,Yujie Wang,Rock H Savage,U Joseph Schoepf,Dan Mu,Bing Zhang
DOI: https://doi.org/10.21037/qims-24-247
2024-08-01
Abstract:Background: The coronary artery calcium score (CACS) has been shown to be an independent predictor of cardiovascular events. The traditional coronary artery calcium scoring algorithm has been optimized for electrocardiogram (ECG)-gated images, which are acquired with specific settings and timing. Therefore, if the artificial intelligence-based coronary artery calcium score (AI-CACS) could be calculated from a chest low-dose computed tomography (LDCT) examination, it could be valuable in assessing the risk of coronary artery disease (CAD) in advance, and it could potentially reduce the occurrence of cardiovascular events in patients. This study aimed to assess the performance of an AI-CACS algorithm in non-gated chest scans with three different slice thicknesses (1, 3, and 5 mm). Methods: A total of 135 patients who underwent both LDCT of the chest and ECG-gated non-contrast enhanced cardiac CT were prospectively included in this study. The Agatston scores were automatically derived from chest CT images reconstructed at slice thicknesses of 1, 3, and 5 mm using the AI-CACS software. These scores were then compared to those obtained from the ECG-gated cardiac CT data using a conventional semi-automatic method that served as the reference. The correlations between the AI-CACS and electrocardiogram-gated coronary artery calcium score (ECG-CACS) were analyzed, and Bland-Altman plots were used to assess agreement. Risk stratification was based on the calculated CACS, and the concordance rate was determined. Results: A total of 112 patients were included in the final analysis. The correlations between the AI-CACS at three different thicknesses (1, 3, and 5 mm) and the ECG-CACS were 0.973, 0.941, and 0.834 (all P<0.01), respectively. The Bland-Altman plots showed mean differences in the AI-CACS for the three thicknesses of -6.5, 15.4, and 53.1, respectively. The risk category agreement for the three AI-CACS groups was 0.868, 0.772, and 0.412 (all P<0.01), respectively. While the concordance rates were 91%, 84.8%, and 62.5%, respectively. Conclusions: The AI-based algorithm successfully calculated the CACS from LDCT scans of the chest, demonstrating its utility in risk categorization. Furthermore, the CACS derived from images with a slice thickness of 1 mm was more accurate than those obtained from images with slice thicknesses of 3 and 5 mm.