Performance of visual, manual, and automatic coronary calcium scoring of cardiac 13N-ammonia PET/low dose CT

Abstract

Background: Coronary artery calcium is a well-known predictor of major adverse cardiac events and is usually scored manually from dedicated, ECG-triggered calcium scoring CT (CSCT) scans. In clinical practice, a myocardial perfusion PET scan is accompanied by a non-ECG triggered low dose CT (LDCT) scan. In this study, we investigated the accuracy of patients’ cardiovascular risk categorisation based on manual, visual, and automatic AI calcium scoring using the LDCT scan. Methods: We retrospectively enrolled 213 patients. Each patient received a 13N-ammonia PET scan, an LDCT scan, and a CSCT scan as the gold standard. All LDCT and CSCT scans were scored manually, visually, and automatically. For the manual scoring, we used vendor recommended software (Syngo.via, Siemens). For visual scoring a 6-points risk scale was used (0; 1-10; 11-100; 101-400; 401-100; > 1 000 Agatston score). The automatic scoring was performed with deep learning software (Syngo.via, Siemens). All manual and automatic Agatston scores were converted to the 6-point risk scale. Manual CSCT scoring was used as a reference. Results: The agreement of manual and automatic LDCT scoring with the reference was low [weighted kappa 0.59 (95% CI 0.53-0.65); 0.50 (95% CI 0.44-0.56), respectively], but the agreement of visual LDCT scoring was strong [0.82 (95% CI 0.77-0.86)]. Conclusions: Compared with the gold standard manual CSCT scoring, visual LDCT scoring outperformed manual LDCT and automatic LDCT scoring.