Optimizing AI-Derived CAC Volume Thresholds on Non-Gated Chest CT for Detection of Actionable Coronary Artery Calcium: A Validation Against Agatston Scoring and Radiologist Reporting.
Authors
Affiliations (6)
Affiliations (6)
- Department of Internal Medicine, Division of Cardiology.
- Gray Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel.
- School of Medicine.
- Department of Radiology, Division of Cardiothoracic Imaging, University of Texas Southwestern Medical Center, Dallas.
- Department of Radiology, Cardiothoracic Imaging Division, Mayo Clinic, Jacksonville, FL.
- Department of Thoracic Imaging, UT MD Anderson Cancer Center, Houston, TX.
Abstract
Coronary artery calcium (CAC) visually estimated on non-electrocardiogram (ECG)-gated chest CTs is reported inconsistently, limiting identification of high-risk individuals. Although automated AI-derived CAC scoring offers the potential to standardize reporting and improve reliability, this method warrants rigorous validation. Three hundred thirty-two patients (59% female, median age 62) who underwent a non-gated chest CT within 1-year of a dedicated cardiac CT with Agatston's CAC score (CACscore) were identified between 2016 and 2022. AI-Rad Companion (Siemens) was used to compute CAC volume (AI-CACvol). AI-CACvol performance was compared with radiologists' qualitative CAC reports (H-CACqual) using optimized thresholds for identifying actionable CAC. Model performance was further validated in a postimplementation population. Median AI-CACvol was 20.3 mm³ (IQR 0-188.65), strongly correlated with CACscore (Spearman ρ=0.91, 95% CI: 0.88-0.94); median 31.55 Agatston Units (AU), IQR: 0 to 250). Agreement with CACscore was higher for AI-CACvol versus H-CACqual (weighted kappa 0.88, 95% CI: 0.84, 0.91; vs. 0.85, 95% CI: 0.82, 0.88; P=0.022). At a cutoff of 70 mm³, AI-CACvol demonstrated superior sensitivity compared with H-CACqual for detecting CAC≥100 AU, with greater sensitivity (95% vs. 75%) and comparable specificity (92% vs. 94%). In a validation cohort of 39 patients (64% female, median age 67 y), median CACscore was 43 AU (IQR 0-215) and median AI-CACvol 29.7 mm³ (IQR: 0 to 219.4). The same 70 mm³ cutoff achieved 100% sensitivity and 88% specificity for CAC≥100 AU (n=13). AI-derived CAC volume from non-gated chest CTs correlates strongly with reference CAC scores, showing high sensitivity at actionable thresholds. This supports potential opportunistic screening, though impact on outcomes requires study.