Artificial Intelligence Detection of Aortic Valve Calcification on Routine Noncontrast Chest CT: A Real-World Cohort

Routine noncontrast chest computed tomography (CT) may reveal aortic valve calcification (AVC) outside aortic stenosis evaluation. Its prevalence, burden, and documentation remain uncertain. Artificial intelligence (AI) may enable systematic detection.

The study aims to evaluate AVC prevalence and burden, clinical correlates, radiology documentation, and AI-assisted case finding.

This retrospective single-center study screened 511 consecutive noncontrast, non-electrocardiographically gated chest CT examinations in adults aged ≥50 years. After 15 protocol exclusions and 56 software-designated nonevaluable examinations, 440 underwent quantification by 2 blinded readers. AVC was defined as >10 AU using a 130-Hounsfield-unit threshold adjusted for blood-pool attenuation and image noise when needed. Locked HealthAVC V1.0 was a secondary comparator.

Manual quantification identified AVC in 172 of 440 scans (39.1%; 95% CI: 34.6%-43.7%); 97 of 440 (22.0%) had ≥100 AU. Inter-reader agreement at >10 AU was 99.5% (kappa = 0.990). Age (adjusted OR: 1.92 per 5 years; 95% CI: 1.63-2.26) and male sex (adjusted OR: 2.09; 95% CI: 1.32-3.30) were associated with AVC. Reports documented 27 of 172 positive scans (15.7%; 95% CI: 11.0%-21.9%). AI detected 169 of 172 manually confirmed cases (98.3%) among 440 evaluable scans, including 142 of 145 undocumented cases. Of 496 submitted scans, 56 (11.3%) were nonevaluable. End-to-end sensitivity counting positive rejected scans as missed was 89.9%. Among 86 AI-positive patients with echocardiography, 10 (11.6%) had mild-or-greater aortic stenosis.

AVC was common but infrequently documented on routine chest CT. AI detected most cases among evaluable scans; 11.3% were nonevaluable, yielding 89.9% end-to-end sensitivity. Prospective studies should assess clinical pathways.