Agreement between fully automated artificial intelligence and cardiac radiologists for coronary stenosis and plaque assessment in an asymptomatic screening cohort.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Wonju College of Medicine, Wonju Severance Christian Hospital, Yonsei University, 20 Ilsan-ro, Wonju, 26426, Republic of Korea.
- Department of Radiation Convergence Engineering, Yonsei University, Wonju, Republic of Korea.
- CT Research Collaboration, Siemens Healthineers Ltd, Seoul, Republic of Korea.
- Computed Tomography, Siemens Healthineers AG, Forchheim, Germany.
- Department of Radiology, Wonju College of Medicine, Wonju Severance Christian Hospital, Yonsei University, 20 Ilsan-ro, Wonju, 26426, Republic of Korea. [email protected].
Abstract
This retrospective single-center study evaluated segment-level agreement between experienced cardiac radiologists and a research AI platform for coronary stenosis and plaque assessment on coronary computed tomography angiography (CCTA). Asymptomatic individuals undergoing CCTA at a health screening center were studied, excluding those with prior revascularization, coronary anomaly, or total occlusion. Stenosis grade and plaque composition were independently assessed by the AI platform and two cardiac radiologists at the segment level, with disagreements between the two radiologists resolved by consensus. Agreement was evaluated using Cohen's κ, quadratic weighted κ for ordinal grading, observed agreement, and positive and negative agreement for dichotomized significant stenosis. The cohort comprised 172 participants and 2,582 segments. Significant stenosis was uncommon (18/2,582, 0.7%), with moderate agreement (κ = 0.412; observed agreement 98.5%; positive agreement 0.418; negative agreement 0.992). Agreement was fair for ordinal stenosis grade (κ = 0.397; quadratic weighted κ = 0.573; observed agreement 84.0%) and plaque composition (κ = 0.330; observed agreement 81.8%). Stenosis-grade agreement was lower in plaque-containing than plaque-free segments, an exploratory finding given marked differences in subgroup size and prevalence. Myocardial bridging and image quality were not major descriptive determinants of discordance. In this low-prevalence cohort, AI-based CCTA analysis showed high negative agreement for radiologist-defined significant stenosis but limited concordance for grading and plaque characterization. Reflecting low prevalence and few positive cases, the findings do not establish independent rule-out performance but support prospective evaluation of AI as an expert-supervised pre-reading or triage adjunct, with positive, ambiguous, or plaque-containing findings requiring expert review.