Coronary stenosis assessment is central to invasive treatment decisions in coronary artery disease. Coronary CT angiography offers a non-invasive route, while invasive coronary angiography often guides revascularisation through visual assessment. Both can be limited when stenosis severity is difficult to judge, especially where image quality, calcification or reader experience affect interpretation. A 2026 analysis published in Insights into Imaging compared AI-based CT quantification, manual CT quantification and visual invasive angiography assessment against quantitative coronary angiography. AI-based CT quantification showed stronger performance and closer agreement with the reference method in most comparisons.
Imaging Methods and Patient Selection
The analysis included patients with suspected or known coronary artery disease who underwent coronary CT angiography and invasive coronary angiography within one month. The final cohort included 368 patients after exclusions for previous coronary bypass grafting, previous stenting and severely impaired CT image quality. Coronary vessels and segments below the required diameter threshold were also excluded from the final assessment.
Must Read: AI Sharpens Low-Dose Paediatric CT
Manual CT quantification relied on radiologists measuring coronary stenosis on CT images, using an established segment model. Senior and junior cardiovascular radiologists assessed the images independently, without access to clinical history, invasive angiography results or automated CT quantification.
The AI-based approach used a commercially available system to automate several parts of coronary image analysis. These included vessel segmentation, centreline tracking, branch identification, plaque detection and stenosis assessment. The model generated diameter stenosis values directly from the CT images, without manual adjustment before comparison.
Invasive coronary angiography underwent visual assessment by interventional cardiologists. Quantitative coronary angiography then provided the reference measurement. Each method assessed stenosis severity using the same coronary segment model, allowing comparison across patient, vessel and segment levels.
Performance Across Stenosis Thresholds
Quantitative coronary angiography identified obstructive stenosis at both the 50% and 70% diameter stenosis thresholds. AI-based CT quantification achieved high diagnostic accuracy across the evaluated levels. Its performance was strongest for the 50% threshold and remained high for the more severe 70% threshold.
The AI-based method outperformed manual CT quantification and visual invasive angiography assessment in most comparisons. This advantage appeared across patient, vessel and segment levels. The main exception involved segment-level assessment at the 70% threshold, where senior radiologist manual CT quantification performed close to the AI-based method and the difference was not statistically significant.
Manual CT quantification showed differences between senior and junior readers. Senior radiologists performed better than junior radiologists in identifying obstructive stenosis across the evaluated thresholds and levels. Reader experience therefore remained relevant in manual assessment.
Visual assessment of invasive angiography also showed lower agreement with quantitative coronary angiography than the AI-based CT approach. The visual method tended to produce larger positive differences from the reference measurement, which matters because visual estimation during invasive angiography is widely used to guide revascularisation decisions.
Agreement, Calcium Burden and Occlusion
AI-based CT quantification showed strong correlation with quantitative coronary angiography at patient, vessel and segment levels. It also produced closer agreement with quantitative coronary angiography than manual CT quantification or visual invasive angiography assessment. The mean differences between AI-based CT quantification and the reference method were small across all three levels.
However, some discordance remained. A difference of more than 30% between AI-based CT quantification and quantitative coronary angiography appeared in a small share of lesion-bearing segments. The AI model also showed a specific limitation in total occlusion. Segments classified as total occlusion by quantitative coronary angiography were identified by the AI model as severe stenosis rather than complete obstruction.
Calcium burden influenced CT-based assessment. Patient-level performance for CT-based methods tended to decline as coronary calcium score increased. This trend was especially apparent for manual quantification by junior radiologists. Visual invasive angiography assessment maintained more stable performance across calcium score categories.
The limitations affect interpretation. The cohort did not include patients with previous coronary bypass grafting or stenting. CT images came from five scanner models, but mid-end CT scanners were not included. Patient selection may also have favoured a population with a higher probability of coronary artery disease because all included patients underwent both CT and invasive angiography within one month.
AI-based CT quantification shows high performance for assessing obstructive coronary stenosis when compared with quantitative coronary angiography. It performs better than manual CT quantification and visual invasive angiography assessment in most scenarios, while senior radiologist assessment remains close at segment level for more severe stenosis. The approach may support more consistent CT interpretation, particularly where manual assessment varies by experience. Its limitations remain important, especially in total occlusion, high calcium burden and populations not represented in the cohort. Careful review of AI output remains necessary before treatment decisions.
Source: Insights into Imaging
Image Credit: iStock
References:
Yu L, Yuan M, Dai X et al. (2026) Coronary stenosis assessment: AI-based CT quantification, visual analysis of invasive angiography, and quantitative coronary angiography. Insights Imaging; 17, 134.
