Can artificial intelligence add diagnostic value for human readers in detecting coronary artery stenosis on coronary CT angiography? A Bayesian network meta-analysis.
Authors
Affiliations (4)
Affiliations (4)
- Heilongjiang University of Chinese Medicine, Harbin 150040, Heilongjiang, China.
- Graduate School, Heilongjiang University of Chinese Medicine, Harbin 150040, Heilongjiang, China.
- Department of Medicine, Heilongjiang Academy of Chinese Medicine Sciences, Harbin 150036, Heilongjiang, China.
- Heilongjiang University of Chinese Medicine, Harbin 150040, Heilongjiang, China; The Second Department of Cardiovascular Diseases, First Affiliated Hospital, Heilongjiang University of Chinese Medicine, Harbin 150040, Heilongjiang, China. Electronic address: [email protected].
Abstract
The incremental diagnostic value of artificial intelligence (AI) in coronary CT angiography (CCTA) remains uncertain. We compared five CCTA strategies: automated stenosis assessment software (COR Analyzer), stand-alone CCTA-based deep learning (DL), junior readers, senior readers, and AI-assisted readers, for detecting anatomical coronary artery stenosis against invasive coronary angiography. PubMed, Embase, Web of Science, and the Cochrane Library were searched from inception through July 9, 2026. Twenty-two observational diagnostic accuracy studies were included (19 retrospective and three prospective; 21 for ≥50% stenosis and six for ≥70% stenosis). In the primary patient-level analysis at ≥50% stenosis, pooled sensitivity for DL and AI-assisted readers was 0.91 each (95% credible interval, 0.87 to 0.94 and 0.86 to 0.94), versus 0.80 (0.73-0.87) for junior and 0.87 (0.82-0.91) for senior readers; AI assistance increased sensitivity by 0.10 versus junior readers (0.06-0.15) and 0.03 versus senior readers (0.01-0.06). Specificity was similar across DL and reader strategies (0.64-0.67) but markedly lower for COR Analyzer (0.25 [0.15 to 0.38]). In the smaller ≥70% network, AI-assisted reading showed the highest estimated sensitivity (0.80 [0.70 to 0.88]). Findings were robust to outlier exclusion; certainty was low to very low for most ≥50% outcomes. Stand-alone DL and AI-assisted reading had similar estimated diagnostic performance. AI assistance improved sensitivity most clearly for less-experienced readers. These accuracy comparisons did not establish formal equivalence or determine whether stand-alone DL could replace human interpretation in clinical practice.