Background and aims: Adenoma detection rate (ADR) is the principal quality indicator of colonoscopy and is inversely associated with post-colonoscopy colorectal cancer (PCCRC). Artificial intelligence (AI)-based computer-aided detection (CADe) systems built on deep convolutional neural networks provide real-time visual alerts for suspected polyps. We aimed to comprehensively synthesise randomised and primary non-randomised evidence published in the last 15 years on the efficacy, effectiveness and diagnostic performance of CADe-assisted colonoscopy for colorectal polyp detection. Methods: The study strictly adhered to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) 2020 guidelines. Eligible designs were randomised controlled trials (RCTs), comparative studies, cohorts and case series; systematic reviews and meta-analyses were excluded from the evidence base and considered only in the discussion. Risk of bias was assessed with RoB 2, ROBINS-I and the JBI case-series checklist. Random-effects meta-analyses used restricted maximum-likelihood heterogeneity estimation with Hartung–Knapp–Sidik–Jonkman (HKSJ) confidence intervals and 95% prediction intervals (PI). Certainty of evidence was rated with GRADE. Results: Sixty-three studies were included (43 RCTs, 12 non-randomised comparative studies and 8 case/validation series; more than 79,000 colonoscopies). In 34 parallel-group RCTs (n = 34,010), CADe increased ADR from 41.4% to 47.4% (risk ratio [RR] 1.16 (95% CI 1.11–1.22); I² = 66%; 95% PI 0.94–1.43; low certainty). CADe halved the adenoma miss rate in tandem RCTs (RR 0.49 (95% CI 0.32–0.75); moderate certainty), increased polyp detection (RR 1.21 (95% CI 1.08–1.35)) and adenomas per colonoscopy (mean difference 0.27 (95% CI 0.15–0.38)), but did not increase detection of advanced adenoma/advanced neoplasia (RR 1.04 (95% CI 0.95–1.14)) or, with statistical certainty, serrated lesions (RR 1.48 (95% CI 0.80–2.73)). The relative benefit declined with higher control-arm ADR (meta-regression P < 0.001; R² = 67%), from RR 1.45 when baseline ADR was below 30% to RR 1.08 when it was 45% or higher, and attenuated in more recent and FIT-positive trials. In non-randomised real-world cohorts the pooled effect was smaller and non-significant (odds ratio 1.20 (95% CI 0.98–1.46); very low certainty). Small-study effects were present (Egger P = 0.004), although the trim-and-fill-adjusted estimate remained favourable (RR 1.11). Stand-alone validation series reported per-lesion or per-image sensitivities of approximately 87–100%. Conclusions: CADe-assisted colonoscopy consistently increases the detection of predominantly diminutive, non-advanced adenomas and reduces miss rates, with the greatest gains among lower-performing endoscopists. Its effect on advanced neoplasia, serrated lesions and, ultimately, PCCRC incidence remains unproven, and benefits are attenuated in high-performing and real-world settings. Implementation should be targeted, monitored and coupled with strategies that mitigate automation bias and deskilling.
Copyrights © 2026