Artificial Intelligence-assisted Versus Conventional Colonoscopy for Adenoma and Polyp Detection Rates: A Meta-analysis of Recent Randomized and Quasi-randomized Trials

doi:10.14218/CSP.2026.00015

Publications > Journals > Cancer Screening and Prevention> Article Full Text

Original Article
OPEN ACCESS

Artificial Intelligence-assisted Versus Conventional Colonoscopy for Adenoma and Polyp Detection Rates: A Meta-analysis of Recent Randomized and Quasi-randomized Trials

Gianmarco Adinolfi^1,* and
Valeria Milia²

Author information

Cancer Screening and Prevention 2026

doi: 10.14218/CSP.2026.00015

Abstract

Background and objectives

Artificial intelligence (AI)-based computer-aided detection (CADe) has been associated with improved adenoma detection during colonoscopy. However, prior meta-analyses synthesized earlier trials, and whether the benefit remains consistent in recent contemporary trials is uncertain. This meta-analysis aimed to estimate the effects of current-generation AI-assisted colonoscopy on adenoma detection rate (ADR) as the primary outcome and polyp detection rate (PDR) as the secondary outcome in randomized and quasi-randomized trials published from August 1, 2024, to August 18, 2026, without re-pooling trials included in earlier comprehensive meta-analyses.

Methods

MEDLINE/PubMed, Embase, CENTRAL, Scopus, and Google Scholar were searched for peer-reviewed parallel-group randomized and quasi-randomized trials enrolling adults undergoing screening, surveillance, or diagnostic colonoscopy and comparing real-time AI-based CADe-assisted with conventional high-definition white-light colonoscopy. Tandem designs were excluded. The primary and secondary outcomes were ADR and PDR, respectively. Random-effects risk ratios with 95% confidence intervals (CIs) were calculated; heterogeneity and leave-one-out sensitivity were assessed.

Results

Seventeen trials comprising 15,242 patients were included for ADR; 12 trials comprising 8,665 patients reported extractable PDR data. The pooled risk ratio was 1.14 (95% CI 1.09–1.20; I² = 53.8%) for ADR and 1.13 (95% CI 1.07–1.20; I² = 63.4%) for PDR.

Conclusions

Keywords

Adenoma detection rate, Computer-aided detection, Colonoscopy, Artificial intelligence, Meta-analysis, Colorectal cancer screening.

Introduction

Colorectal cancer (CRC) is a common malignancy, with most cases arising from precancerous adenomatous and serrated polyps,¹ and it represents one of the leading causes of cancer-related death worldwide.² Established risk factors for conventional adenomas and CRC include age, male sex, family history, obesity, physical inactivity, and red meat intake,¹ as well as cigarette smoking, which has been consistently associated with an increased risk of colorectal adenomas and cancer.³ Serrated polyps can additionally progress to CRC through a pathway distinct from the conventional adenoma–carcinoma sequence.⁴ Colonoscopy is widely used as a primary or follow-up screening test for CRC and has been shown to reduce cancer incidence and mortality through the detection of tumors at an earlier, more treatable stage and through removal of precancerous adenomas.^5,6

The effectiveness of colonoscopy in preventing CRC is closely linked to the quality of the examination, and the adenoma detection rate (ADR) has been established as one of the main quality indicators of screening colonoscopy.⁷ A large population-based study demonstrated that ADR is inversely associated with the risk of interval CRC, advanced-stage interval cancer, and cancer-related death, with each 1% increase in ADR corresponding to an approximate 3% reduction in interval CRC risk.⁶ Despite its clinical relevance, ADR shows substantial variability across endoscopists and countries, with reported rates ranging from 8% to 35%,⁷ and polyp miss rates as high as 27% have been described, largely attributable to both lesion- and operator-related factors.⁸ Consistently, quality-adjusted back-to-back colonoscopy studies have reported adenoma miss rates of approximately 17%, even under optimal examination conditions.⁹ These persistent miss rates highlight the need for more consistent and reliable detection strategies.

In recent years, artificial intelligence (AI)-based computer-aided detection (CADe) systems have been proposed as a tool to reduce operator-dependent variability and improve the diagnostic performance of colonoscopy, providing real-time alerts for suspicious lesions during the withdrawal phase of the procedure.^10,11

The evidence base surrounding CADe-assisted colonoscopy has continued to expand, with newer randomized and quasi-randomized trials evaluating increasingly diverse patient populations, clinical settings, and AI platforms. Previous systematic reviews and meta-analyses have already established a beneficial association between CADe and adenoma detection, predominantly on the basis of trials conducted during the earlier phases of AI-assisted colonoscopy. Consequently, repeated re-pooling of the same historical trials may provide limited additional information regarding the performance of newer CADe technologies.

A clinically relevant question is therefore whether the beneficial effect observed in earlier studies remains consistent in the more recent generation of randomized and quasi-randomized trials conducted with contemporary CADe systems and current colonoscopy practice. Evaluating this more recent evidence separately may also reduce the potential for historical technology, operator experience, and early implementation effects to influence the estimated magnitude of benefit.

Accordingly, the present meta-analysis was designed to evaluate randomized and quasi-randomized trials published between August 1, 2024, and August 18, 2026, following the search period of the most recent comprehensive systematic review used as the reference evidence synthesis. By focusing specifically on this contemporary evidence base, the study aimed to estimate the effect of current-generation CADe-assisted colonoscopy on ADR as the primary outcome and polyp detection rate (PDR) as the secondary outcome, without re-pooling trials already synthesized in earlier comprehensive meta-analyses.

Materials and methods

Study design and rationale

This study was designed as a meta-analysis of recently published randomized and quasi-randomized trials evaluating the effect of contemporary CADe systems on adenoma and polyp detection during colonoscopy. The study deliberately adopted a temporally restricted evidence framework rather than re-pooling the full historical CADe literature.

This approach was chosen because the effectiveness of CADe-assisted colonoscopy has already been evaluated in several comprehensive systematic reviews and meta-analyses, including large syntheses incorporating trials from the earlier phases of AI-assisted colonoscopy, including Makar et al.¹² (k = 28), Soleymanjahi et al.¹³ (k = 44), Eman et al.¹⁴ (k = 48), and Tan et al.¹⁵ (k = 64). Repeating a pooled analysis of the same historical trials would provide limited incremental information. At the same time, a substantial number of randomized and quasi-randomized trials evaluating newer CADe systems and contemporary clinical practice have been published subsequently.

Accordingly, a prespecified publication date window from August 1, 2024 through August 18, 2026 was applied. This temporal restriction was adopted to define a contemporary evidence window, minimize overlap with previously synthesized trials, and specifically capture randomized and quasi-randomized studies evaluating more recent generations of CADe systems and their implementation in current colonoscopy practice (see Section 2.3, publication-date handling).

The objective of this analysis was not to provide a new estimate based on the entire historical evidence base, but rather to determine whether the beneficial effect of CADe on adenoma and polyp detection remains consistent in the most recent generation of randomized and quasi-randomized clinical studies.

The study was reported in accordance with the PRISMA 2020 statement. The protocol was prospectively registered on PROSPERO (registration number: CRD420261446631).

Eligibility criteria (PICOS)

Population (P): Adult patients (≥18 years) undergoing colonoscopy for screening, surveillance, or diagnostic indications.
Intervention (I): Colonoscopy performed with real-time CADe using any commercially available or investigational AI system.
Comparator (C): Conventional high-definition white-light colonoscopy without AI assistance, with or without adjunctive non-AI techniques, provided these were not differentially restricted between arms.
Outcomes (O):

- Primary outcome: ADR — proportion of colonoscopies with ≥1 histologically confirmed adenoma.

- Secondary outcome: Polyp Detection Rate (PDR) — proportion of colonoscopies with ≥1 detected polyp, regardless of histology.

Study design (S): Parallel-group randomized controlled trials (RCTs) and quasi-randomized trials. Tandem (back-to-back) designs were excluded from the primary parallel-arm ADR/PDR pooling, given the differing unit of comparison (within-patient sequential passes vs. between-patient independent groups).

Publication date and search window

The publication date and search windows were prespecified and covered the period from August 1, 2024 through August 18, 2026. Publication date was defined as the date of final journal-issue assignment; where a trial had not yet been assigned to a final issue at the time of the search, the online-ahead-of-print publication date was used instead. The search was performed to identify randomized and quasi-randomized trials published within this predefined period evaluating real-time CADe-assisted colonoscopy compared with conventional colonoscopy.

Studies published before August 1, 2024 or after August 18, 2026 were considered outside the prespecified evidence window and were therefore excluded from the present quantitative synthesis on temporal grounds.

Information sources and search strategy

MEDLINE/PubMed, Embase, CENTRAL, Scopus, and Google Scholar were searched for eligible studies, applying the same date restriction uniformly across all five databases. No language restriction was applied. Only peer-reviewed publications were eligible; conference abstracts, trial registry entries, preprints, and other grey literature were not searched or included. Reference lists of relevant systematic reviews and meta-analyses were reviewed to cross-check the completeness of study identification from the database searches; potentially eligible primary studies were assessed against the same eligibility criteria.

Search terms combined four concept blocks joined with AND: (1) colonoscopy; (2) artificial intelligence / computer-aided detection; (3) adenoma or polyp detection outcomes, further restricted to randomized or quasi-randomized trial designs. The following terms were used, combined with OR within each block and AND across blocks:

Block 1 (colonoscopy): colonoscopy
Block 2 (AI/CADe): artificial intelligence, computer-aided detection, CADe
Block 3 (outcome): adenoma detection rate, ADR, polyp detection rate, PDR, polyp
Block 4 (design): randomized, randomized, randomized controlled trial, quasi-randomized, trial

The same conceptual search strategy was applied across all five databases, with database-specific adaptations of controlled vocabulary, field codes, Boolean syntax, and publication-date filters. The complete database-specific search strategies are provided in Supplementary Table. 1. All databases were last searched on August 18, 2026.

Study selection process

Two reviewers (G.A. and V.M.) independently screened all records after Ryyan-based deduplication and independently performed full-text review of potentially eligible articles; disagreements at both stages were resolved through discussion and consensus between the two reviewers, following the PRISMA 2020 flow diagram.

Data extraction

Two independent reviewers (G.A. and V.M.) screened titles and abstracts, followed by a full-text review of potentially eligible articles. Disagreements were resolved through discussion and consensus. Data were extracted independently by both reviewers using a standardized form. The following information was extracted from each trial:

Study characteristics: first author, publication year, trial name, study design;
Population characteristics: sample size;
Intervention details: colonoscopy performed with real-time CADe using any commercially available or investigational AI system;
Outcomes:
Primary outcome: ADR — proportion of colonoscopies with ≥1 histologically confirmed adenoma.
Secondary outcome: Polyp Detection Rate (PDR) — proportion of colonoscopies with ≥1 detected polyp, regardless of histology.

This resulted in a final analytic dataset of k = 17 studies for ADR and k = 12 studies for PDR.

Risk of bias assessment

Risk of bias for each included randomized controlled trial was assessed independently by two reviewers using the Cochrane Risk of Bias 2 (RoB 2) tool. The assessment covered five domains:

Bias arising from the randomization process;
Bias due to deviations from intended interventions;
Bias due to missing outcome data;
Bias in the measurement of outcomes;
Bias in the selection of the reported result.

Each domain was rated as “low risk,” “some concerns,” or “high risk.” Each study’s risk of bias was evaluated.

Because Lagström et al.²¹ used a quasi-randomized, center-level allocation based on alternating 2-week periods of CADe and control rather than true randomization, this study was assessed separately using the Risk Of Bias in Non-randomized Studies – of Interventions (ROBINS-I) tool. The assessment covered seven domains:

Bias due to confounding;
Bias in selection of participants into the study;
Bias in classification of interventions;
Bias due to deviations from intended interventions;
Bias due to missing data;
Bias in measurement of outcomes;
Bias in selection of the reported result.

Each ROBINS-I domain was rated as “low,” “moderate,” “serious,” or “critical” risk of bias, with an additional “no information” category where applicable. An overall risk-of-bias judgment was then assigned according to the ROBINS-I framework.

Effect measures and data synthesis

Risk ratios (RRs) with 95% confidence intervals (CIs) were used as the effect measure for both outcomes, calculated using a random-effects model with restricted maximum likelihood (REML) estimation in `metafor`,¹⁶ given the a priori expectation of clinical and methodological heterogeneity across CADe systems, populations, and settings — a modeling choice made independent of, and prior to, observation of the actual heterogeneity in the data.

Analyses were cross-validated in Python (PyMARE package, REML estimator), consistent with the metafor implementation. Confidence intervals used the standard Wald method, with a Knapp-Hartung adjustment computed as a sensitivity analysis given the moderate number of studies.¹⁷ Prediction intervals were calculated following Higgins et al.¹⁸ Study weights shown in Tables 1 and 2 are random-effects (REML-based) weights rather than fixed-effect weights,^19-35 which would understate the influence of heterogeneity, particularly for PDR. Statistical heterogeneity was assessed using Cochran’s Q, I², and τ².

Table 1

Study	CADe (n/N)	Control (n/N)	RR	95% CI	Weight (RE)
Park 2024¹⁹	153/437	103/368	1.25	1.02-1.54	4%
Alali 2025²⁰	24/51	19/51	1.26	0.80-2.00	1.1%
Lagström 2025²¹	236/399	184/395	1.27	1.11-1.45	6.5%
Schauer 2026²²	243/383	225/393	1.11	0.99-1.24	7.3%
Seager 2024 (COLO-DETECT)²³	555/980	477/986	1.17	1.08-1.27	8.7%
EAGLE 2025²⁴	180/417	152/424	1.20	1.02-1.42	5.2%
Hsu 2026²⁵	395/675	363/681	1.10	1.00-1.21	8.3%
Thiruvengadam 2024²⁶	234/550	189/550	1.24	1.06-1.44	5.8%
Maas 2024 (DISCOVERY)²⁷	96/250	93/247	1.02	0.81-1.28	3.6%
Spada 2025 (ACCENDO-Colo)²⁸	290/578	235/580	1.24	1.09-1.41	6.7%
Yabuuchi 2025²⁹	254/501	271/497	0.93	0.83-1.05	7.2%
Xu 2025³⁰	78/195	77/195	1.01	0.79-1.29	3.2%
Jeon 2026³¹	260/497	181/501	1.45	1.25-1.67	6.1%
Dos Santos 2025³²	180/354	164/357	1.11	0.95-1.29	5.8%
Spada 2026 (FIT-AI)³³	342/506	301/503	1.13	1.03-1.24	8.3%
Lux 2026 (EndoMind)³⁴	156/452	152/462	1.05	0.87-1.26	4.7%
Davila-Piñón 2026 (PolyDeep Advance 3)³⁵	250/411	240/416	1.05	0.94-1.18	7.4%
Pooled (random-effects)			1.14	1.09–1.20	100%

CADe, computer-aided detection; CI, confidence interval; RE, random-effects; RR, risk ratio (inverse-variance weights based on REML τ²; see Effect measures and data synthesis). Heterogeneity: Cochran’s Q = 34.65, df = 16, P = 0.0044; I² = 53.8%, τ² = 0.0061 (REML). Knapp-Hartung-adjusted 95% CI: 1.08–1.21. REML, restricted maximum likelihood.

Table 2

Study	CADe (n/N)	Control (n/N)	RR	95% CI	Weight (RE)
Park 2024¹⁹	271/437	191/368	1.19	1.06-1.35	8.1%
Alali 2025²⁰	40/51	29/51	1.38	1.04-1.82	3.0%
Lagström 2025²¹	279/400	222/395	1.24	1.11-1.38	9.0%
Schauer 2026²²	298/383	286/393	1.07	0.99-1.16	10.7%
EAGLE 2025²⁴	251/417	225/424	1.13	1.01-1.28	8.4%
Maas 2024 (DISCOVERY)²⁷	138/250	127/247	1.07	0.91-1.27	6.1%
Xu 2025³⁰	131/195	111/195	1.18	1.01-1.38	6.4%
Jeon 2026³¹	359/497	273/501	1.33	1.20-1.46	9.7%
Dos Santos 2025³²	240/354	225/357	1.08	0.97-1.20	9.1%
Spada 2026 (FIT-AI)³³	401/506	350/503	1.14	1.06-1.23	11.2%
Lux 2026 (EndoMind)³⁴	243/452	252/462	0.99	0.87-1.11	8.3%
Davila-Piñón 2026 (PolyDeep Advance 3)³⁵	286/411	286/416	1.01	0.92-1.11	10.0%
Pooled (random-effects)			1.13	1.07–1.20	100%

CADe, computer-aided detection; CI, confidence interval; RE, random-effects; RR, risk ratio (inverse-variance weights based on REML τ²; see Section 2.8). Heterogeneity: Cochran’s Q = 30.04, df = 11, P = 0.0016; I² = 63.4%, τ² = 0.0054 (REML). Knapp-Hartung-adjusted 95% CI: 1.07–1.20. REML, restricted maximum likelihood.

Small-study effects and potential publication bias were assessed for both outcomes using Egger’s regression test, in which the standardized effect size (log RR divided by its standard error (SE)) was regressed on precision (the inverse of the SE); a regression intercept significantly different from zero was taken as evidence of funnel plot asymmetry.

Sensitivity analyses — leave-one-out

A leave-one-out sensitivity analysis was prespecified for both ADR and PDR. Each study was sequentially excluded, and the pooled random-effects RR, 95% CI, and I² were recalculated to assess the influence of individual studies on the magnitude and statistical significance of the overall effect estimate and on between-study heterogeneity.

In addition to the pre-specified leave-one-out analysis, a post-hoc exploratory sensitivity analysis was conducted following completion of the risk-of-bias assessment, in response to the quasi-randomized, cluster-based allocation design identified in Lagström et al., in which randomization occurred at the center level using alternating 2-week blocks rather than individual patient-level allocation. Because Lagström et al. did not report an intracluster correlation coefficient (ICC) for either outcome, a principled, trial-specific design-effect calculation was not possible a priori. As a conservative bounding approach, the reported variance of this trial’s effect estimate was doubled and both pooled models were re-estimated.

Patient and public involvement

No patients or members of the public were involved in the design, conduct, reporting, or dissemination of this research. It represents a secondary analysis of publicly available data from previously published randomized and quasi-randomized controlled trials.

Results

Study selection

The initial search identified 185 potentially relevant articles. After removal of duplicates (n = 10), 175 records remained for title and abstract screening. Of these, 14 articles were excluded because they involved populations other than adults undergoing screening, surveillance, or diagnostic colonoscopy. 161 full-text articles were assessed for eligibility. Of these, 101 were excluded because they did not use a randomized or quasi-randomized comparative study setting and 39 because outcome data were missing. Ultimately, 21 randomized and quasi-randomized controlled trials met all inclusion criteria and were included in the review, and 17 of these were included in the quantitative synthesis (Fig. 1).

Fig. 1 PRISMA flow diagram for the selection of randomized and quasi-randomized trials included in this review.

PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses; RCT, randomized controlled trial.

Characteristics of included studies

Seventeen studies met inclusion criteria for the ADR outcome, collectively enrolling 15,242 patients.^19-35 Twelve of these studies reported extractable PDR data suitable for pooling (Table 3).^19-35 The included trials were published between 2024 and 2026, comprising four studies from 2024, seven from 2025, and six from 2026, reflecting the temporally restricted evidence base defined a priori for this synthesis. Study designs were predominantly parallel-group randomized controlled trials, with one quasi-randomized cluster design (Lagström 2025).²¹

Table 3

Study	Year	Total N	ADR CADe (events/N)	ADR Control (events/N)	ADR RR (95% CI)	PDR CADe (events/N)	PDR Control (events/N)	PDR RR (95% CI)	AI system
Park 2024¹⁹	2024	805	153/437	103/368	1.25 (1.02–1.54)	271/437	191/368	1.19 (1.06–1.35)	RetinaNet-based
Alali 2025²⁰	2025	102	24/51	19/51	1.26 (0.80–2.00)	40/51	29/51	1.38 (1.04–1.82)	CAD EYE (Fujifilm)
Lagström 2025²¹	2025	795	236/399	184/395	1.27 (1.11–1.45)	279/400	222/395	1.24 (1.11–1.38)	GI Genius (Medtronic)
Schauer 2026²²	2026	776	243/383	225/393	1.11 (0.99–1.24)	298/383	286/393	1.07 (0.99–1.16)	ENDO-AID (Olympus)
Seager 2024 (COLO-DETECT)²³	2024	1966	555/980	477/986	1.17 (1.08–1.27)	—	—	—	GI Genius (Medtronic)
EAGLE 2025²⁴	2025	841	180/417	152/424	1.20 (1.02–1.42)	251/417	225/424	1.13 (1.01–1.28)	CADDIEv1.4 (Odin Medical LTD)
Hsu 2026²⁵	2026	1356	395/675	363/681	1.10 (1.00–1.21)	—	—	—	aetherAI Endo
Thiruvengadam 2024²⁶	2024	1100	234/550	189/550	1.24 (1.06–1.44)	—	—	—	GI Genius (Medtronic)
Maas 2024 (DISCOVERY)²⁷	2024	497	96/250	93/247	1.02 (0.81–1.28)	138/250	127/247	1.07 (0.91–1.27)	DISCOVERY (PENTAX Medical)
Spada 2025 (ACCENDO-Colo)²⁸	2025	1158	290/578	235/580	1.24 (1.09–1.41)	—	—	—	ENDO-AID OIP-1 (Olympus)
Yabuuchi 2025²⁹	2025	998	254/501	271/497	0.93 (0.83–1.05)	—	—	—	EndoBRAIN-AY E v1.2.0 (Cybernet Systems Corp)
Xu 2025³⁰	2025	390	78/195	77/195	1.01 (0.79–1.29)	131/195	111/195	1.18 (1.01–1.38)	Custom-developed CADe Software
Jeon 2026³¹	2026	998	260/497	181/501	1.45 (1.25–1.67)	359/497	273/501	1.33 (1.20–1.46)	ENAD-DET-01 (Ainex Corporation)
Dos Santos 2025³²	2025	711	180/354	164/357	1.11 (0.95–1.29)	240/354	225/357	1.08 (0.97–1.20)	CAD EYE (Fujifilm)
Spada 2026 (FIT-AI)³³	2026	1009	342/506	301/503	1.13 (1.03–1.24)	401/506	350/503	1.14 (1.06–1.23)	GI Genius (Medtronic)
Lux 2026 (EndoMind)³⁴	2026	914	156/452	152/462	1.05 (0.87–1.26)	243/452	252/462	0.99 (0.87–1.11)	EndoMind (Research -developed system, University Hospital Wurzburg Group)
Davila-Piñón 2026 (PolyDeep Advance 3)³⁵	2026	827	250/411	240/416	1.05 (0.94–1.18)	286/411	286/416	1.01 (0.92–1.11)	PolyDeep (Galicia Sur/University of Vigo research group)

ADR, adenoma detection rate; AI, artificial intelligence; CADe, computer-aided detection; CI, confidence interval; PDR, polyp detection rate; RR, risk ratio.

Risk-of-bias assessment using the Cochrane RoB 2 tool classified nine randomized trials as low risk and seven as raising some concerns; the quasi-randomized Lagström study was judged to have serious overall risk of bias using ROBINS-I. Among the randomized trials, concerns were most commonly attributable to the randomization process domain (Park 2024, Xu 2025, and Spada 2026b) or to deviations from intended interventions and missing outcome data (EAGLE 2025, Hsu 2026, Spada 2025/ACCENDO-Colo, Lux 2026/EndoMind). Several trials exhibited notable methodological features warranting mention: Lagström 2025 employed a cluster-randomized quasi-experimental design with an associated denominator imbalance between arms; Park 2024 showed an unequal arm-size allocation; Xu 2025 reported multiple baseline imbalances between groups; Dos Santos 2025 was conducted by a single endoscopist, limiting generalizability; Schauer 2026 was funded by the device manufacturer (Olympus America), which also provided the CADe equipment free of charge; ; and Spada 2026b (FIT-AI) reported comparatively weak allocation concealment. Two studies, Lux 2026 (EndoMind) and Davila-Piñón 2026 (PolyDeep Advance 3), evaluated not-for-profit CADe systems, distinguishing them from the predominantly commercially developed platforms assessed elsewhere in the pool.

Four studies retained for qualitative synthesis only (Codesido-Prado 2026,³⁶ Chang 2026,³⁷ Leggett 2026/DEEP2,³⁸ Robles de la Osa 2026³⁹) were excluded from the quantitative synthesis because raw counts were unextractable or outcome-reporting formats were incompatible.

Risk-of-bias assessment

Risk of bias was assessed using the RoB 2 tool across five domains⁴⁰: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in measurement of the outcome, and bias in selection of the reported result. Judgments were made at the outcome level and categorized as low risk, some concerns, or high risk of bias according to the RoB 2 framework (Table 4).^19,20,22-35

Table 4

Study	D1: Randomization process	D2: Deviations from intended interventions	D3: Missing outcome data	D4: Measurement of the outcome	D5: Selection of the reported result	Overall risk of bias
Park 2024¹⁹	Some concerns	Low risk	Some concerns	Low risk	Low risk	Some concerns
Alali 2025²⁰	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Schauer 2026²²	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Seager 2024 (COLO-DETECT)²³	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
EAGLE 2025²⁴	Low risk	Low risk	Some concerns	Low risk	Low risk	Some concerns
Hsu 2026²⁵	Low risk	Low risk	Some concerns	Low risk	Low risk	Some concerns
Thiruvengadam 2024²⁶	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Maas 2024 (DISCOVERY)²⁷	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Spada 2025 (ACCENDO-Colo)²⁸	Low risk	Some concerns	Some concerns	Low risk	Low risk	Some concerns
Yabuuchi 2025²⁹	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Xu 2025³⁰	Some concerns	Low risk	Low risk	Low risk	Low risk	Some concerns
Jeon 2026³¹	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Dos Santos 2025³²	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk
Spada 2026 (FIT-AI)³³	Some concerns	Low risk	Low risk	Low risk	Low risk	Some concerns
Lux 2026 (EndoMind)³⁴	Low risk	Some concerns	Low risk	Low risk	Low risk	Some concerns
Davila-Piñón 2026 (PolyDeep Advance 3)³⁵	Low risk	Low risk	Low risk	Low risk	Low risk	Low risk

Overall, the methodological quality of the included trials was considered acceptable. No trial was judged to be at overall high risk of bias based on the available information. However, several studies raised specific methodological concerns, principally related to post-randomization exclusions, quasi-randomized allocation, baseline imbalance, or allocation concealment.

Park et al.¹⁹ was judged to have some concerns, primarily because of post-randomization imbalance between study arms associated with differential withdrawal, particularly in the control group. Differential post-randomization withdrawals between study arms raised some concerns in the domain of missing outcome data, as the extent and imbalance of exclusions could potentially be related to the outcome.

Funding and manufacturer involvement were also considered as potential sources of bias. For Schauer et al.,²² the reported industry involvement was evaluated in relation to the RoB 2 domains, particularly the selection of the reported result and other aspects of trial conduct and reporting. No evidence was identified that industry involvement influenced the randomization process, deviations from intended interventions, outcome assessment, or selective reporting; therefore, industry sponsorship alone did not warrant upgrading the RoB 2 judgment.

Because Lagström et al.²¹ used quasi-randomized allocation based on alternating 2-week intervention periods rather than individual randomization, risk of bias for this study was assessed separately using the ROBINS-I tool (Table 5). For ADR, the study was judged to have serious overall risk of bias, primarily because the time-block allocation could introduce confounding by calendar period, center, endoscopist characteristics, and changes in patient case-mix over time. Additional moderate concerns related to post-allocation exclusions and incomplete outcome data. The measurement of ADR was considered at low risk of bias because adenoma status was confirmed histologically.

Table 5

Study	D1 Confounding	D2 Selection	D3 Classification	D4 Deviations	D5 Missing data	D6 Measurement	D7 Reported result	Overall
Lagström 2025²¹	Serious	Low	Low	Moderate	Moderate	Low	Low	Serious

ROBINS-I, Risk Of Bias In Non-randomized Studies of Interventions.

EAGLE and Hsu et al.²⁵ were considered to have some concerns related principally to missing outcome data and post-randomization exclusions. In EAGLE, 973 participants were randomized, whereas 841 were included in the final analyzed population; although the exclusions were reported with specific reasons, the extent of post-randomization attrition warranted a cautious judgment. In the Hsu trial, 1,457 participants were randomized and 1,356 were included in the final analysis, leading to a similar concern regarding outcome completeness. EAGLE used block randomization with electronic allocation procedures and blinded pathological assessment, whereas the Hsu trial used a computer-generated 1:1 randomization sequence and intention-to-treat analysis.

Spada et al.²⁸ (ACCENDO-Colo) 2025 was judged to have some concerns because of post-randomization exclusions and the open-label design. Although randomization was performed using a central computer-generated sequence, 70 of 1,228 randomized participants were subsequently excluded from the mITT analysis.

Xu et al.³⁰ was judged to have some concerns because of multiple baseline imbalances between study groups. Spada et al.³³ (FIT-AI) 2026 was also classified as having some concerns because of limitations in allocation concealment. Lux et al.³⁴ was judged to have some concerns because endoscopists could not be blinded to CADe allocation. This limitation is inherent to CADe-assisted colonoscopy and was not considered sufficient to indicate high risk of bias by itself. Importantly, the Lux trial used centralized randomization and was conducted as a multicenter randomized controlled study.

The remaining trials were judged to have low overall risk of bias based on the available methodological information. In particular, the use of open-label designs was not automatically interpreted as high risk because blinding of endoscopists is generally infeasible in CADe-assisted colonoscopy. Where outcome ascertainment relied on histopathological confirmation, blinded pathological assessment further reduced the potential for bias in outcome classification.

Manufacturer involvement or industry-provided equipment was considered separately as a potential source of concern but was not, in the absence of evidence of influence on trial conduct, analysis, or reporting, sufficient by itself to warrant a higher RoB 2 judgment. Accordingly, the manufacturer-supported Schauer trial was not classified as having increased RoB 2 solely on the basis of industry involvement.

Overall, the risk-of-bias assessment indicated that the contemporary evidence base was predominantly of acceptable methodological quality, with some concerns concentrated in specific aspects of randomization, allocation concealment, post-randomization exclusions, baseline comparability, and unavoidable lack of blinding in procedural CADe trials.

Among the 17 trials included in the quantitative ADR analysis, 9 (52.9%) were judged to be at low overall risk of bias, 7 (41.2%) to have some concerns, and 1 (5.9%) to be at serious overall risk of bias. The latter was the quasi-randomized study by Lagström et al.,²¹ which was assessed separately using ROBINS-I because of its time-block allocation design. The main methodological concerns across the included studies related to quasi-randomized allocation and potential confounding, post-randomization exclusions or incomplete outcome data, baseline imbalances, limitations in allocation concealment, and the inability to blind endoscopists to CADe allocation.

Publication bias assessment

Egger’s regression test showed no evidence of significant small-study effects for either outcome. For ADR (k = 17), the regression intercept was 0.40 (SE = 1.12; 95% CI −1.98 to 2.78), corresponding to t(15) = 0.36, P = 0.725. For PDR (k = 12), the intercept was 1.14 (SE = 1.71; 95% CI −2.66 to 4.94), corresponding to t(10) = 0.67, P = 0.520. These non-significant intercepts, together with the visual symmetry of the funnel plots (Figs. 2 and 3), do not support the presence of substantial publication bias or small-study effects in either pooled analysis, although the limited number of included trials — particularly for PDR — constrains the statistical power of this assessment.

Fig. 2 Funnel plot for adenoma detection rate (ADR).

CADe, computer-aided detection; RoB, risk of bias; RR, risk ratio; SE, standard error.

Fig. 3 Funnel plot for polyp detection rate (PDR).

CADe, computer-aided detection; RoB, risk of bias; RR, risk ratio; SE, standard error.

Effect of CADe on ADR

Pooling of all seventeen trials (k = 17) using a random-effects model with REML estimation of τ² showed that CADe-assisted colonoscopy significantly increased ADR compared with conventional colonoscopy (RR 1.14, 95% CI 1.09–1.20, P < 0.0001; Fig. 4). This pool showed moderate statistical heterogeneity (Cochran’s Q = 34.65, df = 16, P = 0.0044; I² = 53.8%, τ² = 0.0061). A Knapp-Hartung adjustment yielded a similar 95% CI (1.08–1.21).

Fig. 4 Forest plot – Adenoma Detection Rate (ADR).

CI, confidence interval; REML, restricted maximum likelihood.

Effect of CADe on PDR

Twelve trials (k = 12) reported PDR as an extractable dichotomous proportion. Pooling of the remaining twelve trials using a random-effects model with REML estimation of τ² showed that CADe significantly increased PDR (RR 1.13, 95% CI 1.07–1.20, P < 0.0001; Fig. 5). As with ADR, this estimate showed moderate-to-substantial statistical heterogeneity (Cochran’s Q = 30.04, df = 11, P = 0.0016; I² = 63.4%, τ² = 0.0054), explored further through the prespecified leave-one-out sensitivity analysis described below. A Knapp-Hartung adjustment yielded an unchanged 95% CI (1.07–1.20).

Fig. 5 Forest plot for polyp detection rate (PDR).

CI, confidence interval; REML, restricted maximum likelihood.

Sensitivity analysis: leave-one-out

A leave-one-out sensitivity analysis was performed for both outcomes to assess the influence of individual studies on the pooled estimate and on observed heterogeneity.

The ADR pooled estimate proved robust in direction and approximate magnitude: exclusion of any single trial produced a pooled RR within the narrow range of 1.12–1.16, remaining statistically significant in every iteration. Excluding Yabuuchi 2025 or Jeon 2026 produced the largest reductions in I² (to 33.2% and 35.6%, respectively), but heterogeneity remained moderate in both cases and stayed at 53–57% following exclusion of every other individual trial. This indicates that the heterogeneity observed in the expanded ADR pool does not stem from a single outlying study but instead reflects broader, multi-source between-study variance.

Excluding Jeon 2026 produced the largest reduction in I² (from 63.4% to 44.5%), but heterogeneity remained moderate rather than being eliminated, and removal of any other individual trial left I² at 58–67%. This indicates that the heterogeneity in the PDR pooled estimate reflects distributed, multi-source variance across the evidence base rather than a single influential trial — consistent with the pattern observed for ADR.

As a post-hoc exploratory analysis addressing the cluster-based allocation design of Lagström et al., the sensitivity analysis doubled the variance of Lagström et al. to account for its quasi-randomized, cluster-based design; this left both pooled estimates essentially unchanged (ADR: RR 1.14, 95% CI 1.08–1.20, I² = 52.0%; PDR: RR 1.13, 95% CI 1.07–1.19, I² = 61.3%).

Discussion

This meta-analysis of seventeen RCTs and quasi-randomized trials, comprising 15,242 participants contributed to the ADR analysis (7,636 CADe and 7,606 control), demonstrated that CADe-assisted colonoscopy significantly increased both the adenoma detection rate (RR 1.14, 95% CI 1.09–1.20) and the polyp detection rate (RR 1.13, 95% CI 1.07–1.20; k = 12, 8,665 patients) compared with conventional colonoscopy. Both outcomes showed moderate-to-substantial statistical heterogeneity (I² = 53.8% for ADR and I² = 63.4% for PDR); a prespecified leave-one-out sensitivity analysis, detailed below, indicated that this heterogeneity reflects distributed, multi-source between-study variance rather than the influence of any single trial.

Interpretation and comparison with existing evidence

These findings are broadly consistent with previously published meta-analyses evaluating AI-assisted colonoscopy. Shiha et al.⁴¹ reported a pooled ADR benefit of similar magnitude across 12 RCTs and 11,340 patients, while Hassan et al.,⁴² in a synthesis of 21 trials, likewise confirmed a consistent ADR increase with CADe (44.0% vs. 35.9%). More recent, larger syntheses corroborate both the magnitude and the heterogeneity pattern observed here: Tan et al.,¹⁵ pooling 30 studies, reported a comparable ADR effect (RR 1.20, 95% CI 1.14–1.26) alongside substantial heterogeneity (I² = 76.6%), and Sultany et al.,⁴³ in a 12-trial synthesis, similarly reported high heterogeneity for both ADR (I² = 81%) and PDR (I² = 79%). The pooled effect size observed in the present analysis therefore falls within the range reported by these larger contemporary syntheses, reinforcing the robustness of the association despite the increase in observed heterogeneity. The present analysis is based exclusively on trials published in 2024–2026, offering an updated perspective spanning an expanded and geographically diverse set of trial settings and incorporating newer-generation, as well as nonprofit and academically developed, CADe platforms not evaluated in earlier syntheses.

Heterogeneity interpretation

The pool showed moderate-to-substantial heterogeneity for both ADR (I² = 53.8%) and PDR (I² = 63.4%). Leave-one-out sensitivity analysis (Tables 6 and 7) demonstrated that no single trial was responsible^19-35: the pooled RR remained within a narrow range (ADR 1.12–1.16; PDR 1.11–1.15) regardless of which study was excluded, and I² fell substantially only after removing Yabuuchi et al.²⁹ or Jeon et al.³¹ from the ADR pool (to 33.2% and 35.6%, respectively) or Jeon et al.³¹ from the PDR pool (to 44.5%), never approaching the near-zero values; this indicates that heterogeneity in the current evidence base is distributed across multiple studies rather than attributable to one outlying trial.

Table 6

Study excluded	RR	95% CI	I²
None (full model)	1.14	1.09–1.20	53.8%
Park 2024¹⁹	1.14	1.08–1.20	55.7%
Alali 2025²⁰	1.14	1.08–1.20	56.5%
Lagström 2025²¹	1.13	1.08–1.19	53.0%
Schauer 2026²²	1.15	1.09–1.21	56.4%
Seager 2024²³	1.14	1.08–1.20	56.1%
EAGLE 2025²⁴	1.14	1.08–1.20	56.1%
Hsu 2026²⁵	1.15	1.09–1.21	55.9%
Thiruvengadam 2024²⁶	1.14	1.08–1.20	55.1%
Maas 2024²⁷	1.15	1.09–1.21	55.5%
Spada 2025²⁸	1.14	1.08–1.20	54.4%
Yabuuchi 2025²⁹	1.16	1.11–1.21	33.2%
Xu 2025³⁰	1.15	1.09–1.21	55.6%
Jeon 2026³¹	1.12	1.08–1.17	35.6%
Dos Santos 2025³²	1.14	1.09–1.21	56.5%
Spada 2026³³	1.14	1.08–1.21	56.7%
Lux 2026³⁴	1.15	1.09–1.21	55.7%
Davila-Piñón 2026³⁵	1.15	1.09–1.21	54.2%

CI, confidence interval; RR, risk ratio.

Table 7

Study excluded	RR	95% CI	I²
None (full model)	1.13	1.07–1.20	63.4%
Park 2024¹⁹	1.13	1.06–1.19	65.7%
Alali 2025²⁰	1.13	1.07–1.19	64.3%
Lagström 2025²¹	1.12	1.06–1.19	62.6%
Schauer 2026²²	1.14	1.08–1.21	64.4%
EAGLE 2025²⁴	1.13	1.07–1.20	66.7%
Maas 2024²⁷	1.14	1.07–1.20	66.3%
Xu 2025³⁰	1.13	1.07–1.20	66.3%
Jeon 2026³¹	1.11	1.06–1.16	44.5%
Dos Santos 2025³²	1.14	1.07–1.21	65.8%
Spada 2026³³	1.13	1.06–1.21	66.6%
Lux 2026³⁴	1.15	1.09–1.21	59.8%
Davila-Piñón 2026³⁵	1.15	1.09–1.21	58.3%

CI, confidence interval; RR, risk ratio.

A plausible qualitative explanation lies in the technological diversity of the included trials: the seventeen studies employed eleven distinct CADe systems, ranging from established commercial platforms (GI Genius, ENDO-AID, CAD EYE) to newer commercial systems and, in two cases, non-commercial, academically developed platforms (EndoMind, PolyDeep). Because most individual platforms in the present pool are represented by only one or two trials, a formal platform-based subgroup analysis was not adequately powered and was not undertaken.

Clinical implications

Both ADR and PDR showed a comparable degree of heterogeneity, and both outcomes should therefore be interpreted with similar caution regarding generalizability across clinical settings and platforms. ADR nonetheless retains particular clinical importance because, unlike PDR, it specifically captures the detection of adenomatous (precancerous) lesions and is the outcome most directly linked to interval cancer risk and mortality,^6,44 whereas PDR reflects detection of any polyp, including non-neoplastic lesions. The consistency of the point estimate throughout the leave-one-out sensitivity analysis - despite the statistical heterogeneity - continues to support a genuine, clinically meaningful average benefit of CADe on both quality indicators.

Limitations

Several limitations should be acknowledged. First, although this meta-analysis includes seventeen trials, comparable in scale to several recently published meta-analyses, the evidence base remains smaller than the largest available syntheses, and statistical power for subgroup or platform-specific analyses remains limited; consequently, only the leave-one-out sensitivity analysis, the sole approach prespecified in the protocol, was performed. Second, the included trials employed eleven different CADe systems, nine of which were represented by only a single trial; this fragmentation substantially limits the generalizability of the pooled estimate to any single commercial or research-developed platform and precluded a formally powered platform-based subgroup or meta-regression analysis. Third, one trial (Lagström et al.²¹) employed a quasi-randomized, cluster-based allocation design, introducing a potential risk of temporal selection bias despite stable referral criteria across periods. Because this trial did not report an intracluster correlation coefficient, its contribution to the primary pooled estimates could not be adjusted for clustering using a principled, trial-specific design effect; it was therefore included using its unadjusted patient-level event counts, consistent with the other included trials. A post-hoc sensitivity analysis doubling this trial’s variance as a conservative bounding approximation left both pooled estimates materially unchanged, but this adjustment was exploratory and not based on an empirically derived ICC, and should be interpreted as a robustness check rather than a definitive correction for the clustering effect. Fourth, PDR could not be pooled for five of the seventeen trials (Seager et al., Hsu et al., Thiruvengadam et al., Spada et al. [ACCENDO-Colo], and Yabuuchi et al.), as this outcome was not reported as a standalone proportion in the original publications, reducing the precision of the secondary outcome estimate relative to ADR. Fifth, and most notably, the substantial statistical heterogeneity observed for both outcomes could not be fully explained through the prespecified leave-one-out sensitivity analysis; residual heterogeneity of this magnitude, reflected in prediction intervals crossing the null for both outcomes, warrants caution in extrapolating the pooled estimates to individual clinical settings. Future updates of this synthesis, incorporating a larger number of trials per CADe platform, should consider preregistered subgroup analyses or meta-regression by platform, risk of bias, and publication year to formally test these as potential moderators. Finally, this meta-analysis was restricted to ADR and PDR as detection-based quality indicators; although the well-established inverse relationship between ADR and interval CRC incidence supports the clinical relevance of these surrogate outcomes,^6,44 none of the included trials were designed or powered to evaluate downstream clinical endpoints such as interval cancer incidence or long-term mortality, which remain the ultimate measures of clinical benefit.

Conclusions

This meta-analysis supports an average improvement in adenoma and polyp detection with AI-assisted colonoscopy; however, moderate-to-substantial heterogeneity and variability across settings and platforms warrant cautious interpretation rather than an unqualified recommendation for routine adoption. Therefore, these findings suggest that the beneficial effect of CADe observed in earlier evidence syntheses persists in the contemporary evidence base, but they do not establish a uniform treatment effect across all systems or clinical settings. Future research should focus on evaluating the impact of CADe on longer-term clinical outcomes, including interval colorectal cancer incidence and cancer-related mortality, as well as on cost-effectiveness and comparative performance across the expanding range of commercial and research-developed platforms.

Supporting information

Supplementary material for this article is available at https://doi.org/10.14218/CSP.2026.00015 .

Supplementary Table. 1

Database-specific search strategies

(DOCX)

Click here for additional data file.

Declarations

Acknowledgments

During the preparation of this work, the authors used AI-based tools (OpenAI GPT-4 and Anthropic Claude) to assist with language editing and with organizing and drafting text related to data extraction. These tools were not used to perform, generate, or verify statistical analyses; all statistical computations were conducted independently by the authors using the metafor package in R. After using these tools, the authors reviewed and edited all content and take full responsibility for the content of this publication.

Funding

This study received no external funding.

Conflict of interest

The authors declare no conflicts of interest relevant to this publication.

Author contributions

Investigation (GA, VM), project administration (GA), formal analysis (GA), conceptualization (GA), validation (VM), and writing—review and editing (GA, VM). All authors have made a significant contribution to this study and have approved the final manuscript.

Ethical statement

Not applicable.

Data sharing statement

The original contributions presented in this study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

References

1	Sninsky JA, Shore BM, Lupu GV, Crockett SD. Risk Factors for Colorectal Polyps and Cancer. Gastrointest Endosc Clin N Am 2022;32(2):195–213 View Article PubMed/NCBI

2	Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74(3):229–263 View Article PubMed/NCBI

3	Giovannucci E, Martínez ME. Tobacco, colorectal cancer, and adenomas: a review of the evidence. J Natl Cancer Inst 1996;88(23):1717–1730 View Article PubMed/NCBI

4	Hagen R, Srivastava A, Anderson JC. The serrated pathway and colorectal cancer: what the gastroenterologist should know. Expert Rev Gastroenterol Hepatol 2025;19(6):593–606 View Article PubMed/NCBI

5	Zauber AG, Winawer SJ, O'Brien MJ, Lansdorp-Vogelaar I, van Ballegooijen M, Hankey BF, et al. Colonoscopic polypectomy and long-term prevention of colorectal-cancer deaths. N Engl J Med 2012;366(8):687–696 View Article PubMed/NCBI

6	Corley DA, Jensen CD, Marks AR, Zhao WK, Lee JK, Doubeni CA, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med 2014;370(14):1298–1306 View Article PubMed/NCBI

7	Adler A, Aminalai A, Aschenbeck J, Drossel R, Mayr M, Scheel M, et al. Latest generation, wide-angle, high-definition colonoscopes increase adenoma detection rate. Clin Gastroenterol Hepatol 2012;10(2):155–159 View Article PubMed/NCBI

8	Zhao S, Wang S, Pan P, Xia T, Chang X, Yang X, et al. Magnitude, Risk Factors, and Factors Associated With Adenoma Miss Rate of Tandem Colonoscopy: A Systematic Review and Meta-analysis. Gastroenterology 2019;156(6):1661–1674.e11 View Article PubMed/NCBI

9	Ahn SB, Han DS, Bae JH, Byun TJ, Kim JP, Eun CS. The Miss Rate for Colorectal Adenoma Determined by Quality-Adjusted, Back-to-Back Colonoscopies. Gut Liver 2012;6(1):64–70 View Article PubMed/NCBI

10	Kröner PT, Engels MM, Glicksberg BS, Johnson KW, Mzaik O, van Hooft JE, et al. Artificial intelligence in gastroenterology: A state-of-the-art review. World J Gastroenterol 2021;27(40):6794–6824 View Article PubMed/NCBI

11	Parsa N, Byrne MF. Artificial intelligence for identification and characterization of colonic polyps. Ther Adv Gastrointest Endosc 2021;14:26317745211014698 View Article PubMed/NCBI

12	Makar J, Abdelmalak J, Con D, Hafeez B, Garg M. Use of artificial intelligence improves colonoscopy performance in adenoma detection: a systematic review and meta-analysis. Gastrointest Endosc 2025;101(1):68–81.e8 View Article PubMed/NCBI

13	Soleymanjahi S, Huebner J, Elmansy L, Rajashekar N, Lüdtke N, Paracha R, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med 2024;177(12):1652–1663 View Article PubMed/NCBI

14	Eman FNU, Gyaneshwari FNU, Kumari R, Jabeen A, Kumari K, Mansha FNU, et al. Artificial Intelligence-Driven Colonoscopy: A Systematic Review and Network Meta-Analysis on System Performance for Colorectal Neoplasia Detection. JGH Open 2026;10(3):e70372 View Article PubMed/NCBI

15	Tan S, Zeng P, Liu S, Yang Y, Chen S, Zhang W, et al. Effectiveness of artificial intelligence-assisted colonoscopy in detecting and diagnosing colorectal tumors: a systematic review and network meta-analysis. Int J Colorectal Dis 2025;40(1):218 View Article PubMed/NCBI

16	Viechtbauer W. Conducting Meta-Analyses in R with the metafor Package. J Stat Softw 2010;36(3):1–48 View Article

17	Röver C, Knapp G, Friede T. Hartung-Knapp-Sidik-Jonkman approach and its modification for random-effects meta-analysis with few studies. BMC Med Res Methodol 2015;15:99 View Article PubMed/NCBI

18	Higgins JP, Thompson SG, Spiegelhalter DJ. A re-evaluation of random-effects meta-analysis. J R Stat Soc Ser A Stat Soc 2009;172(1):137–159 View Article PubMed/NCBI

19	Park DK, Kim EJ, Im JP, Lim H, Lim YJ, Byeon JS, et al. A prospective multicenter randomized controlled trial on artificial intelligence assisted colonoscopy for enhanced polyp detection. Sci Rep 2024;14(1):25453 View Article PubMed/NCBI

20	Alali AA, Alhashmi A, Alotaibi N, Ali N, Alali M, Alfadhli A. Artificial Intelligence for Adenoma and Polyp Detection During Screening and Surveillance Colonoscopy: A Randomized-Controlled Trial. J Clin Med 2025;14(2):581 View Article PubMed/NCBI

21	Lagström RMB, Bräuner KB, Bielik J, Rosen AW, Crone JG, Gögenur I, et al. Improvement in adenoma detection rate by artificial intelligence-assisted colonoscopy: Multicenter quasi-randomized controlled trial. Endosc Int Open 2025;13:a25215169 View Article PubMed/NCBI

22	Schauer C, van Rijnsoever M, Jafer A, Walmsley R, Wang MTM, Atkinson N. Man Plus Machine: A Randomized Control Trial of Artificial Intelligence Including the Impact of Adjunctive Polyp Detection Techniques. J Gastroenterol Hepatol 2026;41(1):117–127 View Article PubMed/NCBI

Seager A, Sharp L, Neilson LJ, Brand A, Hampton JS, Lee TJW, et al. Polyp detection with colonoscopy assisted by the GI Genius artificial intelligence endoscopy module compared with standard colonoscopy in routine colonoscopy practice (COLO-DETECT): a multicentre, open-label, parallel-arm, pragmatic randomised controlled trial. Lancet Gastroenterol Hepatol 2024;9(10):911–923 View Article PubMed/NCBI

24	Kader R, Hassan C, Lanas Á, Romańczyk M, Romańczyk T, Kotowski B, et al. A novel cloud-based artificial intelligence for real-time detection of colorectal neoplasia - a randomized controlled trial (EAGLE). NPJ Digit Med 2025;9(1):84 View Article PubMed/NCBI

25	Hsu WF, Kuo CY, Yen HH, Lin YM, Chen YN, Sun CK, et al. Computer-Assisted Colonoscopy in High-Adenoma Detection Rate Settings in a High-Risk Population: A Randomized Clinical Trial. JAMA Netw Open 2026;9(4):e264881 View Article PubMed/NCBI

Thiruvengadam NR, Solaimani P, Shrestha M, Buller S, Carson R, Reyes-Garcia B, et al. The Efficacy of Real-time Computer-aided Detection of Colonic Neoplasia in Community Practice: A Pragmatic Randomized Controlled Trial. Clin Gastroenterol Hepatol 2024;22(11):2221–2230.e15 View Article PubMed/NCBI

27	Maas MHJ, Rath T, Spada C, Soons E, Forbes N, Kashin S, et al. A computer-aided detection system in the everyday setting of diagnostic, screening, and surveillance colonoscopy: an international, randomized trial. Endoscopy 2024;56(11):843–850 View Article PubMed/NCBI

28	Spada C, Salvi D, Ferrari C, Hassan C, Barbaro F, Belluardo N, et al. A comprehensive RCT in screening, surveillance, and diagnostic AI-assisted colonoscopies (ACCENDO-Colo study). Dig Liver Dis 2025;57(3):762–769 View Article PubMed/NCBI

29	Yabuuchi Y, Hosotani K, Morihisa Y, Fujio Y, Oshikawa D, Oshita M, et al. Effect of Computer-Aided Detection During Colonoscopy on Adenoma Detection Rate in a Community Hospital Setting: Randomized Controlled Trial. Dig Endosc 2025;37(11):1215–1223 View Article PubMed/NCBI

30	Xu X, Ba L, Lin L, Song Y, Zhao C, Yao S, et al. Evaluation efficacy and accuracy of a real-time computer-aided polyp detection system during colonoscopy: a prospective, multicentric, randomized, parallel-controlled study trial. Surg Endosc 2025;39(11):7417–7427 View Article PubMed/NCBI

31	Jeon HJ, Keum B, Jeong ES, Kim SE, Moon CM, Lee B, et al. Clinical Efficacy of Real-Time Artificial Intelligence-Assisted Colonoscopy in Colorectal Polyp Detection: A Prospective Multicenter Randomized Controlled Trial. Gut Liver 2026;20(1):97–106 View Article PubMed/NCBI

32	Dos Santos CEO, Leggett C, Sharma P, Dos Santos GM, Sanmartin IDA, Pereira-Lima JC. White light imaging versus artificial intelligence-assisted white light imaging for colorectal neoplasia detection: a randomised trial. Rev Gastroenterol Peru 2025;45(4):359–366 View Article PubMed/NCBI

Spada C, Cesaro P, Fuccio L, Salvi D, Ferrari C, Barbaro F, et al. Impact of Artificial Intelligence for Detection of Precancerous Colonic Lesions in a Fecal Immunochemical Blood Test-Based Organized Screening Program in Italy: A Randomized Control Trial. United European Gastroenterol J 2026;14(1):e70176 View Article PubMed/NCBI

34	Lux TJ, Saßmannshausen Z, Kafetzis I, Banck M, Krenzer A, Fitting D, et al. Artificial intelligence assisted colorectal lesion detection in private practices a randomized controlled study. NPJ Digit Med 2026;9(1):284 View Article PubMed/NCBI

Davila-Piñón P, Díez-Martín AI, Nogueira-Rodríguez A, Fdez-Riverola F, Glez-Peña D, Reboiro-Jato M, et al. Computer-assisted versus standard colonoscopy for adenoma detection in a population-based colorectal cancer screening program: a randomized clinical trial. Endoscopy 2026 View Article PubMed/NCBI

36	Codesido-Prado L, Davila-Piñón P, Almazán R, De Castro L, Couto-Worner I, Mejuto R, et al. Computer-aided detection in surveillance colonoscopy: a population-based randomized trial. Endoscopy 2026 View Article PubMed/NCBI

37	Chang PW, Nguyen DD, Kong N, Wang D, Wang S, Ong J, et al. Impact of artificial intelligence-assisted colonoscopy on gastroenterology fellow performance: a pragmatic randomized controlled trial. Gastrointest Endosc 2026;103(5):1043–1051.e3 View Article PubMed/NCBI

Leggett CL, Plowman RS, Surace L, Gorospe E, Lachter J, Ben-Ami Shor D, et al. Computer-aided polyp detection multicenter international randomized controlled study with a focus on community clinics: Gastroenterology Artificial INtelligence system for detecting colorectal polyps. Gastrointest Endosc 2026;104(2):277–285.e3 View Article PubMed/NCBI

Robles de la Osa D, Santos Fernández J, Pérez Urra C, Espinel Pinedo P, Bulnes Labrador CB, Martín Ibáñez C, et al. Efficacy of an artificial intelligence system for lesion detection and characterization (CADe and CADx) during colonoscopy following positive faecal immunochemical test in a colorectal cancer screening programme: A randomized clinical trial. Colorectal Dis 2026;28(3):e70426 View Article PubMed/NCBI

40	Sterne JAC, Savović J, Page MJ, Elbers RG, Blencowe NS, Boutron I, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ 2019;366:l4898 View Article PubMed/NCBI

41	Shiha MG, Oka P, Raju SA, David Tai FW, Ching HL, Thoufeeq M, et al. Artificial intelligence-assisted colonoscopy for adenoma and polyp detection: an updated systematic review and meta-analysis. IGIE 2023;2(3):333–343.e8 View Article PubMed/NCBI

42	Hassan C, Spadaccini M, Mori Y, Foroutan F, Facciorusso A, Gkolfakis P, et al. Real-Time Computer-Aided Detection of Colorectal Neoplasia During Colonoscopy : A Systematic Review and Meta-analysis. Ann Intern Med 2023;176(9):1209–1220 View Article PubMed/NCBI

43	Sultany A, Chikatimalla R, Rao A, Omar MA, Shaar A, Ali H, et al. Real-Time Artificial Intelligence Versus Standard Colonoscopy in the Early Detection of Colorectal Cancer: A Systematic Review and Meta-Analysis. Healthcare (Basel) 2025;13(19):2517 View Article PubMed/NCBI

44	Pilonis ND, Spychalski P, Kalager M, Løberg M, Wieszczy P, Didkowska J, et al. Adenoma Detection Rates by Physicians and Subsequent Colorectal Cancer Risk. JAMA 2025;333(5):400–407 View Article PubMed/NCBI

Copyright © 2026 Author(s). This is an Open Access article distributed under the terms of the Creative Commons Attribution-Noncommercial 4.0 License (CC BY-NC 4.0), permitting all non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

About this Article

Cite this article

Adinolfi G, Milia V. Artificial Intelligence-assisted Versus Conventional Colonoscopy for Adenoma and Polyp Detection Rates: A Meta-analysis of Recent Randomized and Quasi-randomized Trials. Cancer Screen Prev. Published online: Sep 29, 2026. doi: 10.14218/CSP.2026.00015.

Copy

Export to RIS

Export to EndNote

Article History

Received	Revised	Accepted	Published
July 21, 2026	August 24, 2026	September 11, 2026	September 29, 2026

DOI http://dx.doi.org/10.14218/CSP.2026.00015

Cancer Screening and Prevention
pISSN 2993-6314
eISSN 2835-3315

Next Article >

PDF

Full Article

28 Article Accesses	Citation counts are provided from Dimensions. The counts may vary by service, and are reliant on the availability of their data. Counts will update daily once available.
8 PDF Download

Publications > Journals > Cancer Screening and Prevention> Article Full Text

Artificial Intelligence-assisted Versus Conventional Colonoscopy for Adenoma and Polyp Detection Rates: A Meta-analysis of Recent Randomized and Quasi-randomized Trials

Abstract

Background and objectives

Methods

Results

Conclusions

Keywords

Introduction

Materials and methods

Study design and rationale

Eligibility criteria (PICOS)

Publication date and search window

Information sources and search strategy

Study selection process

Data extraction

Risk of bias assessment

Effect measures and data synthesis

Sensitivity analyses — leave-one-out

Patient and public involvement

Results

Study selection

Characteristics of included studies

Risk-of-bias assessment

Publication bias assessment

Effect of CADe on ADR

Effect of CADe on PDR

Sensitivity analysis: leave-one-out

Discussion

Interpretation and comparison with existing evidence

Heterogeneity interpretation

Clinical implications

Limitations

Conclusions

Supporting information

Supplementary Table. 1

Declarations

Acknowledgments

Funding

Conflict of interest

Author contributions

Ethical statement

Data sharing statement

References

About this Article

Table of Contents

Artificial Intelligence-assisted Versus Conventional Colonoscopy for Adenoma and Polyp Detection Rates: A Meta-analysis of Recent Randomized and Quasi-randomized Trials