CoChat
Sign Up
Cited by 0Colorectal Cancer Screening and DetectionAI in cancer detectionArtificial Intelligence in Healthcare and Education

Structured Integration of an Artificial Intelligence-Based System for the Optical Diagnosis of Colorectal Polyps

Open access
Kang, Hae Yeon, Kang, Soonwhan, Chung, Goh Eun, Baek, Dong Hoon, Lee, Hong Sub, Park, Jinbae, Yang, Sun Young, Lim, Seon Hee +3 more·2026·Gut and Liver
DOI 10.5009/gnl250331PMC PMC12800665
Open this paper in CoChat — free accountOpen PDF

Abstract

Background/Aims: Recent advances in computer-aided diagnosis (CADx) systems have demonstrated expert-level accuracy in the optical diagnosis of colorectal polyps. High-confidence (HC) diagnoses have been defined as those made within 3 seconds without hesitation, and these systems have been shown to improve diagnostic accuracy. We aimed to evaluate the performance of endoscopists with varying levels of experience in diagnosing colorectal polyps with the assistance of a new CADx system applying the 3-second rule and without artificial intelligence assistance. Methods: study, 35 endoscopists assessed 100 polyps (51 adenomas, 39 hyperplastic polyps, 10 sessile serrated lesions) using narrow-band imaging video clips on an online platform. Assessments consisted of individual endoscopist diagnosis and CADx-assisted diagnosis. HC assignments followed the 3-second rule in both phases. Performance metrics included HC accuracy, HC rate, and adherence to the Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) and Simple Optical Diagnosis Accuracy (SODA) thresholds. Results: HC diagnostic accuracy improved from 78.3% (95% confidence interval [CI], 76.6% to 80.0%) to 89.8% (95% CI, 88.6% to 90.9%) with CADx assistance (p<0.001). The proportion of HC predictions increased from 64.2% to 75.4% (p<0.001). Novice endoscopists showed marked improvement with CADx (74.1% vs 88.8%; p<0.001). CADx-assisted diagnoses nearly met SODA and PIVI thresholds under the 3-second rule. Additional analysis demonstrated that CADx assistance significantly improved interobserver agreement and ground truth, particularly for novices (κ=0.37 to κ=0.65; p<0.001). Conclusions: Integrating CADx with the 3-second rule significantly enhances the performance of endoscopists in the optical diagnosis of colorectal polyps, with the greatest benefit observed among novice endoscopists.

AI snapshot

Written from the paper’s full text.

TL;DR

In a multicenter ex vivo study, adding a CADx system to the 3-second rule improved endoscopists’ high-confidence optical diagnosis of colorectal polyps and reduced decision time. The largest benefit was seen among novice endoscopists.

Key findings

  • CADx assistance increased high-confidence diagnostic accuracy from 78.3% (95% CI, 76.6%–80.0%) to 89.8% (95% CI, 88.6%–90.9%; p<0.001) and high-confidence prediction rate from 64.2% to 75.4% (p<0.001); overall accuracy rose from 74.4% to 85.6% (p<0.001), and mean decision time fell from 3.3 to 2.8 seconds (p<0.001).
  • Standalone CADx achieved 90.0% accuracy (95% CI, 89.0%–91.0%), 95.1% sensitivity, 82.1% specificity, 89.2% PPV, and 91.4% NPV.
  • Novices improved most with CADx: high-confidence accuracy 74.1% vs 88.8% (p<0.001); overall agreement with ground truth improved from κ=0.44 to κ=0.69, with novices improving from κ=0.37 to κ=0.65 (p<0.001).
  • CADx-assisted diagnoses nearly met SODA and PIVI thresholds under the 3-second rule; all groups achieved the PIVI NPV threshold with CADx, but novices did not reach the SODA “resect-and-discard” specificity threshold even with CADx.

Methods

  • Multicenter ex vivo study: 35 endoscopists (6 trained experts, 11 experts, 18 novices) assessed 100 polyps (51 adenomas, 39 hyperplastic polyps, 10 sessile serrated lesions) using 10-second NBI video clips on an online platform, in endoscopist-alone and CADx-assisted phases separated by a 1-month washout with reordered clips.
  • High-confidence assignments used the 3-second rule; outcomes included sensitivity, specificity, PPV, NPV, accuracy, high-confidence rate, decision time, and PIVI/SODA thresholds.
  • Analyses used paired comparisons (McNemar’s test, paired t-test/Wilcoxon signed-rank test), Moskowitz-Pepe method for PPV/NPV, multivariable logistic regression for decision time and accuracy, and Cohen’s kappa with a matched-pair permutation test.

Limitations

  • Stated: not conducted in a real-time clinical setting; fixed 10-second video clips may not represent actual colonoscopy; downstream clinical outcomes such as polyp misclassification, unnecessary resections, or follow-up implications were not assessed; only one CADx system (ENAD-CADx) was evaluated, so generalizability to other AI platforms is uncertain.
  • Evident from the text: the ex vivo online platform and video-clip format limit ecological validity, and external validation in real-time clinical settings is needed.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.5009/gnl250331

Cites

37 of 37 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

9 of 22 cited sources checked · 1 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Achieving PIVI and SODA competencies depends on the accuracy of high-confidence diagnoses, as these criteria are solely based on HC cases.

    “Achieving PIVI and SODA competencies depends on the accuracy of high-confidence (HC) diagnoses, as these criteria are solely based on HC cases.2,8”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [2] source
    • cites [8] source

      Figure 2 (page not specified in our copy) defines both SODA competence standards using sensitivity and specificity for high-confidence optical diagnoses, supporting that part of the claim. The Introduction mentions PIVI competence criteria but does not establish in the passages found that they are based solely on high-confidence cases, so the claim about both sets of criteria is not fully supported. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Premisecitation

    Attaining proficiency in optical diagnosis and meeting the SODA or PIVI standards can be particularly challenging for novices.

    “Attaining proficiency in optical diagnosis and meeting the accurate standards established by Simple Optical Diagnosis Accuracy (SODA) or Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) can be particularly challenging for novices.2,7,8”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [2] source
    • cites [7] source
    • cites [8] source

      The Introduction (page unavailable) says proficiency requires training and meeting competence standards, and that PIVI criteria are difficult to implement in daily practice. It does not say novices particularly struggle to attain proficiency or meet either standard; the Conclusion and future prospects instead describes SODA standards as achievable. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Premisecitation

    Integration of CADx into clinical practice has not consistently enhanced the diagnostic accuracy of endoscopists.

    “However, their integration into clinical practice has not consistently enhanced the diagnostic accuracy of endoscopists.14-16”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [14] source
    • cites [15] source

      The Results (Study outcomes; page unavailable) report 79.4% accuracy for POLAR versus 83.0% for endoscopists, and the Discussion describes their accuracy as comparable. This supports the narrower point that CADx did not outperform endoscopists in this real-time comparison. The study compared their diagnoses; it did not test whether using CADx enhanced endoscopists’ own accuracy after integration into clinical practice. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [16] source
  • Premisecitation

    The limited clinical uptake of diagnose-and-leave and resect-and-discard strategies is due to their suboptimal accuracy.

    “However, the limited clinical uptake of these strategies is due to their suboptimal accuracy.3-6”
    Quote in the paperSubstantial claim3 of 4 sources checked
    • cites [3] source
    • cites [4] source

      The Discussion (page unavailable) says high performance is important for acceptance of resect-and-discard and that leaving polyps in place requires a very high negative predictive value. The paper also reports that surveillance-interval agreement fell below the recommended threshold in its community-practice NBI study. It does not measure clinical uptake or establish that suboptimal accuracy is the cause of limited uptake of both strategies generally. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [5] source

      The Discussion reports that studies involving less-experienced endoscopists have not consistently met accuracy benchmarks, supporting performance as one obstacle. But it also identifies lack of incentives, auditing resources, and quality monitoring as barriers to widespread implementation; it does not establish that limited clinical uptake of both strategies is due to suboptimal accuracy alone. Page numbers are unavailable in our copy. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [6] source

      The Discussion (page unavailable) finds that optical diagnosis by non-experts was not accurate enough to replace histology, and the clinical-practice section says it cannot currently be recommended for routine use. These passages support suboptimal accuracy as a barrier, but the paper does not establish that limited clinical uptake of both strategies is due to that cause. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Findingown data

    When decision time exceeded 1.73 seconds, the odds ratio for accuracy decreased from 0.73 to 0.41.

    “When the decision time exceeded 1.73 seconds, the odds ratio for accuracy decreased from 0.73 to 0.41, supporting the implementation of a time-based threshold for HC assignments.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 8
  • May combine two claims. 18
  • The quote only partly states this claim. 1

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    Colorectal polyps were histologically characterized as adenoma, hyperplastic polyp, or sessile serrated lesion by expert pathologists according to the revised Vienna and World Health Organization classification.

    “Colorectal polyps were histologically characterized as adenoma (AD), HP, or sessile serrated lesion (SSL) by expert pathologists according to the revised Vienna and World Health Organization classification.3,4”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [3] source
    • cites [4] source

      The Methods (In Vivo Phase of the Study; Statistical Analysis, page unavailable) say that pathology reports supplied polyp diagnoses and distinguish adenomas from sessile serrated adenomas; the paper also discusses hyperplastic polyps. Our copy does not say that expert pathologists made the diagnoses or that they used the revised Vienna and WHO classifications. (LLM, conf=0.83)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationcitation

    The limited impact of CADx may stem from the absence of standardized workflows, the complexity of deep learning algorithms, and nuanced interactions between artificial intelligence and human operators.

    “This limited impact may stem from the absence of standardized workflows, the complexity of deep learning algorithms, and nuanced interactions between artificial intelligence (AI) and human operators.17”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [17] source

      The Introduction discusses how over-reliance, under-reliance, and uncertainty about AI reliability can hinder human–AI collaboration, supporting the point about nuanced interactions. It does not identify absent standardized workflows or the complexity of deep-learning algorithms as reasons for CADx’s limited impact. The Discussion instead reports that the lack of algorithmic transparency did not disrupt physicians’ performance in this study. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    CADx systems designed to distinguish neoplastic from nonneoplastic polyps have demonstrated accuracies of 78.8% to 93.2% in clinical trials.

    “These systems, designed to distinguish neoplastic from nonneoplastic polyps, have demonstrated accuracies of 78.8% to 93.2% in clinical trials.13”
    Quote in the paperNot a substantial claim1 of 1 sources checked

  • May combine two claims.

    Findingown data

    Overall, agreement with ground truth was moderately high for endoscopist-alone diagnoses (κ=0.44; 95% CI, 0.41 to 0.47) and significantly improved with CADx assistance (κ=0.69; 95% CI, 0.66 to 0.71; p<0.001).

    “Overall, agreement with ground truth was moderately high for endoscopist-alone diagnoses (κ=0.44; 95% CI, 0.41 to 0.47) and significantly improved with CADx assistance (κ=0.69; 95% CI, 0.66 to 0.71; p<0.001).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    CADx improved the overall and HC diagnostic accuracy among trained experts but was associated with lower HC rates and longer decision-making times.

    “CADx improved the overall and HC diagnostic accuracy among trained experts but was associated with lower HC rates and longer decision-making times.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The CADx system had a PPV of 89.2% (95% CI, 87.9% to 90.5%) and an NPV of 91.4% (95% CI, 89.7% to 92.9%).

    “The PPV was 89.2% (95% CI, 87.9% to 90.5%), and the NPV was 91.4% (95% CI, 89.7% to 92.9%).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The novice group had the lowest baseline agreement (κ=0.37) but showed the greatest relative improvement with CADx support, reaching a high level of agreement (κ=0.65).

    “The novice group had the lowest baseline agreement (κ=0.37) but showed the greatest relative improvement with CADx support, reaching a high level of agreement (κ=0.65).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Trained experts demonstrated high agreement at baseline (κ=0.63), which further improved to κ=0.75 with CADx.

    “When stratified by group, trained experts demonstrated high agreement at baseline (κ=0.63), which further improved to κ=0.75 with CADx.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingreasoning

    For SODA, a high specificity was observed for the 'resect-and-discard' strategy; however, novices did not achieve the 80% benchmark.

    “For SODA, a high specificity was observed for the”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The CADx system had a sensitivity of 95.1% (95% CI, 94.1% to 96.0%) and a specificity of 82.1% (95% CI, 79.9% to 84.1%).

    “The sensitivity and specificity of the CADx system were 95.1% (95% CI, 94.1% to 96.0%) and 82.1% (95% CI, 79.9% to 84.1%), respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Beginners often find it challenging to differentiate between high-confidence and low-confidence decisions owing to the absence of a clear definition or standardized process, resulting in variability and uncertainty in optical diagnosis.

    “However, beginners often find it challenging to differentiate between HC and low-confidence decisions owing to the absence of a clear definition or standardized process, resulting in variability and uncertainty in optical diagnosis.2”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [2] source

  • May combine two claims.

    Interpretationown data

    CADx support significantly reduced interobserver variability, with the greatest effect in novice endoscopists.

    “The present analysis revealed that CADx support significantly reduced interobserver variability, with the greatest effect in novice endoscopists.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    CADx assistance significantly improves optical diagnostic accuracy for less experienced endoscopists, but this improvement was not statistically significant in the trained expert group.

    “Our study demonstrated that while CADx assistance significantly improves optical diagnostic accuracy for less experienced endoscopists, this improvement was not statistically significant in the trained expert group.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The relationship between decision time and diagnostic accuracy was evaluated using a multivariate logistic regression model adjusted for polyp characteristics including size, location, and morphology.

    “The relationship between decision time and diagnostic accuracy was evaluated using a multivariate logistic regression model adjusted for polyp characteristics, including size, location, and morphology.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Trained experts may experience a cognitive load when a CADx suggestion conflicts with their initial assessment, leading to longer decision-making time and hesitation to blindly follow the AI.

    “Trained experts, who already possess high diagnostic accuracy and rely on established diagnostic patterns, may experience a cognitive load when a CADx suggestion conflicts with their initial assessment. This can lead to a longer decision-making time and a hesitation to blindly follow the AI, as reflected in the non-significant change in HC rate and decision time in our study.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    This method decreases interobserver variability, optimizes optical diagnostic strategies, and may enhance the cost-effectiveness of polyp management.

    “This method decreases interobserver variability, optimizes optical diagnostic strategies, and may enhance the cost-effectiveness of polyp management.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The participating endoscopists were blinded to the histological distribution and diagnoses.

    “The participating endoscopists were blinded to the histological distribution and diagnoses.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    The second phase of the study took place after a 1-month washout period with video clips reordered randomly to minimize recall bias.

    “To minimize recall bias, the second phase took place after a 1-month washout period with video clips reordered randomly.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The 3-second rule functions as a self-regulatory mechanism, enabling experienced endoscopists to make frequent HC diagnoses while promoting careful assessment among less experienced practitioners and maintaining diagnostic accuracy across expertise levels.

    “This pattern suggests the 3-second rule functions as a self-regulatory mechanism, enabling experienced endoscopists to make frequent HC diagnoses while promoting careful assessment among less experienced practitioners and maintaining diagnostic accuracy across expertise levels.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Integration of CADx with a structured, time-based confidence framework improved both diagnostic accuracy and consistency across experience levels.

    “Integration of CADx with a structured, time-based confidence framework improved both diagnostic accuracy and consistency across experience levels.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Combining the 3-second rule with CADx provides a standardized framework for the optical diagnosis of colorectal polyps, enhancing confidence level determination and diagnostic accuracy.

    “In conclusion, combining the 3-second rule with CADx provides a standardized framework for the optical diagnosis of colorectal polyps, enhancing confidence level determination and diagnostic accuracy.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The 3-second rule organizes HC assignments, whereas CADx improves performance through objective analysis.

    “The 3-second rule organizes HC assignments, whereas CADx improves performance through objective analysis.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Interobserver agreement was assessed using Cohen's kappa with 95% Wald CIs.

    “Interobserver agreement was assessed using Cohen’s kappa with 95% Wald CIs.”
    Quote in the paperSubstantial claim
23 of 23

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Conflicting significance claims for trained experts

    Major

    The Results state that both overall and high-confidence accuracy improved significantly in all three groups, which includes trained experts. The Discussion says the accuracy improvement in the trained expert group was not statistically significant.

1 of 1

Cited works that were retracted

Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.

  • Cites a retracted work: 10.1057/s41599-023-01787-8

    Info

    The cited work was retracted on 2026-02-03 (notice 10.1057/s41599-026-06602-8), not before this paper was published (2025-11-27).

    Cited work 10.1057/s41599-023-01787-8 · retracted 2026-02-03 · notice 10.1057/s41599-026-06602-8 · this paper published 2025-11-27

1 of 1

Cites

1 retracted

Flagged first, then by citations

Cited by

Still being checked.

Related

Authors
Kang, H. Y., Kang, S., Chung, G. E., Baek, D. H., Lee, H. S., Park, J., Yang, S. Y., Lim, S. H., Choi, J. M., Kim, J., & Bae, J. H.
Affiliations
  • Seoul National University (KR)
  • Ainex Corporation, Ltd., Seoul, Korea
  • Pusan National University Hospital (KR)
  • Inje University Busan Paik Hospital (KR)
Corresponding author
Bae, Jung Ho
ORCID
Venue
Gut and Liver
Publisher
The Editorial Office of Gut and Liver
Published
2025-11-27
Volume
20
Issue
1
Pages
86-96
ISSN
  • 1976-2283 (ISSN-L)
  • 2005-1212
DOI
10.5009/gnl250331
PMC
PMC12800665
OpenAlex
W4416767674
Semantic Scholar
777657145615d806e7075a80b4d9d940893143a6
Cited by
0
References
37
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Gut and Liver (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Research Foundation of Korea (RS-2024-00356090)
  • Ministry of Science and ICT, South Korea (RS-2024-00356090)
  • Ainex Corporation (90-2025-0002)
Topics
  • Health Sciences › Medicine › Oncology › Colorectal Cancer Screening and Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
computer-aided diagnosis, colorectal polyps, optical diagnostics, narrow-band imaging, endoscopist experience, adenoma, hyperplastic polyps, sessile serrated lesions, interobserver agreement, artificial intelligence in endoscopy
MeSH terms
Adenoma, Artificial Intelligence, Clinical Competence, Colonic Polyps, Colonoscopy, Diagnosis, Computer-Assisted, Female, Humans, Male, Middle Aged, Colorectal Neoplasms, Narrow Band Imaging

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-03.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-03.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review