Computer-assisted detection of colorectal polyps: a narrative review of clinical utility, ongoing limitations, and opportunities for advancement

Introduction

Background

Colorectal cancer (CRC) remains a major global health challenge as the second leading cause of cancer-related deaths worldwide and the third most common cancer worldwide (1). Colonoscopy is a safe screening modality that reduces the incidence and mortality of CRC, as the only tool which both allows for detection and removal of precancerous and cancerous lesions (2). However, colonoscopy has its limitations, and interval CRCs can still arise despite screening. More than 1,700 to 9,200 interval CRCs are predicted from 2019 to 2029 in the United States due to missed adenomas during screening colonoscopy (3). Despite advances in CRC screening, the burden of CRC persists, with over 1.9 million new cases and more than 900,000 deaths globally in 2022 (1,4).

Rationale and knowledge gap

Although there are proven benefits of screening colonoscopy to reduce CRC incidence and mortality, significant challenges remain in ensuring consistently high-quality examinations. Although the rate of interval CRC is the most relevant patient-centered outcome after CRC screening, adenoma detection rate (ADR) is used as the standard proxy for colonoscopy quality. ADR is the endoscopist-dependent percent of colonoscopies that detect at least one histologically proven adenoma or carcinoma, and it is considered the gold standard quality indicator of colonoscopy (5). Each 1% increase in ADR decreases the rate of interval CRC by 3% and decreases mortality by 5% (5). Current guidelines in the United States recommend physicians have ADRs of at least 30% for average-risk males and 20% for average-risk females (6). Similarly, as a secondary quality metric, polyp detection rate (PDR) is commonly cited. Prior studies have shown that a PDR of less than 20% is associated with an increased risk of interval CRC (7). In addition to polyp-specific and anatomic challenges in achieving sufficient ADR, there remains significant inter-operator variability in CRC detection rates with colonoscopy due to differences in endoscopist skill (8). Various techniques have been studied to reduce adenoma miss rate (AMR) and improve ADR and PDR, including real-time alerts of patient positioning during colonoscope withdrawal, prolonging withdrawal time (WT), administering split-dose bowel preparation, utilizing an in-person observer, and incorporating artificial intelligence (AI) (9,10).

AI holds tremendous potential in endoscopy and colorectal surgery from polyp detection during colonoscopy to predicting lymph node metastases in CRC based on whole slide imaging (11). In colonoscopy, computer-assisted polyp detection (CADe) systems leverage deep learning software algorithms to automatically highlight polyps in real-time, acting as a “second observer”. CADe systems have demonstrated the ability to increase ADR and PDR, reduce AMR, and enhance consistency across operators to improve diagnostic accuracy and colonoscopy quality in multiple studies (12-14). However, there remains heterogeneity among randomized controlled trials (RCTs) that have assessed the efficacy of CADe, and long-term effectiveness data does not yet exist (15-17).

Objectives

The rapid proliferation of commercial CADe systems has outpaced the accumulation of real-world evidence and the development of consensus-based practices (18). This targeted narrative review aims to bridge these gaps by synthesizing current knowledge of CADe technologies, summarizing available clinical evidence and professional society recommendations, and outlining priorities for future research and standardization. Our goal is to equip providers, specifically gastroenterologists and colorectal surgeons, with the knowledge needed to make informed decisions about incorporating CADe into their screening practices. We present this article in accordance with the Narrative Review reporting checklist (available at https://tgh.amegroups.com/article/view/10.21037/tgh-25-116/rc).

Methods

A comprehensive literature search was performed in PubMed, limiting results to articles published in English from the inception of the database up to and including July 31, 2025. We included observational studies, RCTs, systematic and narrative reviews, meta-analyses, guidelines, consensus conferences, case reports, and comparative studies. Key search phrases included “computer-assisted detection of colorectal polyps”, “CADe in colonoscopy”, “AI-enhanced colonoscopy”, and “artificial intelligence in colonoscopy”. Eligible studies were synthesized into a narrative review summarizing current evidence on CADe in colonoscopy, its limitations, and future directions. The search was conducted independently by two authors (M.L.D. and S.M.L.), and results were manually cross-checked for consistency. Since this is a targeted review of CADe technology, some RCTs and systematic reviews that did not report AMR, ADR, or WT were excluded. The search strategy is detailed in Table 1.

Results

Background of AI in endoscopy

AI refers to the field of computer science focused on creating systems capable of performing tasks that typically require human intelligence. These tasks include pattern recognition, decision-making, natural language understanding, and image interpretation (19). In the context of endoscopy, through machine learning and deep learning algorithms, AI enables the development of real-time tools for image analysis, such as CADe systems that can identify polyps instantaneously during colonoscopy. By mimicking human visual cognition and continuously improving through data exposure, AI systems in endoscopy aim to augment clinician performance, reduce variability, and improve diagnostic outcomes (20).

In response to the challenges of colonoscopy, AI entered the endoscopy scene in 2021 when the United States Food and Drug Administration (FDA) approved GI Genius, which uses CADe to improve detection of polyps in real-time (21). The regulatory approval and commercial rollout of CADe systems marked a major turning point in the integration of AI into clinical endoscopy. Since 2021, several other approved CADe systems have been developed and approved for use worldwide including SKOUT, ENDO-AID, and EndoAngel (Table 2). More recently, the first cloud-based CADe platform, CADDIE, has been approved by the FDA. As of 2025, there are more than a dozen CADe systems commercially available. These regulatory milestones have not only validated the safety, feasibility, and performance of CADe technologies but also accelerated their integration into routine clinical practice worldwide.

Beyond colorectal polyp detection, AI has also been applied to computer-assisted diagnosis (CADx), enabling real-time optical characterization of polyps to help differentiate neoplastic from non-neoplastic lesions. Systems that combine CADe and CADx for real-time detection and histological diagnosis are now entering the market and have the potential to further improve efficiency and accuracy of screening colonoscopies (22,23).

Technical aspects and variability of CADe systems

CADe in colonoscopy is designed to assist endoscopists by analyzing the colonoscope’s video feed in real-time. Using advanced image recognition algorithms, the system processes each frame to identify visual patterns suggestive of polyps. When a potential lesion is detected, CADe software generates an on-screen visual cue, such as a bounding box or marker, drawing the endoscopist’s attention to the area. CADe systems are built upon deep learning architectures, most commonly convolutional neural networks (CNNs), which are trained to identify patterns associated with polyps from large datasets of annotated endoscopic images or videos (24). The quality, diversity, and scale of these training datasets play a critical role in the generalizability and accuracy of the final model (25). Systems trained on image-based datasets may capture static features effectively but often struggle with motion blur, partial views, or dynamic lighting changes encountered during real-time procedures (24). Conversely, video-based training, including temporal and contextual information across frames, allows for better modeling of polyp movement, transient appearances, and background variability (26). To improve performance, training datasets must be curated with meticulously labeled data, often requiring expert endoscopist annotation of lesion boundaries and classifications. However, lack of standardization in labeling protocols and variability in imaging sources between regions and platforms may introduce biases that affect performance across patient populations and device types (26).

CADe systems consistently perform well in detecting diminutive polyps, which are frequently missed in standard practice. Several studies have demonstrated CADe’s high sensitivity for polyps less than five millimeters in size, which contributes to significantly improved ADR. For adenomas mostly less than five millimeters in size, Liu et al. reported an ADR of 0.24 without CADe and 0.39 with CADe (P<0.001) (27,28). However, although CADe consistently improves ADR across multiple studies, most of the increase may come from diminutive polyps (less than five millimeters in size). As such, the extent to which CADe-driven ADR improvements translate into meaningful reductions in interval CRC rates remains an area of debate. Importantly, the clinical significance of CADe may depend less on ADR and more on its ability to enhance detection of clinically relevant lesions, such as advanced adenomas and sessile serrated lesions. Accuracy of CADe may vary across polyp morphologies, such as flat and sessile lesions, which often exhibit subtle surface changes and lack overt protrusion making them more challenging for both human and algorithmic detection.

While some advanced CADe systems trained on diverse polyp types have shown improved specificity and sensitivity for these challenging lesions, others may underperform if such examples were underrepresented in training data. One study that processed previously recorded colonoscopy videos using two CADe systems, GI Genius and Endo-AID, showed that both systems did not detect sessile serrated adenomas in the right colon in two separate patients (29). In one of the two cases, the missed sessile serrated polyp by CADe would have caused a 7-year delay in patient follow-up. Furthermore, GI Genius and Endo-AID systems had lower sensitivity for flat polyps when compared to pedunculated and elevated lesions (51.70% vs. >80%) (29). Another study by Sinonquel et al. demonstrated that CADe was not superior to human polyp detection (sensitivity 94.6% vs. 96.0%), but it outperformed humans when restricted to adenomas, highlighting the variability of CADe (30). This highlights the need for continuous system refinement using varied and pathologically confirmed training datasets.

The overall performance of CADe systems is influenced by several procedural factors, notably bowel preparation quality, WT, and imaging modality. Poor bowel preparation can obscure lesions or produce false signals due to residual stool and debris, reducing both sensitivity and specificity (27). Some CADe systems incorporate preprocessing filters to reduce these artifacts, but their effectiveness varies (31). Adequate WT remains a cornerstone of high-quality colonoscopy, and studies have shown that CADe can enhance detection across a range of withdrawal speeds. However, its benefits are most pronounced with adequate inspection times (32).

The type of imaging modality also plays an essential role in colonoscopy. Most CADe systems are designed for white light endoscopy (WLE), which remains the global standard for colonoscopy. However, a growing number of CADe platforms are being optimized to work with narrow-band imaging (NBI) or linked color imaging (LCI), enhancing contrast and mucosal details. NBI emphasizes the microstructure and capillaries on mucosal surfaces by modifying the center wavelength and bandwidth of light into narrow-band illumination within the hemoglobin absorption band (33). LCI provides a significant color difference between the lesion and surrounding mucosa by separating the red color to enhance slight color differences (33). Integrating CADe with advanced imaging modalities holds promise but requires further validation and regulatory alignment (29). Since CADe can only identify lesions within the field of view, adjunctive strategies such as NBI and LCI as well as mechanical exposure devices such as transparent caps and Endocuff attachments have been investigated to enhance mucosal visualization. Emerging data suggest that combining CADe with these approaches may yield additive improvements in ADR by expanding mucosal exposure and optimizing mucosal contrast (34).

CADe and clinical outcomes

Over the past five years, a growing body of prospective trials has evaluated the clinical performance and efficacy of CADe systems, a few of which are highlighted in Table 3 (9,12,35-38). Two multi-center RCTs by Repici et al. and Wang et al. using GI Genius and EndoScreener, respectively, established the clinical utility of using CADe in real-time colonoscopy (12,35). These studies randomized patients to undergo standard high-definition colonoscopy with or without CADe assistance and assessed outcomes such as ADR, PDR, WT, and FP rates.

The multicenter trial of GI Genius reported a significant ADR increase [54.8% in the CADe group vs. 40.4% in the control group; relative risk (RR) 1.3, 95% confidence interval (CI): 1.14–1.45] with no increase in WT, supporting an effective reduction in missed adenomas (35). Wang et al. reported a similar improvement in ADR from 28% to 34% [odds ratio (OR) 1.36, 95% CI: 1.03–1.79] with the use of the EndoScreener (12,35). Xu et al. performed a multicenter RCT with the non-FDA-approved Eagle Eye system and demonstrated an increase in ADR (32.4% vs. 39.9%, P<0.001) (37). Another study evaluated adenomas per colonoscopy (APC) as the primary endpoint, rather than ADR. It showed a significant increase in APC with the use of the CADe system SCOUT (0.83 without CADe vs. 1.05 with CADe, P=0.002) (38). Other studies report relative ADR increases ranging from 20% to 50% with CADe systems compared to colonoscopy without CADe, with pronounced gains in the detection of polyps less than five millimeters in size and non-pedunculated polyps (13,39).

Improvements in PDR have paralleled these ADR findings, with CADe systems aiding in the recognition of hyperplastic and nonneoplastic polyps. Wang et al. reported improvements in polyp detection with CADe compared to standard colonoscopy (0.29 to 0.45 polyps detected per colonoscopy, P<0.001) (32). Meta-analyses have also shown significant increases in PDR with CADe compared to standard colonoscopy (34.6% to 50.3%, P<0.01) (13).

These improvements in ADR and PDR were achieved without significantly increasing WT, indicating a potential to use CADe without compromising clinical efficiency. Most RCTs reported minimal or no clinically meaningful prolongation of WT attributable to CADe. Xu et al. demonstrated increased WT with CADe (7.78 vs. 8.25 minutes, P=0.004), and Wang et al. similarly reported increased WT with CADe (6.39 vs. 6.89 minutes, P<0.001) (32,37). When biopsy time is excluded from WT, there was no significant difference in WT in the Wang et al. study (6.07 minutes without CADe vs. 6.18 minutes with CADe, P=0.15) (32). When small increases in WT occurred, they were typically explained by endoscopist inspection and/or biopsy of AI-flagged regions rather than delays related to the CADe system itself, such as latency, setup or takedown requirements, or troubleshooting needs (32).

Meta-analyses and pooled randomized data corroborate these improvements related to ADR, PDR, and WT with CADe (13,40). Hassan et al. showed a pooled relative risk of 1.44 (95% CI: 1.27–1.62, P<0.01) for ADR in the CADe groups with similar improvements in PDR (13). Another meta-analysis including 11 studies and 6,856 patients showed higher ADR with CADe compared to standard colonoscopy (OR 1.51, P=0.003) (39). One meta-analysis reported only minimal changes in WT, with a pooled mean effect size of 15 seconds across eight studies, suggesting negligible workflow disruption while still providing enhanced detection support (39). These studies demonstrate that CADe may be an effective, efficient adjunct for improving colonoscopy quality while minimally impacting procedure time.

False positive (FP) rate is an inherent limitation of any detection system and can carry distinct clinical implications. The exact definition of FP in the application of AI for polyp detection remains ambiguous across studies (41). In general, FPs refer to computer prompts indicating polyps that the endoscopist does not consider to be polyps and are often triggered by artifacts such as folds or debris (42). Despite discrepancies in the exact definition, FPs can lead to unnecessary polypectomies, prolonged procedure time, and operator alarm fatigue (42-44). Across multiple RCTs, there were significant increases in biopsies of non-neoplastic polyps in CADe groups compared to control groups (41). Most commercial systems maintain an average FP rate between 1.2% and 5.6% for CADe systems (45). One study has shown that the effectiveness of CADe appears to decrease when the FP rate exceeds five FPs per minute, as this diminishes the clinical value (46). Recent improvements in algorithm robustness, real-time polyp characterization with the integration of CADx, and improved image preprocessing may help reduce FP rates, but no system currently achieves 100% specificity (43).

The use of CADe may be particularly advantageous for trainee education and non-expert endoscopists. Evidence comparing ADRs between early career endoscopists and experienced endoscopists suggests that the use of CADe could help standardize performance across levels of expertise (47,48). As a “second observer” to trainees, CADe has the potential to reinforce meticulous inspection habits by drawing attention to subtle lesions through frequent alerts, thereby promoting an AI-driven culture of thorough evaluation. Nonetheless, questions remain, as long-term data are lacking on the impact of CADe-assisted training on endoscopists’ careers, procedure WTs, and the overall cost implications.

On the contrary, there is growing concern that the use of CADe systems may predispose endoscopists, particularly trainees, to over-reliance on AI alerts, possibly leading to deskilling over time. This “automation bias” may reduce vigilance and independent detection capabilities (49,50). Expert consensus echoes the concern that overreliance on CADe could diminish clinical autonomy and erode diagnostic skills if not paired with rigorous training and oversight (51). However, on the other hand, CADe has been shown to serve as an equalizing tool to reduce operator variability and even as a teaching tool in some instances (16). Some studies demonstrate that CADe usage may elevate PDRs uniformly across centers, independent of the baseline operator ADR (35). Nevertheless, CADe cannot replace the role of endoscopists, as human verification remains essential to evaluate the alerts and determine appropriate lesion management.

Challenges with generalizability, integration, and cost

Although evidence supporting the adoption of CADe in population-based screening is steadily growing, uncertainties remain regarding its generalizability, integration into existing systems, long-term patient benefit, potential patient burden from increased adenoma detection of unclear clinical significance, and the possibility of higher healthcare costs.

Many CADe systems are trained and validated within narrowly defined datasets from high-volume academic centers in Asia, Europe, and North America, often under optimal bowel preparation and expert supervision (52). This raises concern about generalizability to diverse patient populations, practice settings, and healthcare systems. Furthermore, variability in polyp prevalence, morphology distribution, and procedural protocols may influence real-world performance. External validation studies are limited and may involve the same vendors or overlapping datasets used in algorithm development, which may overestimate generalizability.

A recent multicenter, crowd-sourced deep learning challenge, EndoCV2021, highlighted high inter-center variability in detection metrics, underscoring generalizability concerns of AI systems (53). Despite teams of gastroenterologists and computational experts applying a wide range of algorithms to train and sequence multicenter dataset, performance on out-of-sample, unseen datasets remained limited as reflected by low dice similarity coefficients (53). Conversely, a self-supervised approach using a Masked Siamese Network demonstrated that certain CADe models maintain high detection performance when applied to entirely different geographic datasets and imaging modalities (e.g., Israeli-trained models tested on Japanese NBI images) (54). These findings indicate both the challenges and strategies for improving external validation. Large-scale, multicenter, prospective evaluations across diverse populations are needed to confirm CADe performance in routine practice.

Real-world reproducibility of CADe RCTs also poses challenges. While RCTs provide critical evidence on the efficacy of CADe under controlled conditions, they may not fully capture the complexities of routine endoscopic practice. Blinding is not feasible in most CADe trials, and trial environments often involve high-volume centers and experienced endoscopists, which can limit generalizability. Real-world studies, by contrast, reflect broader patient populations, varying operator experience, and heterogenous practice settings, offering important complementary insights into the effectiveness of CADe in everyday clinical practice. One recent meta-analysis of 12 real-world CADe studies found statistically significant but clinically minimal improvement in ADR with CADe, while another meta-analysis of eight real-world studies did not show an increase in ADR or APC (55,56). This conflicting evidence emphasizes the importance of integrating both trial-based and real-world data when interpreting the clinical impact and potential of CADe.

In terms of system compatibility, adopting CADe is often complicated by harmonization issues across diverse endoscopy systems (57). Many CADe systems are proprietary and optimized for specific hardware, which constrains flexibility for centers using multi-vendor or legacy equipment (58). Although some developers are exploring platform-agnostic or software-only models, technical and regulatory hurdles remain significant impediments to seamless integration and widespread adoption of CADe.

In addition to system integration issues, introducing AI into clinical workflows also raises complex ethical and legal issues. CADe influences clinical decisions, such as polyp detection and surveillance planning, creating ambiguous lines of responsibility when errors occur. Clinicians must understand that CADe tools are advisory, and, as the clinician, they retain ultimate responsibility for diagnosis and management (59). Broader concerns include algorithmic bias, data privacy, lack of transparency (e.g., “black box models”), informed consent, and equitable access. Calls for robust governance frameworks focusing on transparency, accountability, and inclusivity are increasingly emphasized in healthcare AI discourse (60-62).

By increasing the number of detected adenomas, CADe has the potential to reclassify patients into shorter recommended surveillance intervals under current guideline frameworks. This may increase short-term demand for surveillance colonoscopy and pathology services while potentially improving long-term cancer prevention, although long-term data does not currently exist (18). Conversely, the combination of CADe (detection) with validated CADx (diagnosis and characterization) and “resect-and-discard” strategies could mitigate some costs and surveillance burdens by allowing in-situ decision making for diminutive lesions. CADx could enable real-time optical characterization of polyps and differentiate neoplastic from non-neoplastic lesions. Policy decisions about whether and how additional small adenomas detected by CADe should alter surveillance recommendations will require consensus from professional societies informed by outcomes data showing that the incremental lesions detected meaningfully change cancer risk (13,14).

Health-economic evaluations to date suggest CADe can be cost-effective under plausible scenarios, especially where baseline ADRs are low, procedure volumes are high, or per-procedure CADe costs are modest. Modeling studies and microsimulation analyses indicate that modest per-procedure costs for CADe can be offset by reductions in interval cancers and associated treatment costs, though results are sensitive to assumptions about ADR (63). For CADe to be cost-effective based on one semi-Markov microsimulation study, ADR would need to increase to at least 30% with CADe, or each colonoscopy with CADe would need to cost less than $579 (63). Using these models, the use of CADe with an ADR threshold of 22% resulted in an overall cost savings of $7 per person, while there was an overall cost savings of $335 with an ADR threshold of 41% (63). Robust registry data and longitudinal follow-up studies are required to confirm the modeled benefits (14).

Guidelines and recommendations of routine use of CADe

Current national and international guidelines do not overtly recommend the routine use of CADe in colonoscopy (16,17,58). The American Gastroenterological Association recognized the potential of CADe to continue improving as an iterative AI application; however, the association ultimately issued no recommendation for or against the use of CADe (58). The British Medical Journal issued a weak recommendation against the routine use of CADe, citing uncertainty around long-term outcomes, potential overdiagnosis, and increased patient burden (17). In contrast, the European Society of Gastrointestinal Endoscopy provided a weak recommendation in favor of the use of CADe, noting its potential but limited benefits (64). Similarly, the Asia-Pacific Consensus group expressed weak support for its routine use in screening colonoscopies, provided it is available and cost effective (Table 4) (51).

Future directions

Future directions include developing more robust and generalizable CADe algorithms, studying long-term patient outcomes, evaluating CADe’s utility in early-onset CRC (EOCRC) screening, improving system compatibility, and integrating CADx with other AI capabilities.

EOCRC, defined as CRC diagnosed before age 50, has demonstrated a global increase in incidence, with cases more than doubling between the years 1990 and 2019 (65). Although updated screening guidelines in several countries have lowered the recommended initiation age for screening colonoscopy to 45 years old, important quality concerns remain (66). High ADRs are crucial for screening effectiveness; however, EOCRC is associated with lower ADRs on screening colonoscopy compared to later-onset disease (67). Given the rise in EOCRC incidence, the application of CADe could be a valuable strategy to enhance lesion detection in this understudied population. However, although CADe has potential to improve ADR and thereby contribute to EOCRC prevention, it is important to acknowledge that screening initiation age is a major determinant of EOCRC outcomes at the population level. Earlier initiation of screening likely yields a greater absolute impact on EOCRC and mortality than technological quality improvements alone. Rather than serving as a substitute, CADe should be viewed as a complementary tool that can optimize detection once patients enter the screening pathway. Framing CADe as part of a broader EOCRC prevention strategy, alongside earlier and more equitable access to screening, provides a more balanced perspective on its potential role.

The integration of CADe (detection) with CADx (diagnosis and characterization) represents a significant advancement in AI-assisted colonoscopy, enabling both lesion detection and real-time histology prediction with on-the-spot differentiation of neoplastic from non-neoplastic polyps. When integrated seamlessly, a CADe/CADx system could both flag a lesion and immediately suggest management (e.g., remove, discard, or leave), enabling more efficient and potentially cost-saving workflows. The application of CADx has the potential to reduce the number of unnecessary polypectomies of non-neoplastic polyps. GI Genius v3, a combined CADe and CADx system, when implemented into clinical practice, can provide optical diagnosis to guide “diagnose-and-leave” strategies for diminutive polyps (68,69).

Beyond lesion detection and characterization, emerging AI models now perform real-time semantic segmentation, delineating polyp boundaries during ongoing colonoscopy. This includes AI-driven recommendations for optimal snare placement, estimation of polyp size and depth, and even prediction of difficult-to-resect features. Novel AI models, such as NanoNet, offer lightweight, high-speed polyp segmentation suitable for real-time application, which could facilitate resection planning and improved visual margins when integrated with CADe and CADx systems (70).

AI systems are increasingly capable of quantifying polyp features and providing data-rich profiles that can enable tailored surveillance intervals. AI-generated lesion profiles—combining number, size, histological prediction, and growth dynamics—could inform individualized interval recommendations and optimize prevention while minimizing unnecessary colonoscopies. While personalized recommendation frameworks remain under development, this represents a logical future extension of AI-assisted colonoscopy.

To support continued innovation and validation, the establishment of multicenter registries and benchmarking initiatives will be vital. Shared, annotated datasets across diverse settings can facilitate comparative studies of CADe, help identify population-specific performance gaps, drive transparent benchmarking, and power long-term outcome research. The REAL-Colon dataset stands out as an extensive, multicenter annotated video library comprising 2.7 million frames and 350,000 bounding boxes from diverse colonoscopy recordings (26). This resource is poised to drive improved algorithm development, reproducibility, and cross-platform benchmarking.

Finally, several clinical trials are expanding the evidence base for CADe and CADe/CADx integrated AI systems in real-world settings. For instance, the PolyDeep Advance 3 trial was completed in early 2025, and data analysis is underway to evaluate the clinical validation of combined detection and diagnosis systems (71). Another trial using GI Genius to assess CADe and CADx performance in real-time polyp characterization is estimated to be completed by late 2025 (23). Regulatory bodies are also evolving approaches to AI oversight, exploring adaptive pre-market approvals, post-market performance monitoring, and software-as-a-medical-device pathways (69). These developments promise to streamline innovation while ensuring safety and clinical effectiveness.

Conclusions

Cumulative evidence from multiple RCTs and meta-analyses demonstrates that CADe systems consistently improve key quality indicators in colonoscopy, particularly ADR and PDR, without meaningfully impacting procedure time; however, uncertainties remain in terms of long-term outcomes, patient burden of more frequent surveillance, and potential costs. Benefits of CADe have been observed across diverse practice settings and among both expert and non-expert endoscopists, suggesting broad applicability and the potential for population-level impact on CRC prevention, yet limitations persist. Variability in CADe performance across different polyp morphologies, patient populations, and endoscopy platforms underscores the need for broader external validation, while concerns about over-reliance on AI, deskilling, and FPs highlight the importance of maintaining human oversight. Furthermore, economic analyses and real-world cost-effectiveness data remain limited, and optimal strategies for integrating CADe into workflow without extending procedure times are still being defined. Looking ahead, CADe is poised to become an integral component of a comprehensive AI-assisted endoscopy ecosystem that incorporates CADx for real-time polyp characterization, adapts to diverse clinical environments, and is supported by robust post-market surveillance. To achieve this vision, multicenter registries, equitable access initiatives, and adaptive regulatory frameworks will be essential to ensure global benefit, reducing disparities in CRC screening, and improve long-term outcomes. When implemented responsibly, CADe holds the potential to help deliver high-quality, consistent, and equitable colonoscopy for all patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://tgh.amegroups.com/article/view/10.21037/tgh-25-116/rc

Peer Review File: Available at https://tgh.amegroups.com/article/view/10.21037/tgh-25-116/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tgh.amegroups.com/article/view/10.21037/tgh-25-116/coif). M.L.H. has provided several hours of expert medical consultant for Iterative Health to validate and improve usability for their CADe platform Skout. A.K.M. serves as a paid consultant to Medtronic for the Hugo robotic platform outside of this submitted work. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Bretthauer M, Løberg M, Wieszczy P, et al. Effect of Colonoscopy Screening on Risks of Colorectal Cancer and Related Death. N Engl J Med 2022;387:1547-56. [Crossref] [PubMed]
3. Zhao S, Wang S, Pan P, et al. Magnitude, Risk Factors, and Factors Associated With Adenoma Miss Rate of Tandem Colonoscopy: A Systematic Review and Meta-analysis. Gastroenterology 2019;156:1661-1674.e11. [Crossref] [PubMed]
4. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
5. Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med 2014;370:1298-306. [Crossref] [PubMed]
6. Rex DK, Schoenfeld PS, Cohen J, et al. Quality indicators for colonoscopy. Gastrointest Endosc 2015;81:31-53. [Crossref] [PubMed]
7. Macken E, Van Dongen S, De Brabander I, et al. Post-colonoscopy colorectal cancer in Belgium: characteristics and influencing factors. Endosc Int Open 2019;7:E717-27. [Crossref] [PubMed]
8. Rex DK, Anderson JC, Butterly LF, et al. Quality indicators for colonoscopy. Gastrointest Endosc 2024;100:352-81. [Crossref] [PubMed]
9. Shaukat A, Tuskey A, Rao VL, et al. Interventions to improve adenoma detection rates for colonoscopy. Gastrointest Endosc 2022;96:171-83. [Crossref] [PubMed]
10. Köksal AŞ, Kalkan IH, Torun S, et al. A simple method to improve adenoma detection rate during colonoscopy: altering patient position. Can J Gastroenterol 2013;27:509-12. [Crossref] [PubMed]
11. Schultz KS, Linhares SM, Mongiu AK. What Every Colorectal Surgeon Should Know About Artificial Intelligence. Dis Colon Rectum 2025;68:1249-54. [Crossref] [PubMed]
12. Wang P, Liu X, Berzin TM, et al. Effect of a deep-learning computer-aided detection system on adenoma detection during colonoscopy (CADe-DB trial): a double-blind randomised study. Lancet Gastroenterol Hepatol 2020;5:343-51. [Crossref] [PubMed]
13. Hassan C, Spadaccini M, Iannone A, et al. Performance of artificial intelligence in colonoscopy for adenoma and polyp detection: a systematic review and meta-analysis. Gastrointest Endosc 2021;93:77-85.e6. [Crossref] [PubMed]
14. Lou S, Du F, Song W, et al. Artificial intelligence for colorectal neoplasia detection during colonoscopy: a systematic review and meta-analysis of randomized clinical trials. EClinicalMedicine 2023;66:102341. [Crossref] [PubMed]
15. Soleymanjahi S, Huebner J, Elmansy L, et al. Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis. Ann Intern Med 2024;177:1652-63. [Crossref] [PubMed]
16. Linhares SM, Schultz KS, Mongiu AK. Computer aided polyp detection has limited clinical efficacy. BMJ 2025;389:r732. [Crossref] [PubMed]
17. Foroutan F, Vandvik PO, Helsingen LM, et al. Computer aided detection and diagnosis of polyps in adult patients undergoing colonoscopy: a living clinical practice guideline. BMJ 2025;388:e082656. [Crossref] [PubMed]
18. Halvorsen N, Hassan C, Correale L, et al. Benefits, burden, and harms of computer aided polyp detection with artificial intelligence in colorectal cancer screening: microsimulation modelling study. BMJ Med 2025;4:e001446. [Crossref] [PubMed]
19. Choi RY, Coyner AS, Kalpathy-Cramer J, et al. Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol 2020;9:14. [Crossref] [PubMed]
20. Letafatkar N, El-Sehrawy AAMA, Prasad K, et al. Artificial intelligence in endoscopy and colonoscopy: a comprehensive bibliometric analysis of global research trends. Front Med (Lausanne) 2025;12:1532640. [Crossref] [PubMed]
21. Schultz KS, Hughes ML, Akram WM, et al. Artificial intelligence for the colorectal surgeon in 2024 – A narrative review of Prevalence, Policies, and (needed) Protections. Semin Colon Rectal Surg 2024;35:512-20.
22. Polydeep Advance I: Prospective Diagnostic Tests Trial with a Paired Study Design. 2022. Available online: https://clinicaltrials.gov/study/NCT05514301
23. Accuracy of Real Time Characterization in Artificial Intelligence-assisted Colonoscopy - A Prospective Quality Assurance Study. 2023. Available online: https://clinicaltrials.gov/study/NCT05754229
24. Jaspers TJM, Boers TGW, Kusters CHJ, et al. Robustness evaluation of deep neural networks for endoscopic image analysis: Insights and strategies. Med Image Anal 2024;94:103157. [Crossref] [PubMed]
25. Sun K, Wang Y, Qu R, et al. Comprehensive application of artificial intelligence in colorectal cancer: A review. iScience 2025;28:112980. [Crossref] [PubMed]
26. Biffi C, Antonelli G, Bernhofer S, et al. REAL-Colon: A dataset for developing real-world AI applications in colonoscopy. Sci Data 2024;11:539. [Crossref] [PubMed]
27. Liu WN, Zhang YY, Bian XQ, et al. Study on detection rate of polyps and adenomas in artificial-intelligence-aided colonoscopy. Saudi J Gastroenterol 2020;26:13-9. [Crossref] [PubMed]
28. Chen PJ, Lin MC, Lai MJ, et al. Accurate Classification of Diminutive Colorectal Polyps Using Computer-Aided Analysis. Gastroenterology 2018;154:568-75. [Crossref] [PubMed]
29. Troya J, Sudarevic B, Krenzer A, et al. Direct comparison of multiple computer-aided polyp detection systems. Endoscopy 2024;56:63-9. [Crossref] [PubMed]
30. Sinonquel P, Eelbode T, Pech O, et al. Clinical consequences of computer-aided colorectal polyp detection. Gut 2024;73:1974-83. [Crossref] [PubMed]
31. Ramzan M, Raza M, Khan ZF, et al. A review on computer-aided diagnostic system to classify the disorders of the gastrointestinal tract. Eur J Med Res 2025;30:674. [Crossref] [PubMed]
32. Wang P, Berzin TM, Glissen Brown JR, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut 2019;68:1813-9. [Crossref] [PubMed]
33. Nagai M, Suzuki S, Minato Y, et al. Detecting colorectal lesions with image-enhanced endoscopy: an updated review from clinical trials. Clin Endosc 2023;56:553-62. [Crossref] [PubMed]
34. Spadaccini M, Hassan C, Rondonotti E, et al. Combination of Mucosa-Exposure Device and Computer-Aided Detection for Adenoma Detection During Colonoscopy: A Randomized Trial. Gastroenterology 2023;165:244-251.e3. [Crossref] [PubMed]
35. Repici A, Badalamenti M, Maselli R, et al. Efficacy of Real-Time Computer-Aided Detection of Colorectal Neoplasia in a Randomized Trial. Gastroenterology 2020;159:512-520.e7. [Crossref] [PubMed]
36. Glissen Brown JR, Mansour NM, Wang P, et al. Deep Learning Computer-aided Polyp Detection Reduces Adenoma Miss Rate: A United States Multi-center Randomized Tandem Colonoscopy Study (CADeT-CS Trial). Clin Gastroenterol Hepatol 2022;20:1499-1507.e4. [Crossref] [PubMed]
37. Xu H, Tang RSY, Lam TYT, et al. Artificial Intelligence-Assisted Colonoscopy for Colorectal Cancer Screening: A Multicenter Randomized Controlled Trial. Clin Gastroenterol Hepatol 2023;21:337-346.e3. [Crossref] [PubMed]
38. Shaukat A, Lichtenstein DR, Somers SC, et al. Computer-Aided Detection Improves Adenomas per Colonoscopy for Screening and Surveillance Colonoscopy: A Randomized Trial. Gastroenterology 2022;163:732-41. [Crossref] [PubMed]
39. Aslam MF, Bano S, Khalid M, et al. The effectiveness of real-time computer-aided and quality control systems in colorectal adenoma and polyp detection during colonoscopies: a meta-analysis. Ann Med Surg (Lond) 2023;85:80-91. [Crossref] [PubMed]
40. Shah S, Park N, El Hage Chehade N, et al. Artificial intelligence and computer aided detection (CADe) systems improve adenoma miss rates, adenoma detection rates and polyp detection rates: a systematic review and meta-analysis. Gastrointestinal Endoscopy 2022;95:AB231-2.
41. Hsieh YH, Tang CP, Tseng CW, et al. Computer-Aided Detection False Positives in Colonoscopy. Diagnostics (Basel) 2021;11:1113. [Crossref] [PubMed]
42. Hassan C, Badalamenti M, Maselli R, et al. Computer-aided detection-assisted colonoscopy: classification and relevance of false positives. Gastrointest Endosc 2020;92:900-904.e4. [Crossref] [PubMed]
43. Okumura T, Imai K, Misawa M, et al. Evaluating false-positive detection in a computer-aided detection system for colonoscopy. J Gastroenterol Hepatol 2024;39:927-34. [Crossref] [PubMed]
44. Soons E, Rath T, Hazewinkel Y, et al. Real-time colorectal polyp detection using a novel computer-aided detection system (CADe): a feasibility study. Int J Colorectal Dis 2022;37:2219-28. [Crossref] [PubMed]
45. Papachrysos N, Smedsrud PH, Ånonsen KV, et al. A comparative study benchmarking colon polyp with computer-aided detection (CADe) software. DEN Open 2025;5:e70061. [Crossref] [PubMed]
46. Zhang C, Yao L, Jiang R, et al. Assessment of the role of false-positive alerts in computer-aided polyp detection for assistance capabilities. J Gastroenterol Hepatol 2024;39:1623-35. [Crossref] [PubMed]
47. Biscaglia G, Cocomazzi F, Gentile M, et al. Real-time, computer-aided, detection-assisted colonoscopy eliminates differences in adenoma detection rate between trainee and experienced endoscopists. Endosc Int Open 2022;10:E616-21. [Crossref] [PubMed]
48. Orzeszko Z, Gach T, Necka S, et al. The implementation of computer-aided detection in an initial endoscopy training improves the quality measures of trainees' future colonoscopies: a retrospective cohort study. Surg Endosc 2025;39:5276-86. [Crossref] [PubMed]
49. Campion JR, O'Connor DB, Lahiff C. Human-artificial intelligence interaction in gastrointestinal endoscopy. World J Gastrointest Endosc 2024;16:126-35. [Crossref] [PubMed]
50. Sinagra E, Rossi F, Raimondo D. Use of Artificial Intelligence in Endoscopic Training: Is Deskilling a Real Fear? Gastroenterology 2021;160:2212. [Crossref] [PubMed]
51. Koh FH, Li JW, Wong SH, et al. Asia-Pacific consensus on the use of artificial intelligence in colorectal cancer screening and surveillance. iGIE 2026; In press. [Crossref]
52. Repici A, Spadaccini M, Antonelli G, et al. Artificial intelligence and colonoscopy experience: lessons from two randomised trials. Gut 2022;71:757-65. [Crossref] [PubMed]
53. Ali S, Ghatwary N, Jha D, et al. Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci Rep 2024;14:2032. [Crossref] [PubMed]
54. Shor J, McNeil C, Intrator Y, et al. Predicting the generalization of computer aided detection (CADe) models for colonoscopy. Discov Artif Intell 2024;4:85.
55. Wei MT, Fay S, Yung D, et al. Artificial Intelligence-Assisted Colonoscopy in Real-World Clinical Practice: A Systematic Review and Meta-Analysis. Clin Transl Gastroenterol 2024;15:e00671. [Crossref] [PubMed]
56. Patel HK, Mori Y, Hassan C, et al. Lack of Effectiveness of Computer Aided Detection for Colorectal Neoplasia: A Systematic Review and Meta-Analysis of Nonrandomized Studies. Clin Gastroenterol Hepatol 2024;22:971-980.e15. [Crossref] [PubMed]
57. El-Sayed A, Lovat LB, Ahmad OF. Clinical Implementation of Artificial Intelligence in Gastroenterology: Current Landscape, Regulatory Challenges, and Ethical Issues. Gastroenterology 2025;169:518-30. [Crossref] [PubMed]
58. Sultan S, Shung DL, Kolb JM, et al. AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy. Gastroenterology 2025;168:691-700. [Crossref] [PubMed]
59. Jovanovic I. AI in endoscopy and medicolegal issues: the computer is guilty in case of missed cancer? Endosc Int Open 2020;8:E1385-6. [Crossref] [PubMed]
60. Ahmed MI, Spooner B, Isherwood J, et al. A Systematic Review of the Barriers to the Implementation of Artificial Intelligence in Healthcare. Cureus 2023;15:e46454. [Crossref] [PubMed]
61. Nasir M, Siddiqui K, Ahmed S. Ethical-legal implications of AI-powered healthcare in critical perspective. Front Artif Intell 2025;8:1619463. [Crossref] [PubMed]
62. Yu S, Lee SS, Hwang H. The ethics of using artificial intelligence in medical research. Kosin Med J. 2024;39:229-37.
63. Thiruvengadam NR, Coté GA, Gupta S, et al. An Evaluation of Critical Factors for the Cost-Effectiveness of Real-Time Computer-Aided Detection: Sensitivity and Threshold Analyses Using a Microsimulation Model. Gastroenterology 2023;164:906-20. [Crossref] [PubMed]
64. Bretthauer M, Ahmed J, Antonelli G, et al. Use of computer-assisted detection (CADe) colonoscopy in colorectal cancer screening and surveillance: European Society of Gastrointestinal Endoscopy (ESGE) Position Statement. Endoscopy 2025;57:667-73. [Crossref] [PubMed]
65. Pan H, Zhao Z, Deng Y, et al. The global, regional, and national early-onset colorectal cancer burden and trends from 1990 to 2019: results from the Global Burden of Disease Study 2019. BMC Public Health 2022;22:1896. [Crossref] [PubMed]
66. US Preventive Services Task Force. Screening for Colorectal Cancer: US Preventive Services Task Force Recommendation Statement. JAMA 2021;325:1965-77. [Crossref] [PubMed]
67. Chang TH, Chong LW, Chang HC, et al. Adenoma detection rate of screening colonoscopy among age 40–75 years: Implications for lowering the age for colorectal cancer screening. Adv in Digestive Med 2024;12:e13410.
68. Cherubini A, Dinh NN. A Review of the Technology, Training, and Assessment Methods for the First Real-Time AI-Enhanced Medical Device for Endoscopy. Bioengineering (Basel) 2023;10:404. [Crossref] [PubMed]
69. Hann A, Troya J, Fitting D. Current status and limitations of artificial intelligence in colonoscopy. United European Gastroenterol J 2021;9:527-33. [Crossref] [PubMed]
70. Jha D, Tomar NK, Ali S, et al. NanoNet: Real-Time Polyp Segmentation in Video Capsule Endoscopy and Colonoscopy. 2021 IEEE 34th International Symposium on Computer-Based Medical Systems (CBMS); 07-09 June 2021; Aveiro, Portugal. IEEE; 2021:37-43.
71. PolyDeep Advance 3: Randomized Clinical Trial Comparing High Definition Colonoscopy With PolyDeep Assisted High Definition Colonoscopy. 2022. Available online: https://clinicaltrials.gov/study/NCT05513261