Computer aided detection and diagnosis of polyps in adult patients undergoing colonoscopy: a living clinical practice guideline

Why is the guideline needed?

Colorectal cancer (CRC) is the third most common cancer and the second leading cause of cancer related death worldwide.123 Most colorectal cancers are adenocarcinomas and arise from precancerous polyps (adenomas or sessile lesions).2 Adenomas are precancerous growths in the lining of the colon or rectum and affect up to 40% of adults by the age of 80 years.4 Adenomas may become cancerous if left untreated, and up to 20% of individuals with a history of adenomas will develop CRC.5 Serrated lesions or polyps can also be premalignant lesions, similar to adenomas. Identifying serrated lesions that can develop into cancer endoscopically or histologically can be difficult. Colonoscopy, which allows for the detection and removal of these polyps (polypectomy), confers protection from the development of CRC.6 However, long term reduction in colorectal cancer depends on the quality of the colonoscopy, including adequate visualisation of the colon, appropriate detection, and complete resection of any precancerous polyps.78

Computer aided detection (CADe) systems are advanced software algorithms designed to assist endoscopists by highlighting potential polyps (including flat or non-polypoid lesions) during colonoscopy.9 These systems leverage artificial intelligence (AI) and machine learning (ML) technologies to analyse real-time video images from the colonoscopy, aiming to enhance the detection rate of polyps. CADe operates by identifying and marking areas of interest for closer examination by the endoscopist, thus acting as a second observer to potentially improve diagnostic accuracy and quality of colonoscopy. A commonly used measure to evaluate the performance of colonoscopy is the adenoma detection rate, which is endoscopist-dependent and varies based on indication, setting, and population.10111213

CADe is being increasingly used in clinical practice and is met with considerable enthusiasm and trepidation by gastroenterologists.14 It also represents the most extensively researched application of AI in medicine, with evaluation through randomised controlled trials (RCTs).15 In September 2023, an RCT addressing CADe use reported on 3213 patients undergoing colonoscopy for positive faecal immunochemical test (FIT+).16 The authors concluded that CADe does not improve the identification of advanced colorectal adenomas that are associated with higher risk of colorectal cancer and mortality.

We performed a systematic review and meta-analysis of all RCTs, evaluating the impact of CADe-assisted colonoscopy for screening, surveillance, and follow-up of FIT+, on all reported outcomes.15 The publication of the largest RCT,16 along with the systematic review, triggered this guideline. Given the rapidly evolving field of AI, during guideline development, we commissioned an updated systematic review focused on trials assessing efficacy of CADe17 and a separate review that examined patients’ values and preferences (supplemental material). The review, together with a microsimulation study,18 informed the recommendation.

Understanding the recommendation

Uncertainties

• Impact of CADe on CRC incidence and CRC-related mortality—In the absence of RCTs addressing CRC incidence and CRC-related mortality directly, the evidence on effects on these critical outcomes were estimated from a modelling study (for details see “How the guideline was made” below and the MAGICapp). Many of the inputs to the modelling study were informed by data predominantly from a European population, though sensitivity analyses from a North American population did not result in major changes to estimates. Direct evidence to support these outcomes may come from an ongoing long term European Union funded RCT evaluating the impact of CADe on patient-important outcomes.21This trial is slated to complete its follow-up by 2036.

• Unblinded RCTs of CADe systems using adenoma detection rate as the primary outcome—The lack of blinding in almost all trials and the use of outcomes judged by endoscopists raises concerns regarding potential bias. Knowing which group they belong to (CADe-assisted or routine colonoscopy) could influence the endoscopists’ performance and, consequently, the outcomes of the trials. The absence of blinding introduces the possibility that belief in the efficacy of CADe might lead to overreliance on the tool.22

• Generalisability of CADe efficacy across endoscopists—CADe should not be a substitute for careful and thorough examination. Any benefit of CADe is contingent upon the mucosa that is visualised during colonoscopy.2223 The technology requires colonoscopy to be performed by a trained endoscopist who ensures full mucosal visualisation of the colon. It is uncertain if trial findings are generalisable to endoscopists with varying expertise levels. However, to date, we have found no credible evidence to suggest that CADe's efficacy may be influenced by the endoscopist's skill set.

• Lack of transparency in approval of CADe devices for use in practice—Many CADe tools are developed by private companies, and the specific algorithms they use to flag or classify medical images are often proprietary. Because they are proprietary, there’s limited public information about how these algorithms work or how they are trained. In addition, once a product has initial regulatory clearance (for example, through the FDA in the United States), there is often leeway for the developer to update or modify its algorithms without necessarily reapplying for a full regulatory review. This means that over time, the product’s performance could change—potentially for better, but also possibly for worse—without formal regulatory reassessment or transparent reporting of the new clinical evidence.

• Uncertainty in inferences on patients’ values and preferences—Due to the lack of relevant studies on values and preferences (see supplemental material), the panel made inferences about what most patients would want, as outlined in “How the guideline was made.” Future research is needed to better understand the values and preferences of patients undergoing colonoscopy. Given the uncertain evidence, all panel members agreed that the recommendation should be weak. To inform the final recommendation, we asked the panel members to vote on whether to suggest against or for the routine use of CADe for adults undergoing colonoscopy for any indication (symptoms, screening, or surveillance), with one of the following statements to encapsulate the rationale and values and preferences informing the recommendation:

  • “Recommendation against.” The benefits on the critical outcomes of CRC incidence, and post-colonoscopy cancer incidence remain very uncertain. For colorectal cancer-related mortality, the evidence is of low certainty suggesting a trivial or no benefit. The potential burdens—including more frequent surveillance colonoscopies, increased health-related anxiety, and overdiagnosis—are likely to exist for many patients. The uncertainty of any benefits, and the high likelihood that many patients would experience the burdens, led the panel to conclude that majority of well informed patients would not favour CADe assistance. In concluding with a weak recommendation against the routine use of CADe, the panel placed higher value on avoiding the burdens than on the uncertain benefits.

  • “Recommendation for.” Given the uncertain benefits on critical outcomes, the panel believes that majority of well informed patients who have already decided to undergo colonoscopy would favour CADe assistance. This is due to the potential benefits of reduction in colorectal cancer incidence and mortality. The weak recommendation for CADe places greater value on the potential of avoiding colorectal cancer and death, rather than avoiding the potential burden of more frequent surveillance colonoscopies, increased health-related anxiety, and overdiagnosis.

  • Thirteen panel members (60%) voted against, while nine (40%) voted for.

  • Acknowledging the panel’s deliberations—which ultimately yielded a 60%/40% split vote—we arrived at a weak recommendation against CADe. Recommendations must be both actionable and transparent about their rationale, even when evidence is low or very low certainty. The BMJ Rapid Recommendations approach therefore necessitates clarifying how the panel weighed anticipated benefits, harms, burdens, and uncertainties in patient values and preferences. On balance, these considerations led to a weak recommendation against CADe-assisted colonoscopy. We recognise that this may not fully capture the true values and preferences of all patients undergoing colonoscopy, and we will update our guidance as more robust evidence becomes available.

• Microsimulation model using a 10 year follow-up period—We selected a 10 year horizon for our modelling study for three main reasons. First, it reflects the longest and most robust randomised trial data currently available for colonoscopy screening (for example, the NordICC trial6) and aligns with emerging evidence on the relationship between adenoma detection rate and post-colonoscopy colorectal cancer.24 Second, it matches the standard interval recommended by several CRC screening guidelines for average risk individuals (including those from the US Preventive Services Task Force25 and the US Multi-Society Task Force26). Third, follow-up periods shorter than 10 years would likely underestimate the long term impact of colonoscopy on CRC incidence and mortality, whereas extending beyond a decade would require more speculative assumptions about disease progression and adherence of patients to recommended screening or follow-up schedule. We recognise that CRC can have a prolonged latency period, and some panel members expressed concern that 10 years might still be relatively short. Nonetheless, our modelling inputs reflect the best available evidence and common screening intervals for average risk adults. Additionally, the average patient age in our simulations is approximately 60 years, so a 10 year follow-up extends into the early 70s—a period when CRC incidence starts to markedly increase.27 As a living guideline, we plan to update our recommendation if new evidence on longer term outcomes emerges that might alter our understanding of CADe’s impact beyond 10 years.

• Age range in microsimulation study and generalisability—We acknowledge that the microsimulation model informing our estimates of CRC incidence, mortality, and potential harms was based on individuals aged 60 69, while our recommendation applies to all adults ≥18 years. This discrepancy may introduce uncertainty because of population indirectness. However, we did not downgrade our certainty in the conclusions drawn from the microsimulation study for indirectness, as we have no compelling evidence that CADe’s relative efficacy varies by age. Younger adults face an even lower baseline risk of CRC and mortality, suggesting that the absolute effect of CADe would be smaller—further underscoring minimal or no net benefit. Therefore, the panel concluded that the recommendation remains valid for the broader adult population, while acknowledging this as an area of uncertainty that we will revisit if new data emerge.

Implementation and adaptation of the guideline

While the performance of a CADe system is dependent on how it was developed and validated, the actual application and utility of the intervention is reliant on how end users (endoscopists) ultimately engage with it.222328 According to one study, when given a choice, endoscopists “turned on” the CADe system in only 52% of procedures with varying amounts of time spent on repeat mucosal inspection in response to a visual indicator (or bounding box).29 Provider attitude and trust is an important factor in how much and how often CADe is used.

This is a living guideline, and we will continuously monitor for new CADe trials or major updates to existing data. If a new trial indicates a substantially different effect on patient-important outcomes or if technological changes alter CADe’s performance, the microsimulation model will be revised, and the panel will reassess the recommendation. This ensures that our living guideline remains current as evidence evolves. Additional evidence may increase our understanding of individual values and preferences, provider trust and adoption, and cost effectiveness or budget impact. While this guideline focuses on the individual patient level, future updates will take a societal and system level approach and consider formal cost-effectiveness analysis studies accounting for the upfront costs associated with buying (or subscribing) to CADe systems, and the increased burden of colonoscopies at a population level.

This guideline is the result of a collaborative approach to guideline development with MAGIC, the American Gastroenterological Association (AGA),30 and the European Society of Gastroenterology (ESGE) (doi:10.1055/a-2543-0370). By leveraging a shared methodological framework (GRADE) and adhering to Institute of Medicine’s strict standards for trustworthy recommendations,31 we synthesised the most up-to-date evidence to streamline the guideline development process. This collaboration increased efficiency, allowed the sharing of evidence profiles and evidence-to-decision tables, and promoted transparency in panel judgments, with adaptations for the North American and European context.

While our panel used the same evidence base as the AGA and the ESGE, we reached a different conclusion—namely, a weak recommendation against routine CADe. All three guidelines made weak recommendations, with panel members agreeing that the net benefit is uncertain. The key distinction lay in how each panel judged patient values and preferences. Our panel placed a relatively higher weight on avoiding the potential burdens of additional surveillance, overdiagnosis, and anxiety for patients, given minimal or no proven benefit for critical outcomes such as CRC incidence and mortality. In contrast, the AGA and ESGE panels placed a higher value on the potential—though uncertain—benefits of CADe. Although we diverge in our final recommendations, we share a recognition that individual decisions may differ based on how patients and clinicians weigh uncertain benefits versus likely burdens. As a living guideline, we acknowledge that new data—especially from large RCTs addressing CRC incidence, mortality, and overdiagnosis—may lead to revisions in our recommendation, which we will update and refine as more evidence becomes available.