Highlights
In this pragmatic cluster randomized quality improvement study across Veterans Health Administration facilities, availability of computer-aided detection (CADe) during colonoscopy was associated with a significant increase in adenoma detection compared with usual care.
The absolute adenoma detection rate (ADR) increased by 4.2 percentage points at CADe sites, from 50.7% to 54.9%, whereas ADR fell slightly at control sites, from 51.8% to 51.1%.
In adjusted mixed-effects analysis, CADe availability increased the odds of adenoma detection by 22% (odds ratio 1.22; 95% confidence interval 1.15-1.28), with no apparent effect on withdrawal time or detection of other lesion types.
The benefit of CADe did not appear confined to lower-performing endoscopists; baseline ADR quintile was not associated with the magnitude of effect.
Background and Clinical Context
Colonoscopy remains the cornerstone of colorectal cancer prevention because detection and removal of adenomas can interrupt the adenoma-carcinoma sequence. Among quality metrics in colonoscopy, adenoma detection rate is particularly important. Higher endoscopist ADR has been consistently associated with lower risks of interval colorectal cancer and death from post-colonoscopy colorectal cancer. For this reason, strategies that improve ADR have direct clinical relevance, even when the intervention does not yet demonstrate cancer outcome data.
Computer-aided detection systems use artificial intelligence, typically deep learning-based image analysis, to identify suspected polyps in real time during colonoscopy. Randomized controlled trials have generally shown that CADe improves adenoma detection, especially for small or subtle lesions. However, a persistent question has been whether these gains translate beyond controlled trial settings into routine practice, where workflow variation, endoscopist heterogeneity, competing operational pressures, and inconsistent uptake may dilute efficacy.
The current study by Dominitz and colleagues addresses that gap. Rather than testing CADe in a tightly controlled explanatory trial, the investigators evaluated the impact of device availability across real-world Veterans Health Administration facilities using a cluster randomized design. This is an important distinction: for health systems and policy leaders, effectiveness in ordinary care is more actionable than efficacy under idealized conditions.
Study Design
Design and Setting
This was a pragmatic quality improvement study using cluster randomization at the facility level within the Veterans Health Administration. Facilities randomly selected for implementation were equipped with CADe devices between October 2022 and February 2023. Control facilities were assigned a random deployment date, allowing pre- and post-period comparisons that help account for secular trends.
Participants and Colonoscopy Volume
Colonoscopy data were analyzed from 269 endoscopists at 42 facilities with CADe and 547 endoscopists at 97 control facilities. At CADe sites, there were 71,594 pre-deployment colonoscopies and 35,399 post-deployment colonoscopies. At control sites, there were 151,792 pre-deployment colonoscopies and 75,415 post-deployment colonoscopies. The scale of the study is a major strength, providing robust estimates across a broad range of operators and practice environments.
Exposure, Comparator, and Outcome Assessment
The exposure of interest was CADe availability at the facility. The comparator was usual care at control facilities during analogous time periods. The primary clinical outcome was adenoma detection during colonoscopy. The authors used mixed-effects logistic regression to estimate the independent association between CADe availability and colonoscopy outcomes, an appropriate analytic strategy given clustering by facility and endoscopist as well as repeated observations over time.
Why This Design Matters
Cluster randomization is especially suitable for interventions deployed at the level of an endoscopy unit rather than an individual patient. It reduces contamination that would occur if the same endoscopist alternated unpredictably between AI-assisted and non-assisted practice. It also better reflects how health systems actually adopt digital tools: by equipping sites, training staff, and embedding technology into workflow.
Key Findings
Primary Outcome: Adenoma Detection Improved
The core result is straightforward and clinically meaningful. At CADe sites, ADR rose from 50.7% before deployment to 54.9% after deployment, an absolute increase of 4.2 percentage points (95% confidence interval 3.48-4.74). By contrast, control sites experienced a small decline in ADR from 51.8% to 51.1%, an absolute change of -0.7 percentage points (95% confidence interval -1.16 to -0.28).
In multivariable mixed-effects logistic regression, CADe availability was independently associated with higher adenoma detection, with an odds ratio of 1.22 (95% confidence interval 1.15-1.28). The confidence interval is narrow and excludes the null, indicating a statistically robust finding. From a practical standpoint, a 22% relative increase in the odds of adenoma detection across a mature national endoscopy system is not trivial. Because ADR is already high in many VA settings, improving it further suggests that CADe may add value even in experienced hands.
Effect Across Endoscopist Performance Levels
An especially notable result is that baseline endoscopist ADR quintiles were not associated with the impact of CADe availability. This suggests that the benefit was not restricted to low detectors. In other words, CADe may function less as a remedial tool for underperformers and more as a broadly useful second observer. That interpretation has operational implications. If validated further, selective deployment only to low-ADR clinicians may underestimate the system-wide benefit.
No Signal for More Nonadenomatous Findings
CADe availability was not associated with increased detection of other lesions. That point matters because a common concern is that AI may simply raise alert frequency without meaningful enrichment for clinically relevant neoplasia, potentially increasing unnecessary polypectomy or pathology use. The available summary suggests the observed gain was specific to adenoma detection rather than a nonspecific increase in all visualized abnormalities.
No Change in Withdrawal Time
The study also found no association between CADe availability and withdrawal time. This is reassuring. If CADe had improved ADR only by substantially prolonging procedures, its scalability and cost-effectiveness would be less attractive. Preserved withdrawal time implies the system may improve inspection efficiency or lesion recognition without materially slowing workflow. Of course, average time metrics do not fully capture procedural burden, but the finding supports feasibility in routine practice.
Clinical Interpretation
These results are important because they address a major implementation question: does CADe still work when introduced at scale outside a conventional randomized efficacy trial? In this study, the answer appears to be yes. The real-world effectiveness signal aligns with prior randomized trials showing improved ADR, and it contrasts with earlier observational reports that failed to demonstrate meaningful benefit after market adoption.
For clinicians, the practical message is that CADe may serve as a genuine quality-improvement adjunct rather than a purely technological novelty. The increase in ADR is clinically relevant because ADR is a validated surrogate quality metric linked to interval cancer risk. At the same time, the study does not establish that CADe reduces colorectal cancer incidence or mortality. That remains an inference based on the established relationship between ADR and cancer outcomes, not a directly measured endpoint here.
For endoscopy unit leaders, the absence of longer withdrawal times is encouraging because implementation barriers often hinge on throughput, room turnover, and staffing burden. A technology that improves a key quality metric without obvious time penalties is easier to justify operationally, assuming acquisition and maintenance costs are acceptable.
Strengths of the Study
The study has several major strengths. First is its pragmatic design, which enhances external validity. Second is its scale: hundreds of endoscopists and more than 330,000 colonoscopies across intervention and control periods. Third is the use of cluster randomization and mixed-effects modeling, which improves causal inference relative to purely uncontrolled before-after analyses. Fourth, the study examines a meaningful clinical process outcome rather than a technical endpoint divorced from patient care.
The Veterans Health Administration setting also offers advantages for implementation research, including standardized electronic data capture and broad geographic reach. This makes the findings especially useful for integrated health systems considering enterprise-level AI deployment.
Limitations and Remaining Questions
Several limitations deserve attention. The first is generalizability. VA populations are predominantly older and male, and practice patterns in an integrated federal system may differ from those in community or private health care settings. The magnitude of benefit may therefore not translate exactly to all populations.
Second, the intervention studied was CADe availability, not necessarily confirmed use during every eligible procedure. In pragmatic implementation studies, that distinction is expected, but it means the reported effect reflects real-world adoption plus performance, not pure device efficacy. That is valuable for health systems, though it can complicate mechanistic interpretation.
Third, the summary does not provide lesion-level details such as adenoma size, morphology, location, advanced histology, or serrated lesion outcomes. Many prior CADe studies have shown that gains are driven largely by small and diminutive adenomas. If this was also true here, the ultimate effect on long-term cancer prevention may be smaller than the ADR increase alone suggests. On the other hand, even small-lesion detection may improve quality consistency.
Fourth, the authors appropriately note that the impact on cancer outcomes remains undetermined. This is not a shortcoming unique to this study; rather, it reflects the long time horizon required to show interval cancer reduction. Similarly, the possibility of endoscopist deskilling remains unresolved. If clinicians become overly reliant on CADe, visual vigilance could theoretically decline when the system is absent or when lesions fall outside algorithmic strengths. That concern is plausible but currently unproven.
How This Fits With Existing Evidence
The broader literature on CADe in colonoscopy has generally shown improved ADR in randomized trials. Multiple meta-analyses have reported higher polyp and adenoma detection with AI assistance, especially for diminutive lesions. What has been less consistent is whether those benefits persist in routine practice, where user behavior, workflow integration, and training variability may erode efficacy.
This study helps bridge that evidence gap. Its findings suggest that the discrepancy between trial efficacy and real-world effectiveness may not be intrinsic to CADe itself, but rather dependent on implementation quality and system context. In a structured health system with organized deployment, measurable gains appear achievable.
Current colonoscopy quality frameworks emphasize high ADR, careful mucosal inspection, adequate bowel preparation, and complete resection. CADe should probably be viewed as an adjunct to, not a replacement for, those fundamentals. It does not correct poor bowel prep, incomplete examination, or suboptimal therapeutic technique. Nor does it substitute for thoughtful surveillance interval decisions.
Implications for Practice and Policy
From a practice standpoint, the study supports consideration of CADe as a quality-improvement tool in colonoscopy programs seeking to improve ADR. Health systems may reasonably ask which settings stand to benefit most, whether universal deployment is preferable to targeted implementation, and how performance should be monitored after rollout. These questions remain open, but the present data support meaningful effectiveness at the system level.
From a policy perspective, the study strengthens the case for evaluating AI tools using pragmatic designs rather than relying exclusively on pre-market performance studies. Real-world deployment studies should assess not only detection metrics but also downstream pathology utilization, surveillance intensity, patient outcomes, workflow effects, and cost-effectiveness. Those data will be essential for reimbursement and procurement decisions.
Conclusion
In this large cluster randomized study within the Veterans Health Administration, availability of computer-aided detection during colonoscopy was associated with a significant improvement in adenoma detection compared with usual care. The adjusted effect size was substantial, the result was consistent across endoscopist baseline ADR strata, and there was no evident increase in withdrawal time or nonadenomatous lesion detection.
For clinicians and health systems, the main message is that CADe can improve a validated colonoscopy quality metric under real-world conditions. However, whether this translates into fewer interval cancers, lower colorectal cancer mortality, better cost-effectiveness, or unintended operator deskilling remains to be determined. The next phase of evidence generation should move beyond detection metrics toward patient-centered and system-level outcomes.
Funding and Trial Registration
ClinicalTrials.gov: NCT05888623.
The abstract provided does not specify funding details. Readers should consult the full Gastroenterology publication for complete funding and disclosure information.
References
1. Dominitz JA, Gawron AJ, McKee GB, Hoggatt KJ, Kaltenbach T. Impact of Availability of Computer-Aided Detection Devices on Adenoma Detection During Colonoscopy: A Cluster Randomized Study. Gastroenterology. 2026-06-05. PMID: 42250891.
2. Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014;370(14):1298-1306.
3. 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(2):512-520.e7.
4. Wang P, Berzin TM, Brown JRG, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut. 2019;68(10):1813-1819.
5. Hassan C, Wallace MB, Sharma P, et al. New artificial intelligence system: first validation study versus experienced endoscopists for colorectal polyp detection. Gut. 2020;69(5):799-800.