Real-world computer-aided detection for adenoma detection: promising but context-dependent Authors: Yunho Jung Published: 2026 With great interest, we read the article by Rønborg et al.,1 who evaluated the real-world effect of computer-aided detection (CADe, GI Genius; Medtronic) on adenoma detection rate (ADR)—an outcome previously reported to improve in pooled analyses of randomized controlled trials2—in a prospective observational study with a concurrent comparator group. This design is more rigorous than those of many previous real-world CADe studies that relied on historical controls, and we commend the authors for the pragmatic approach. The authors found only a modest, statistically nonsignificant increase in ADR with CADe use (34.7% vs. 30.5%; risk ratio, 1.12; 95% confidence interval, 0.88–1.43), suggesting that the real-world benefit of CADe may be limited. We would like to raise a few points regarding the study design that may have contributed to this nonsignificant finding. It should be noted that no mucosal exposure device (e.g., Endocuff Vision) was used in either study arm. Both a randomized trial and a real-world retrospective study have shown that combining GI Genius with a mucosal exposure device significantly increases ADR compared with GI Genius alone,3,4 raising the possibility that the absence of such a device may have limited the ceiling of achievable benefit in the present cohort. The CADe module was available for only three months, the same duration used for patient enrollment, providing little time for endoscopists to become familiar with the system and calibrate their reliance on its prompts. In contrast, real-world evaluations in which endoscopists had several months of prior experience with a CADe system before formal assessment have reported significant ADR gains,5,6 suggesting that the degree of prior exposure to a specific system, rather than its mere availability, may be an important and under-reported variable. Moreover, the median Boston bowel preparation scale score was 6, which is generally considered adequate rather than excellent bowel preparation. Since a CADe system can only flag lesions within the mucosa that has already been exposed and adequately cleansed,7 the inclusion of patients with suboptimal bowel preparation may have diluted the apparent benefit of CADe in the overall analysis. Additionally, patients and endoscopists were not blinded to CADe system use, which the authors acknowledge as a potential source of bias. Unblinded, real-world CADe cohorts elsewhere have shown numerically lower ADR with artificial intelligence (AI) assistance, an effect attributed to shortened procedural times and possible overreliance on the system.8 Whether such behavioral compensation contributed to the modest effect size observed in the present study remains an open question. In summary, we congratulate the authors on a rigorous evaluation of CADe that appropriately tempers the expectations set by randomized trials. However, the real-world benefit of CADe may depend as much on the implementation context—including device combination, endoscopists’ cumulative experience with the specific system, bowel preparation quality, and endoscopists’ behavior—as on the underlying detection algorithm itself. Similar concerns regarding implementation and generalizability have been raised for other AI applications in endoscopy, including deep learning-based prediction of lymph node metastasis risk in T1 colorectal cancer, for which generalizability remains limited by dataset quality and single-institution training.9 We agree with Rey10 that realizing the full potential of AI in endoscopy will require moving beyond single-device or single-center evaluations toward larger, multicenter studies that systematically account for these practice-level variables.