Artificial intelligence (AI)-assisted colonoscopy did not significantly improve adenoma detection compared with high-definition white-light endoscopy (HD-WLE) in patients with Lynch syndrome undergoing surveillance at expert centers, according to an international randomized controlled trial published in The Lancet Gastroenterology & Hepatology. The computer-aided diagnosis (CADx) system also did not outperform expert endoscopists in differentiating neoplastic from non-neoplastic colorectal lesions.
Study design and primary findings
The multicenter, open-label CADLY2 superiority trial enrolled adults aged 18 years or older with genetically confirmed Lynch syndrome who were scheduled for surveillance colonoscopy at nine hereditary cancer surveillance centers in Belgium, Germany, the Netherlands, and Spain. Patients were randomly assigned 1:1 to undergo HD-WLE alone or HD-WLE with computer-aided detection (CADe; CAD EYE, Fujifilm). The AI system was used during withdrawal to identify suspected lesions, while its CADx function was evaluated after lesion detection for optical characterization. The primary endpoint was adenoma detection rate, defined as the proportion of patients with at least one histopathologically confirmed adenoma.
Among 757 randomly assigned patients, 733 were included in the full analysis set, including 369 in the HD-WLE group and 364 in the AI-assisted group. The median age was 49 years in the HD-WLE group and 50 years in the AI-assisted group. Baseline characteristics, including pathogenic mismatch repair gene variants, previous colorectal cancer, surveillance intervals, and prior colonoscopies, were balanced between groups.
The adenoma detection rate was 31% with HD-WLE vs. 34% with AI assistance, a difference that was not statistically significant. No significant differences were observed for adenomas per colonoscopy, proximal or distal adenoma detection, flat adenomas, advanced adenomas, sessile serrated lesions, hyperplastic polyps, serrated polyps, colorectal cancer detection, or overall polyp detection. The researchers wrote that because the trial was designed as a superiority study, the absence of a statistically significant difference should not be interpreted as evidence of equivalence between CADe-assisted colonoscopy and HD-WLE.
Exploratory analyses restricted to the largest recruiting centers, smaller recruiting centers, and individual endoscopists were consistent with the primary findings. Variation in adenoma detection across study centers and endoscopists was observed, but exploratory modeling found no evidence that individual endoscopists modified the effect of AI assistance.
Optical diagnosis performance
The study also evaluated the CADx system for real-time optical differentiation of colorectal lesions. When sessile serrated lesions and traditional serrated adenomas were classified as non-neoplastic according to the prespecified analysis plan, the system achieved 86% sensitivity and 91% specificity. Expert optical diagnosis using the Narrow-band Imaging International Colorectal Endoscopic (NICE) and Japan Narrow-band Imaging Expert Team (JNET) classifications demonstrated higher sensitivity of 90% with similar specificity of approximately 92%. When sessile serrated lesions and traditional serrated adenomas were instead classified as neoplastic in a sensitivity analysis, diagnostic sensitivity declined for both the AI system and endoscopists.
Safety outcomes
Procedure-related outcomes were similar between groups. Mean procedure time was 26.5 minutes in both groups, and patient comfort was comparable. Three adverse events occurred in the AI-assisted group, including two mild post-polypectomy bleeding events and one pulmonary embolism or deep venous thrombosis that researchers determined was unrelated to the procedure and attributed to previously undiagnosed ovarian cancer. No adverse events occurred in the conventional colonoscopy group.
Study limitations and interpretation
The researchers wrote that the study was conducted at expert centers with highly experienced endoscopists, which might restrict the generalizability of the findings to settings with lower baseline detection rates, although Lynch syndrome surveillance is likely to be performed at dedicated centers across most Western countries. Additional limitations included local rather than centralized pathology review and the use of histopathology, rather than CADx, as the reference standard for clinical decision-making.
First study author Robert Hüneburg, MD, of the National Center for Hereditary Tumor Syndromes at University Hospital Bonn in Germany, said the findings should not be interpreted as showing that AI is ineffective in colonoscopy or in Lynch syndrome surveillance more broadly. “Rather, they suggest that the incremental benefit of current-generation CADe systems may be modest when baseline procedural quality is already high and examinations are performed by endoscopists experienced in Lynch syndrome surveillance,” he said.
Dr. Hüneburg said the findings should be interpreted in the context of the trial’s high-quality surveillance environment, where the control-group adenoma detection rate was already high, advanced adenomas were detected in 9.2% of patients and mean withdrawal time was approximately 15 minutes.
Clinical implications
Dr. Hüneburg said the study reinforces that meticulous mucosal exposure, adequate inspection time, high-quality bowel preparation, and experience with the subtle and often flat lesions seen in Lynch syndrome remain the foundations of effective surveillance.
“CADe can be used as an adjunctive tool, but currently available systems should not be considered a substitute for high-quality colonoscopy or careful inspection technique,” he said. “Our findings suggest that routine addition of the CADe system evaluated in CADLY2 is unlikely to produce a substantial increase in adenoma detection in expert Lynch syndrome surveillance programs.”
For patients, Dr. Hüneburg said, the results are reassuring in that conventional high-quality colonoscopy performed in an experienced surveillance program remains an appropriate standard of care. “Access to a current-generation CADe system should not determine whether a patient is considered to have received adequate Lynch syndrome surveillance,” he said. “However, AI might still offer greater value in settings with lower baseline detection rates, shorter inspection times, or greater variability between endoscopists. Future systems trained specifically on Lynch syndrome-associated lesions might also perform differently from the system evaluated in CADLY2.”
Dr. Hüneburg said the CADx component of the study also showed that current systems are not yet sufficiently reliable to replace histopathological assessment in Lynch syndrome, particularly because both AI and expert endoscopists had difficulty reliably differentiating sessile serrated lesions. He added that the observed variation between centers and endoscopists highlights the importance of continuously monitoring local colonoscopy quality.
Interpretation and next steps
“The study evaluated one currently available AI platform in specialized expert centers,” Dr. Hüneburg said. “The results should not be extrapolated to all current or future AI systems. Systems trained specifically on the flat, subtle, proximal and serrated lesion phenotypes characteristic of Lynch syndrome might achieve different results.”
Dr. Hüneburg said further studies should evaluate CADe in a broader range of clinical environments and prospectively examine whether baseline endoscopist performance modifies the benefit of AI. “Longer-term research is also needed to determine whether AI affects clinically meaningful outcomes such as post-colonoscopy colorectal cancer incidence rather than adenoma detection alone,” he said.
Expert insight
GI & Hepatology News invited Alex Zhornitskiy, MD, assistant clinical professor in the Division of Gastroenterology and Hepatology, Department of Internal Medicine, at UC Davis Medical Center, who was not involved in the study, to comment on the findings.
“Both study groups had long mean withdrawal times (>14 minutes) and procedures were performed by experienced endoscopists at expert referral centers, likely providing little opportunity for AI to provide additional benefit,” Dr. Zhornitskiy said. “These findings suggest that in the setting of a high-quality exam by an experienced provider, CADe currently offers limited additional value even in high-risk populations such as Lynch syndrome undergoing surveillance colonoscopies — although its utility in less experienced settings or routine clinical practice remains an important area for investigation.”
Robert Hüneburg, MD, reported receiving research funding from Deutsche Krebshilfe and Wilhelm Sander-Stiftung; speaker honoraria from Falk Pharma, Fujifilm, and MSD; and endoscopic equipment on loan from Fujifilm. Other researchers reported research funding, consulting fees, speaker honoraria, equipment loans, advisory roles, or other relationships with Fujifilm, Medtronic, Norgine, Olympus, Pentax, and other commercial entities. The remaining researchers reported no competing interests.
The study was supported through third-party research funding of the National Center for Hereditary Tumor Syndromes at University Hospital Bonn. The CAD EYE system was provided by Fujifilm. Fujifilm had no role in the study design, conduct, data analysis, interpretation, or reporting.