GI Genius AI endoscopy boosts adenoma detection in largest VA study

Cosmo Pharmaceuticals' GI Genius AI-assisted endoscopy system improved adenoma detection in a large real-world study across U.S. Veterans Health Administration facilities, leading to system-wide adoption of computer-aided detection (CADe). For AI builders working on medical imaging or clinical decision support, the study provides evidence that AI tools can deliver consistent benefits when deployed at scale in a large health system.

Real-world evidence from 334,000 colonoscopies

The CADeNCE study analyzed more than 334,000 routine colonoscopies across 139 VA facilities in what researchers describe as the largest cluster-randomized quality improvement study of computer-aided detection in colonoscopy. Centers using the GI Genius system recorded a 22% higher odds of adenoma detection, with the adenoma detection rate (ADR) rising from 50.7% to 54.9%. The odds ratio was 1.22 with a 95% confidence interval of 1.15 to 1.28. Importantly, the improvement was consistent across endoscopists with different experience levels, suggesting the tool works in routine clinical practice, not just in controlled trial conditions.

Following the study, the Veterans Health Administration expanded computer-aided detection across all colonoscopy facilities, a strong institutional endorsement. The GI Genius system, developed by Cosmo and commercialized by Medtronic, provides real-time computer-aided detection of colorectal polyps during colonoscopy.

Why this matters for AI deployment in medicine

For AI builders, the VA study is a case study in real-world evidence gathering and institutional rollout. The scale (334,000 procedures across 139 sites) demonstrates that AI-assisted detection can be integrated into a nationwide health system without requiring uniform physician expertise or specialized settings. The consistent benefit across experience levels addresses a common concern about AI tools: that they only help inexperienced practitioners. Here, even experienced endoscopists saw gains.

This also matters for product teams building AI tools for regulated medical environments. The VA's decision to expand CADe system-wide suggests that payers and large health systems are willing to commit to AI when real-world data supports both clinical benefit and operational feasibility. For teams building similar tools, the study reinforces the importance of cluster-randomized or pragmatic trial designs that reflect actual workflow conditions.

Practical implications for health systems and builders

Health systems evaluating AI-assisted endoscopy should consider deployment workflows, training, and ongoing monitoring to replicate these results. The study design itself is instructive: it compared ADR at sites where CADe was available versus sites without it, rather than randomizing individual procedures. That pragmatic approach made the study feasible at scale but also introduces some caveats about causation.

For AI builders, the study highlights the need to measure not just model accuracy but also workflow integration and physician adoption. The GI Genius system's ability to fit into existing colonoscopy procedures without disrupting throughput was likely a factor in its uptake.

Caveats and open questions

The study reports an association between GI Genius availability and higher adenoma detection, but as a cluster-randomized quality improvement study, it cannot fully rule out confounding factors such as site-level differences in patient populations or endoscopist expertise. The odds ratio and ADR changes depend on study design specifics that should be confirmed with the full peer-reviewed publication.

Additionally, translating higher adenoma detection to improved patient outcomes requires long-term follow-up. Detecting more adenomas is a surrogate endpoint; the ultimate goal is reducing colorectal cancer incidence and mortality. Real-world studies like CADeNCE provide strong signals but not final proof of benefit.

Finally, the evidence here is based on a single vendor's system in a single health system. Generalizability to other CADe systems, different patient populations, or non-VA settings remains an open question. AI builders should treat these results as promising but not universally transferable without validation in their target deployment context.