TRUE-Colon reveals why colonoscopy AI fails in real-world procedures

Computer-aided detection systems for colonoscopy promise to reduce missed polyps, but their real-world performance is often overestimated by lesion-centric benchmarks. The TRUE-Colon protocol, created by Sebastian Doerrich and colleagues and accepted to EndoLINA @ MICCAI 2026, systematically exposes this gap. The researchers benchmarked four real-time architectures—Faster R-CNN, YOLOv8, YOLOv11, and RT-DETR—against both curated datasets (SUN, PICCOLO) and 60 unedited, full-length procedures from the REAL-Colon dataset. The core finding is a consistent transfer asymmetry: models trained strictly on curated clips suffer a severe performance collapse when evaluated on full procedures, while procedure-trained models actually improve rejection of non-polyp content on REAL-Colon and largely retain accuracy on curated benchmarks. This indicates that training on curated clips inflates success metrics and misses clinically crucial long negative stretches and procedure-related artifacts.

Beyond transferability, the study details architecture-specific trade-offs. The Transformer-based RT-DETR detector achieved the strongest sensitivity and earliest, most persistent detections, while convolutional detectors like YOLOv8 and YOLOv11 remained competitive with higher throughput—a critical factor for real-time colonoscopy. The authors argue that training and benchmarking should shift from curated, lesion-centric clips toward full-procedure data and deployment-relevant operating points. By releasing source code, TRUE-Colon provides a standardized way for researchers to measure deployment readiness beyond localization accuracy. For clinicians and AI developers, this is a clear signal that current validation methods may be overstating performance, and that real-world deployment requires rethinking data selection and evaluation metrics to actually reduce clinical miss rates.