Real-time You Only Look Once-based colorectal polyp detection for cancer screening across independent colonoscopy datasets
Eurasian Journal of Medicine and Oncology · 2026-07-16 · 026220238
DOI: 10.36922/ejmo026220238 ↗Abstract
Introduction: Colorectal cancer remains a major cause of cancer-related mortality, although many cases are preventable when precursor polyps are detected and removed during colonoscopy. Real-time computer-aided detection may assist endoscopists, but reliable localization is challenging because polyps vary in size, shape, color, contrast, boundary clarity, and visibility across heterogeneous endoscopic conditions. Objective: This study benchmarked recent You Only Look Once (YOLO)-based detectors for single-class colorectal polyp localization by comparing architectural families and model scales ranging from nano to x-large. Methods: YOLO12, YOLOv13, and YOLO26 models were evaluated using PolypGen2021, Polyp Imaging Classification for Colonoscopy using Optical Learning Outcomes (PICCOLO), and ETIS-Larib. PolypGen2021 was split at the patient level for model development; the predefined patient-level PICCOLO split was preserved, and ETIS-Larib was reserved exclusively for external testing. Performance was assessed using precision, recall, mean average precision (mAP) 50, mAP50–95, parameter count, giga floating-point operations per second (GLOPs), and inference time. Results: On ETIS-Larib, YOLO26x achieved the highest mAP50–95 (0.6475) and the highest recall (0.7736). On the PICCOLO test set, YOLOv13x achieved the highest mAP50–95 of 0.6630, whereas YOLO26x reached the highest precision of 0.9118. Across both test sets, YOLO26x achieved the highest average mAP50–95 (0.6462). YOLO26n provided the most efficient profile, with 2.3750 million parameters, 5.1894 GFLOPs, and inference times below 3 ms per image. Conclusion: YOLO26x was the strongest accuracy-oriented detector, while YOLO26n offered the most practical lightweight alternative for real-time colonoscopy support.
EUVIMED