Multi-Dataset Training Strategy for Robust Polyp Detection With Clinical Validation Insights

Abstract

Colorectal (CRC) represents one of the leading global causes of cancer-related deaths since polyps function as original precursors to cancer development. The identification of colorectal polyps at an early stage directly contributes to lowering CRC incidence while giving patients improved survival outcomes. The proposed research utilizes YOLOv8m through YOLOv11m state-of-the-art YOLO architectures as an advanced framework to enhance real-time detection and classification of polyps. The research used four different colorectal database sets including Kvasir-SEG and CVC-300, CVC-ClinicDB, CVC-ColonDB for training and validating the models across a wide spectrum of polyp image variations. Following 300 training epochs YOLOv8m achieved the best performance with precision at 92.4% and recall at 85.4% while reaching an F1-Score of 88.7% and mAP@0.5 at 91.1%. The testing of YOLOv8m's generalization skills included an external validation on ETIS-LaribPolypDB where it demonstrated reliable detection performance and high accuracy levels. YOLOv8m demonstrates a strong capability for real-time polyp detection in endoscopic examinations through these performance results which showcase its potential to enhance diagnostic support.