공개 용종 검출 데이터셋의 임상적 크기 편향성: 소형 용종 검출 과제에 대한 체계적 재평가

초록

AI-based colonoscopic polyp detection models report high benchmark recall rates, yet clinical performance—particularly for small polyps—often falls short. We investigated whether the structural size bias inherent in public training datasets underlies this performance gap. A total of 6,222 polyp images from five public datasets (Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS-LaribPolypDB, and the AI-Hub Korean colonoscopy dataset) were analyzed. Relative polyp size was defined as the bounding-box-to-image area ratio (%). Small-polyp subsets (≤3% and ≤1% of total image area) served as exclusive test sets. Detection performance was compared between Convolutional Neural Network (CNN) based models (YOLOv4, YOLO-OB, YOLO26-L) and transformer-based models (RT-DETR, RF-DETR) fine-tuned on small polyps. Polyps with an area ratio <0.75% comprised only 3.1% of the integrated dataset and 2.3% of Kvasir-SEG. On small-polyp-specific test sets, CNN-based model recall declined significantly compared with overall benchmark performance, whereas transformer-based models fine-tuned on small polyps demonstrated relatively higher recall under limited conditions. The high recall rates in existing public benchmarks likely reflect data compositions skewed toward large polyps. Most polyp detection models failed to achieve optimal performance on small-polyp-exclusive validation sets, suggesting that benchmark-based evaluations do not adequately reflect the clinically important ability to detect small polyps.

키워드