Deep Learning Instance Segmentation–Driven Detection of Colonic Polyps in Pediatric Colonoscopy

Early and accurate detection of colonic polyps is essential for preventing the progression of colorectal disease, particularly in pediatric patients where lesion characteristics and data availability differ from adult populations. This study presents a deep learning–based instance segmentation framework for automated detection and precise localization of colonic polyps in pediatric colonoscopy images. A pediatric-focused dataset was curated, annotated, and augmented to support robust model training and evaluation. The proposed model achieved strong quantitative performance, attaining a Box mAP@0.50 of 0.966 and a Mask mAP@0.50 of 0.974, with high precision, recall, and a peak F1 score of approximately 0.96, indicating a balanced trade-off between detection sensitivity and reliability. Confusion matrix analysis confirmed effective discrimination between polyp and background regions, correctly identifying approximately 95% of true polyp instances with minimal false positives. The model maintained computational efficiency with a low inference cost of 9.6 GFLOPs and low end-to-end latency, supporting real-time clinical deployment. Qualitative evaluations on unseen images further demonstrated accurate polyp segmentation, while a web-based Gradio application showcased practical real-time inference and accessibility. The proposed approach highlights the potential of instance segmentation–driven deep learning systems to assist pediatric gastroenterologists in early polyp detection and enhance diagnostic confidence during colonoscopy procedures.