Real-Time Polyp Detection Using an Enhanced YOLOv12 Framework With Adaptive Feature Routing and Auxiliary Self-Distillation

Abstract

Early detection of gastric and colorectal polyps during endoscopic procedures is essential for preventing cancer progression; however, manual identification remains challenging due to subtle lesion appearance, size variability, and the fast pace of clinical workflows. YOLO-based object detection models have demonstrated strong potential for real-time computer-aided detection. Still, limitations persist in multiscale representation, small polyp detection, and maintaining accuracy without increasing computational cost. To address these challenges, we developed an enhanced YOLOv12s-based imaging framework tailored for endoscopic polyp detection. The proposed system integrates three key innovations: (1) RouteHead, an adaptive multiscale feature routing mechanism that dynamically fuses features from multiple feature-pyramid levels (P2-P5), enabling improved detection of small and fine-grained lesions; (2) auxiliary detection heads employed during training for self-distillation, enhancing representational learning without affecting inference-time performance; and (3) task-specific loss functions, including α-CIoU for precise localization, edge-aware loss for boundary refinement, and self-distillation loss for deeper feature consistency. The model was evaluated on a comprehensive multicenter colonoscopy dataset comprising 2188 images aggregated from four publicly available benchmarks: Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB. For datasets lacking bounding box annotations, YOLO-format labels were systematically generated from segmentation masks to ensure consistency in annotations across datasets. Training was performed using a stratified data partitioning strategy with a held-out independent test set and fivefold cross-validation to obtain robust performance estimates. The enhanced model achieved a mAP@50–95 of 0.6976, a precision of 0.9300, a recall of 0.9069, and an F1-score of 0.9182, consistently matching or exceeding the baseline YOLOv12s architecture across validation folds and achieving detection accuracy statistically comparable to that of substantially larger detectors at a fraction of their computational cost. Despite improved accuracy, the proposed framework remains computationally efficient with 8.81 million parameters and 7.92 GFLOPs. By combining adaptive multiscale feature routing with self-distillation-based training, the proposed imaging system effectively overcomes key limitations of existing real-time endoscopic detection approaches. The strong balance between detection accuracy, robustness across multiple datasets, and computational efficiency makes the framework suitable for practical computer-aided endoscopy systems, with the potential to enhance early detection rates and reduce cognitive burden on clinicians during routine examinations.