BCP-YOLO: a lightweight and interpretable YOLOv8-based framework for real-time coloreactal polyp detection

Colorectal cancer is one of the most prevalent malignant tumors worldwide. Early screening relies on accurate polyp detection during colonoscopy. Polyps in colonoscopic images exhibit diverse morphologies, indistinct boundaries, and low contrast. Specular reflections, fold occlusions, and imaging artifacts further complicate detection, which fail to meet the requirements of real-time clinical assistance.

To address these challenges, we propose BCP-YOLO (You Only Look Once), a high-precision, relatively lightweight polyp detection framework built upon an improved YOLOv8 architecture, designed to achieve a well-balanced trade-off between detection accuracy and computational efficiency. First, to mitigate complex background interference and improve small polyp detection, a BiFormer module is integrated into the backbone network to enhance focus on salient polyp regions while suppressing noise. To alleviate boundary ambiguity, the CARAFE content-aware upsampling operator is incorporated into the feature fusion stage, to refine lesion boundaries and spatial details. PConv module is employed to optimize network efficiency, reducing computational cost while maintaining detection performance.

Experimental results on the Kvasir-SEG and CVC-ClinicDB datasets demonstrate that BCP-YOLO achieves a mean average precision (mAP_0.5) of 88.5% on Kvasir-SEG, representing a 3.4% improvement over the YOLOv8 baseline. Precision and recall increase by 5.5% and 1.3%, respectively. The model contains 11.7 M parameters and achieves an inference speed of 104.1 frames per second (FPS). Five-fold cross-validation on both datasets validates its strong generalization capability and robustness.

The method provides a high-accuracy and deployable solution for computer-aided diagnosis in real-time colonoscopy, offering significant potential to improve the reliability and efficiency of early colorectal cancer screening.