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

Area of Science:

• Medical imaging analysis
• Computer-aided diagnosis
• Gastroenterology

Background:

• Colorectal cancer is a leading global malignancy.
• Accurate polyp detection during colonoscopy is crucial for early screening.
• Challenges include diverse polyp morphologies, indistinct boundaries, low contrast, and imaging artifacts, hindering real-time clinical assistance.

Purpose of the Study:

• To develop a high-precision, lightweight polyp detection framework for real-time colonoscopy.
• To improve the trade-off between detection accuracy and computational efficiency.
• To address challenges in detecting polyps with complex backgrounds and ambiguous boundaries.

Main Methods:

• Proposed BCP-YOLO framework based on an improved YOLOv8 architecture.
• Integrated BiFormer module in the backbone to enhance salient region focus and suppress noise.
• Incorporated CARAFE upsampling operator for refined lesion boundaries and spatial details.
• Utilized PConv module for network efficiency optimization.

Main Results:

• Achieved 88.5% mAP on Kvasir-SEG, a 3.4% improvement over YOLOv8 baseline.
• Increased precision by 5.5% and recall by 1.3%.
• Model has 11.7M parameters and achieves 104.1 FPS inference speed.
• Demonstrated strong generalization and robustness via five-fold cross-validation.

Conclusions:

• BCP-YOLO offers a high-accuracy, deployable solution for real-time computer-aided diagnosis in colonoscopy.
• The framework has significant potential to enhance the reliability and efficiency of early colorectal cancer screening.