MEA-YOLO: Medical Enhanced Attention YOLO for Simultaneous Polyp and Lumen Detection in Robotic Colonoscopy

Robotic colonoscopy requires accurate simultaneous detection of polyps and lumens for autonomous navigation. However, existing object detectors lack medical domain-specific designs. We present MEA-YOLO (Medical Enhanced Attention YOLO), a unified framework for simultaneous polyp and lumen detection that integrates domain-specific attention mechanisms tailored for colonoscopy imaging. Our novel C2f-MEA module combines efficient multi-scale attention with specialized edge and lumen enhancement branches, addressing the unique visual characteristics of colonoscopic images. Evaluated on the Extended-LDPolyp dataset (37,947 images with dual annotations, video-level split), MEA-YOLO achieves 74.52% mAP@0.5, outperforming YOLOv8n by 2.58% and Faster R-CNN by 41.65%. Task-specific evaluation shows 72.52% AP for polyp detection and 76.52% AP for lumen detection, with consistent improvements in both precision and recall. Ablation studies confirm that medical-specific enhancements are crucial-while generic attention mechanisms degrade performance, our domain-aware design contributes 2.58% mAP improvement. Despite adding only 6.25% parameters overhead, MEA-YOLO maintains real-time performance at 111.5 FPS, demonstrating that architectural innovations guided by clinical insight can achieve superior accuracy without sacrificing efficiency for robotic colonoscopy deployment.