Real-time multi-class detection of colorectal polyps based on the colon-YOLO network

Colorectal adenomatous polyps are key precursors to colorectal cancer (CRC), but their accurate classification during endoscopy remains challenging due to variability in physician expertise and difficulties in detecting certain lesion types, such as sessile serrated adenomas/polyps (SSAP). To address this, we developed Colon-YOLO, a real-time polyp detection network based on YOLOv5, incorporating ConvNeXt for global feature extraction, a SimAM attention mechanism for enhanced 3D feature weighting, and a concentrated feature pyramid attention layer for improved context capture. A novel Soft-SIoU-NMS method was introduced to boost occlusion detection and convergence speed. Evaluated on both remote and edge devices, the model achieved a 6.3 % mAP improvement (IoU=0.5) over YOLOv5, with inference speeds of 120.5 FPS (remote) and 28.57 FPS (edge), meeting real-time clinical needs. This approach enhances polyp detection accuracy, reducing missed diagnoses and supporting CRC prevention.