Comparative analysis and application of deep learning polyp detection in colonoscopy

Colonoscopy is the most common method for early diagnosis of colorectal cancer. The traditional detection method mainly relies on manual detection by doctors, and the detection efficiency is low. By contrast, the deep learning method has high detection precision and accuracy, which is an ideal tool for the disease detection of colonoscopy. Based on this, this paper compares the six models of YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x, Faster RCNN with and without "fpn" structure, and finds out the most effective colonoscopy polyp detection among the six models. Comparing and integrating various factors, it is found that the YOLOv5l model has the best comprehensive detection effect, and its mAP value is generally higher than 98.5%, which is the largest mAP value among the above six models, and the detection time is relatively short, about 0.018s per image. The balance of accuracy and rapidity in colonoscopy can be achieved.