A Real-time, Multi-functional Intestinal Endoscopy AI Detection System

Applying deep learning techniques to computer-assisted colonoscopy diagnosis significantly enhances polyp detection accuracy, thereby contributing to the prevention of colorectal cancer. While prior studies have predominantly focused on pathological tissues during colonoscopy, they often neglect potential interferences such as foam and residue. We collected over 20,000 intestinal endoscopic images and annotated polyps, foam and residue, and applied data augmentation based on the morphological characteristics of these objects. We modified YOLOv5 by incorporating the SE-Block into the backbone and using EIoU as the bounding box regression loss, which highly improved the detecting performance on irregularly shaped objects. Besides, We integrated components for entry-exit detection and speed tracking based on image feature analysis into the model, thus developing a real-time, multi-functional intestinal endoscopy detection system.