Semantic Segmentation of Polyps in Colonoscopy Images Using U-Net

Abstract— One of the main causes of cancer-related fatalities globally is colorectal cancer. Colorectal polyps can be removed and early discovery can stop cancer from developing. Nevertheless, manual polyp detection during a colonoscopy is laborious and prone to mistakes made by people. This work explores the automatic detection of polyps from colonoscopy pictures using deep learning. Using the CVC-Clinic DB dataset, which comprises 612 polyp image frames taken from colonoscopy movies and matching binary mask images displaying the polyp regions, we trained a U-Net convolutional neural network. Images of normal colon architecture and polyps are included in the dataset. The adenoma diagnosis rate of physicians could be enhanced via automated polyp detection, and reduce polyp miss rates during colonoscopy. Our deep learning approach shows promise for developing an intelligent decision support system to aid endoscopists in the early diagnosis of precancerous polyps. This could improve early colorectal cancer prevention and reduce cancer incidence and mortality.