Automatic Polyp Segmentation in Colonoscopy Images Using Single Network Model: SegNet

Abstract

Colorectal cancer is the third most common diagnosed cancer worldwide. Early detection and removal of adenoma during the colonoscopy examination may increase the survival probability. A novel computer-aided tool for automated polyp segmentation in colonoscopy images is described in this work. SegNet, a deep convolutional neural networks has been chosen to map low resolution features with the input resolution for automated pixel-wise semantic polyp segmentation. Publicly available databases, CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB were used to train and to test the model. The outcome demonstrated the proposed method is feasible as it attains an average of 81.78, 92.35% for mean intersection over union, and dice coefficient, respectively for testing on a combination of the aforementioned datasets.