A U-Net Based Lesion Segmentation Method for Computer-Aided Diagnosis in Colorectal NBI Endoscopy

In this paper, we propose a method based on deep learning architecture, U-Net, for detecting lesion from images token by colorectal Narrow Band Imaging (NBI) endoscopy. The proposed method is a part of a Computer-Aided Diagnosis (CAD) system that presents quantitative inference results to doctors when making diagnoses with NBI endoscopy, which aims to reduce the variation and burden of diagnoses due to the experience of diagnosing doctors. As a result, for our test dataset, the mean F-measure score has exceeded 80%.