A Lesion Classification Method Using Deep Learning Based on JNET Classification for Computer-Aided Diagnosis System in Colorectal Magnified NBI Endoscopy

In this study, we develop and evaluate a colorectal endoscopic diagnosis support system using deep learning. We use image dataset about the JNET (The Japan NBI Expert Team) classification based on the findings classification of the colorectal magnified endoscopic images. The JNET classification is classified into four types and taken by clinical doctors at three magnifications. We use low to medium magnification and high magnification rate in this study. In addition, we use a learned CNN called ResNet34 as a classifier. As the classification results for 80 test data, the accuracies of low to medium magnification and high magnification were 95.0% and 97.5%, respectively, and 3 images were classified as the wrong type.