Endoscopic Diagnostic Support System for cT1b Colorectal Cancer Using Deep Learning

Abstract

Objective: This study aimed to use convolutional neural network (CNN), a deep learning software, to assist in cT1b diagnosis. Methods: This retrospective study used 190 colon lesion images from 41 cases of colon endoscopies performed between February 2015 and October 2016. Unenhanced colon endoscopy images (520 × 520 pixels) with white light were used. Images included 14 cTis cases with endoscopic resection and 14 cT1a and 13 cT1b cases with surgical resection. Protruding, flat, and recessed lesions were analyzed. AlexNet and Caffe were used for machine learning. Fine tuning of data to increase image numbers was performed. Oversampling for the training images was conducted to avoid impartiality in image numbers, and learning was carried out. The 3-fold cross-validation method was used. Sensitivity, specificity, accuracy, and area under the curve (AUC) values in the receiver operating characteristic curve were calculated for each group. Results: The results were the average of obtained values. With CNN learning, cT1b sensitivity, specificity, and accuracy were 67.5, 89.0, and 81.2%, respectively, and AUC was 0.871. Conclusion: Quantitative diagnosis is possible using an endoscopic diagnostic support system with machine learning, without relying on the skill and experience of endoscopists. Moreover, this system could be used to objectively evaluate endoscopic diagnoses.