Deep convolutional neural networks for recognition of atrophic gastritis and intestinal metaplasia based on endoscopy images

Patients with atrophic gastritis (AG) or intestinal metaplasia (IM) are at risk for gastric adenocarcinoma. The aim of the study was to explore the accuracy of a deep convolutional neural network (CNN) utilizing endoscopic images for detection of AG and IM. Endoscopic white-light images of pathologically confirmed as atrophic gastritis and/or intestinal metaplasia were collected. A pre-trained CNN model (ResNet-50) was fined-tuned using a training cohort. Diagnostic accuracy was evaluated in an independent testing cohort. Of 973 patients, 779 (80%) including 541 (69%) with AG/IM patients (3007 images) were assigned to the training cohort, 96 (10%) including 63 (66%) with AG/IM patients (375 images) were assigned to the validation cohort, and 98 (10%) including 67 (68%) with AG/IM patients (377 images) were assigned to testing cohort. The area under the curve (AUC) for detection of AG was 0.96 with sensitivity, specificity, and precision of 87.2%, 91.1%, and 85.4%, respectively. AUC for detection of IM was 0.98 with sensitivity, specificity, and precision of 90.3%, 93.7% and 90.9%, respectively. The CNN system based on endoscopic white-light images achieved high diagnostic performance for detection of atrophic gastritis and intestinal metaplasia.