Development and validation of a convolutional neural network model for diagnosing Helicobacter pylori infections with endoscopic images: a multicenter study.

ABSTRACT

BACKGROUND AND AIMS: Insufficient validation limits the generalizability of deep learning in diagnosing Helicobacter pylori infection with endoscopic images. The aim of this study was to develop a deep learning model for the diagnosis of H pylori infection using endoscopic images and validate the model with internal and external datasets. METHODS: A convolutional neural network (CNN) model was developed based on a training dataset comprising 13,403 endoscopic images from 952 patients who underwent endoscopy at Seoul National University Hospital Gangnam Center. Internal validation was performed using a separate dataset comprised of images of 411 individuals of Korean descent and 131 of non-Korean descent. External validation was performed with the images of 160 patients in Gangnam Severance Hospital. Gradient-weighted class activation mapping was performed to visually explain the model. RESULTS: In predicting H pylori ever-infected status, the sensitivity, specificity, and accuracy of internal validation for people of Korean descent were .96 (95% confidence interval [CI], .93-.98), .90 (95% CI, .85-.95), and .94 (95% CI, .91-.96), respectively. In the internal validation for people of non-Korean descent, the sensitivity, specificity, and accuracy in predicting H pylori ever-infected status were .92 (95% CI, .86-.98), .79 (95% CI, .67-.91), and .88 (95% CI, .82-.93), respectively. In the external validation cohort, sensitivity, specificity, and accuracy were .86 (95% CI, .80-.93), .88 (95% CI, .79-.96), and .87 (95% CI, .82-.92), respectively, when performing 2-group categorization. Gradient-weighted class activation mapping showed that the CNN model captured the characteristic findings of each group. CONCLUSIONS: This CNN model for diagnosing H pylori infection showed good overall performance in internal and external validation datasets, particularly in categorizing patients into the never- versus ever-infected groups.