Primary Investigation of Deep Learning Models for Japanese "Group Classification" of Whole-Slide Images of Gastric Endoscopic Biopsy.

Background: Accurate pathological diagnosis of gastric endoscopic biopsy could greatly improve the opportunity of early diagnosis and treatment of gastric cancer. The Japanese "Group classification" of gastric biopsy corresponds well with the endoscopic diagnostic system and can guide clinical treatment. However, severe shortage of pathologists and their heavy workload limit the diagnostic accuracy. This study presents the first attempt to investigate the applicability and effectiveness of AI-aided system for automated Japanese "Group classification" of gastric endoscopic biopsy. Methods: In total, 260 whole-slide images of gastric endoscopic biopsy were collected from Dalian Municipal Central Hospital from January 2015 to January 2021. These images were annotated by experienced pathologists according to the Japanese "Group classification." Five popular convolutional neural networks, i.e., VGG16, VGG19, ResNet50, Xception, and InceptionV3 were trained and tested. The performance of the models was compared in terms of widely used metrics, namely, AUC (area under the receiver operating characteristic curve, i.e., ROC curve), accuracy, recall, precision, and F1 score. Results: Results showed that ResNet50 achieved the best performance with accuracy 93.16% and AUC 0.994. Conclusion: Our results demonstrated the applicability and effectiveness of DL-based system for automated Japanese "Group classification" of gastric endoscopic biopsy.