Expert-level classification of gastritis by endoscopy using deep learning: a multicenter diagnostic trial

Abstract

Background and study aims Endoscopy plays a crucial role in diagnosis of gastritis. Endoscopists have low accuracy in diagnosing atrophic gastritis with white-light endoscopy (WLE). High-risk factors (such as atrophic gastritis [AG]) for carcinogenesis demand early detection. Deep learning (DL)-based gastritis classification with WLE rarely has been reported. We built a system for improving the accuracy of diagnosis of AG with WLE to assist with this common gastritis diagnosis and help lessen endoscopist fatigue. Methods We collected a total of 8141 endoscopic images of common gastritis, other gastritis, and non-gastritis in 4587 cases and built a DL -based system constructed with UNet + + and Resnet-50. A system was developed to sort common gastritis images layer by layer: The first layer included non-gastritis/common gastritis/other gastritis, the second layer contained AG/non-atrophic gastritis, and the third layer included atrophy/intestinal metaplasia and erosion/hemorrhage. The convolutional neural networks were tested with three separate test sets. Results Rates of accuracy for classifying non-atrophic gastritis/AG, atrophy/intestinal metaplasia, and erosion/hemorrhage were 88.78 %, 87.40 %, and 93.67 % in internal test set, 91.23 %, 85.81 %, and 92.70 % in the external test set ,and 95.00 %, 92.86 %, and 94.74 % in the video set, respectively. The hit ratio with the segmentation model was 99.29 %. The accuracy for detection of non-gastritis/common gastritis/other gastritis was 93.6 %. Conclusions The system had decent specificity and accuracy in classification of gastritis lesions. DL has great potential in WLE gastritis classification for assisting with achieving accurate diagnoses after endoscopic procedures.

Fig. 1

Structure of our system.

Fig. 2

Representative images of four kinds of gastritis lesions. a Atrophy. b IM. c Erosion. d Hemorrhage. The first column are originals. The second are segmentation masks. The third shows: green dotted line and white box domain surrounds CNN predicting domain and blue dotted line and red box surrounds manual descripting domain. The fourth are demonstration: classification results and confidence are displayed.

Fig. 3

Flowchart.

Fig. 4

Performance of endoscopists vs. model. a , b , c Accuracy of CNNs and endoscopists in the internal, external, and video test sets, respectively.

Fig. 5

Summary chart.