Diagnosis of routine endoscopic images of gastric lesions through a deep convolutional neural network / 中华消化内镜杂志

ABSTRACT

Objective:To develop a deep convolutional neural network (CNN) to automatically detect gastric lesions in endoscopic images.Methods:A CNN-based diagnostic system was constructed based on ResNet-34 residual network structure and DeepLabv3 structure, and trained by using 17 217 routine gastroscopy images.These images were from 1 121 gastric lesions of five types acquired in Peking University People′s Hospital between 2012 and 2018, namely peptic ulcer (PU), early gastric cancer (EGC) and high-grade intraepithelial neoplasia (HGIN), advanced gastric cancer (AGC), gastric submucosal tumors (SMTs), and normal gastric mucosa without lesions. The trained CNN was evaluated through a test dataset that contained 1 091 routine gastroscopy images of 237 gastric lesions. The accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the CNN were calculated.Results:The accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the CNN-assisted diagnosis of EGC and HGIN were 78.6% (33/42), 84.4% (27/32), 60.0% (6/10), 87.1% (27/31), and 54.5% (6/11), respectively. The accuracy, sensitivity, and specificity of CNN-assisted diagnosis of PU were 90.4% (47/52), 92.7% (38/41), and 81.8% (9/11), respectively, the outcomes of AGC were 88.1% (52/59), 91.8% (45/49), and 70.0% (7/10), respectively, and those of gastric SMTs were 86.0% (43/50), 89.7% (35/39), and 72.7% (8/11), respectively. The CNN′s recognition time for all images of the test set was 42 seconds.Conclusion:The constructed CNN system, as a rapid and accurate auxiliary diagnostic instrument, can detect not only EGC and HGIN but also other gastric lesions.