Gastroenterologist-level detection of gastric precursor lesions and neoplasia with a deep convolutional neural network

Abstract

Background: Gastric precursor lesions and neoplasia with very delicate changes in the gastric mucosa could be easily missed or misdiagnosed under endoscopy. Here we developed an automatic real-time pattern recognition tool based on convolutional neural networks (CNNs) algorithm to help endoscopists in detection of chronic atrophic gastritis (CAG) and gastric cancer (GC) lesions. Methods: The five-convolution-layer ZF model and thirteen-convolution-layer VGG16 model were combined in our neural network A total of 10,014 CAG and 3724 GC annotated images were used in the network training. Another independent set consisted of 50 CAG, 50 GC and 100 negative controls images were used to evaluate the performance of the final network. Results: In CAG detection, the performance of our model was much better than the average performance of the 77 endoscopists in sensitivity, specificity and accuracy (95% versus 74%, 86% versus 82%, 90% versus 78%, respectively). In GC detection, the performance of our model achieved a slightly higher sensitivity (90% versus 87%), but a lower specificity (50% versus 74%) and accuracy (70% versus 80%) than the average performance of the 89 endoscopists. Conclusion: In conclusion, we provided a CNN based computational tool to improve the detection of CAG and GC under endoscopy, and simplify diagnostic procedures.