Objective To develop and validate a model based on deep learning for automatic diagnosis of early gastric cancer （EGC） to improve detection and diagnosis of EGC. Methods A total of 5 159 images （including 1 000 images of EGC and 4 159 images of other benign lesions or normal patients） obtained from May 2014 to December 2016 were collected from endoscopic database in changhai Hospital. Then 4 449 images were selected randomly for a deep convolutional neural network （CNN） training， of which 768 were diagnosed as EGC and 3 681 diagnosed as other benign lesions or normal. The remaining 710 images were used to test the model by comparing with diagnostic results of four endoscopists. Results The deep learning model showed accuracy of 89.4％ （635／710）， sensitivity of 88.8％ （206／232） and specificity of 89.7％ （429／478）for EGC. The mean time required for diagnosis was 0.30±0.02 s. The performance of the model was superior to that of four endoscopists. Conclusion The model based on deep learning has high accuracy，sensitivity and specificity for detecting EGC，which could assist endoscopists in real-time diagnosis.