Objective To improve the detection rate of early esophageal cancer during endoscopy by construction of artificial intelligence assistant diagnosis system. Methods A total of 2 400 esophageal images were collected from Zhongshan Hospital of Fudan University from January 2016 to December 2017, including 1 200 images of early esophageal cancer and 1 200 images of normal esophageal mucosa. The lesions in pictures were marked with rectangular box by using computer program. Among them, 2 000 pictures were divided into the training set and 400 pictures into the test set. An assistant diagnostic model of early esophageal cancer was established by back propagation algorithm in computer deep learning. The training model was tested and the sensitivity and specificity of the system at different cut-off points in the test set was calculated. Receiver operating characteristic (ROC) curve was used to evaluate the performance of the diagnostic model. Results The area under ROC curve (AUC) of the auxiliary diagnostic model was 0996 1. The sensitivity and specificity were satisfactory. Conclusion The deep learning model constructed in this study has good specificity, sensitivity and AUC value in the diagnosis of early esophageal cancer, and can assist endoscopists in real-time diagnosis in clinical examination.