Prediction of lymph node metastasis in early gastric cancer using artificial intelligence technology.

Abstract

289 Background: Early gastric cancer shows lymph node involvement in about 10-15% of patients. Despite the fact, we perform radical lymphadenectomy for all patients because predicting lymph node metastasis has yet to be successful. In this study, we hypothesize that image analysis using artificial intelligence (AI) technology may help solve the problem. Methods: We retrospectively collected 82 patients with clinical T1N0 and pathological node negative and 82 patients with clinical T1N0 and pathological node positives and then divided the 164 patients into a training:validation set in ratio of 9:1. Endoscopic images of the early tumors were analyzed by transfer learning using AlexNet, a deep neural network containing 5 convolutional layers and 3 fully-connected layers. The model was validated with newly-collected 40 images from 20 clinical T1N0 and pathological node negative and 20 patients with clinical T1N0 and pathological node positives as a test set. For comparison, three methods of prediction were implemented: prediction at random, by logistic regression, and by skilled endoscopists. Results: The AI predicted LNM with accuracy of 80.9% in the validation set and 66.9% in the test set. (48.3% for node negative cancers and 85.4% for node positive cancers) On the other hand, prediction at random, by logistic regression, and by 2 endoscopists resulted in 50.3%, 50.0%, and 47.5%, respectively. Conclusions: Although the accuracy still needs to be improved, image analysis using the AI technology resulted in the best prediction of lymph node metastasis, indicating that AI is a promising technology for the diagnosis of lymph node metastasis in early gastric cancer.