BACKGROUND AND AIMS: Diagnosing the invasion depth of gastric cancer (GC) is necessary to determine the optimal method of treatment. Although the efficacy of evaluating macroscopic features and EUS has been reported, there is a need for more accurate and objective methods. The primary aim of this study was to test the efficacy of novel artificial intelligence (AI) systems in predicting the invasion depth of GC. METHODS: A total of 16,557 images from 1084 cases of GC for which endoscopic resection or surgery was performed between January 2013 and June 2019 were extracted. Cases were randomly assigned to training and test datasets at a ratio of 4:1. Through transfer learning leveraging a convolutional neural network architecture, ResNet50, 3 independent AI systems were developed. Each system was trained to predict the invasion depth of GC using conventional white-light imaging (WLI), nonmagnifying narrow-band imaging (NBI), and indigo-carmine dye contrast imaging (Indigo). RESULTS: The area under the curve of the WLI AI system was .9590. The lesion-based sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of the WLI AI system were 84.4%, 99.4%, 94.5%, 98.5%, and 92.9%, respectively. The lesion-based accuracies of the WLI, NBI, and Indigo AI systems were 94.5%, 94.3%, and 95.5%, respectively, with no significant difference. CONCLUSIONS: These new AI systems trained with multiple images from different angles and distances could predict the invasion depth of GC with high accuracy. The lesion-based accuracy of the WLI, NBI, and Indigo AI systems was not significantly different.
BACKGROUND AND AIMS: Diagnosing the invasion depth of gastric cancer (GC) is necessary to determine the optimal method of treatment. Although the efficacy of evaluating macroscopic features and EUS has been reported, there is a need for more accurate and objective methods. The primary aim of this study was to test the efficacy of novel artificial intelligence (AI) systems in predicting the invasion depth of GC. METHODS: A total of 16,557 images from 1084 cases of GC for which endoscopic resection or surgery was performed between January 2013 and June 2019 were extracted. Cases were randomly assigned to training and test datasets at a ratio of 4:1. Through transfer learning leveraging a convolutional neural network architecture, ResNet50, 3 independent AI systems were developed. Each system was trained to predict the invasion depth of GC using conventional white-light imaging (WLI), nonmagnifying narrow-band imaging (NBI), and indigo-carmine dye contrast imaging (Indigo). RESULTS: The area under the curve of the WLI AI system was .9590. The lesion-based sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of the WLI AI system were 84.4%, 99.4%, 94.5%, 98.5%, and 92.9%, respectively. The lesion-based accuracies of the WLI, NBI, and Indigo AI systems were 94.5%, 94.3%, and 95.5%, respectively, with no significant difference. CONCLUSIONS: These new AI systems trained with multiple images from different angles and distances could predict the invasion depth of GC with high accuracy. The lesion-based accuracy of the WLI, NBI, and Indigo AI systems was not significantly different.