Background: Endoscopic ultrasonography (EUS) is useful for the differential diagnosis of subepithelial lesions (SELs); however, not all of them are easy to distinguish. Gastrointestinal stromal tumors (GISTs) are the commonest SELs, are considered potentially malignant, and differentiating them from benign SELs is important. Artificial intelligence (AI) using deep learning has developed remarkably in the medical field. This study aimed to investigate the efficacy of an AI system for classifying SELs on EUS images.
Methods: EUS images of pathologically confirmed upper gastrointestinal SELs (GIST, leiomyoma, schwannoma, neuroendocrine tumor [NET], and ectopic pancreas) were collected from 12 hospitals. These images were divided into development and test datasets in the ratio of 4:1 using random sampling; the development dataset was divided into training and validation datasets. The same test dataset was diagnosed by two experts and two non-experts.
Results: A total of 16,110 images were collected from 631 cases for the development and test datasets. The accuracy of the AI system for the five-category classification (GIST, leiomyoma, schwannoma, NET, and ectopic pancreas) was 86.1%, which was significantly higher than that of all endoscopists. The sensitivity, specificity, and accuracy of the AI system for differentiating GISTs from non-GISTs were 98.8%, 67.6%, and 89.3%, respectively. Its sensitivity and accuracy were significantly higher than those of all the endoscopists.
Conclusion: The AI system, classifying SELs, showed higher diagnostic performance than that of the experts and may assist in improving the diagnosis of SELs in clinical practice.
Keywords: Artificial intelligence; Deep learning; Endoscopic ultrasonography; Gastrointestinal tumors; Subepithelial lesion.
© 2021. The Author(s) under exclusive licence to The International Gastric Cancer Association and The Japanese Gastric Cancer Association.