S1296 Deep Learning for Detection and Differentiation of Subepithelial Lesions: A Multicentric Transatlantic Endoscopic Ultrasound Study

Introduction: Subepithelial lesions (SELs) are commonly identified during esophagogastroduodenoscopy. Nevertheless, their diagnostic approach is complex, and patient management is dependent on appropriate anatomopathological diagnosis. In this context, endoscopic ultrasound offers the possibility of both lesion characterization and sampling, being a first line procedure in patients with suspected upper gastrointestinal tract SELs over 10-millimeter diameter. However, endoscopic ultrasound (EUS) is operator dependent and has a suboptimal diagnostic accuracy. This multicentric study aimed to develop a deep learning model for detection of upper gastrointestinal tract SEL, with distinction between leiomyoma and gastrointestinal stromal tumors (GIST). Methods: A convolutional neural network (yolo11net) was developed based on 57190 images from 65 patients from across 2 continents and 5 countries (Portugal, United States of America, Spain, Brazil, United Kingdom). The procedures were performed with 5 different EUS devices. A binary CNN identified and distinguished leiomyoma and GIST. The total data set was divided into a training, validation and testing dataset with a 70%/20%/10% ratio. The model was evaluated through its accuracy and area under receiver-operating characteristic curve (AUC-PR). Results: Pleomorphic SEL were identified with 91.8% accuracy. The model had an AUC-PR of 0.992 for GIST and 0.982 for leiomyoma. Conclusion: This was the first worldwide study capable of detecting and differentiating the commonest SEL in EUS. AI-driven has potential to revolutionize the approach to SEL, contributing for more accurate differential diagnosis while guiding EUS-guided tissue sampling. The inclusion of data from 5 different EUS devices enhances model interoperability, while integrating data from 5 countries mitigates the risk of demographic bias. Larger multicentric studies are needed for technology implementation.