S97 Deep Learning for Lymph Node Detection and Malignancy Risk Stratification in Endoscopic Ultrasound: A Multicenter Proof of Concept Study

Introduction: Endoscopic ultrasound (EUS) is important for accurate lymph node (LN) characterization, being essential to guide treatment in suspected malignant involvement. Features like hypoechoic structures, sharply demarcated borders, rounded contours and size over 10 mm suggest malignancy, with 80-100% accuracy when combined, but this complete profile is uncommon. While EUS-guided fine needle aspiration/biopsy aids diagnosis, its accuracy can be limited. Artificial intelligence shows potential in improving diagnostic precision in EUS. However, data on convolutional neural network (CNN) application to EUS for LN malignancy prediction remains scarce. This multicenter study aims to evaluate a CNN’s effectiveness in predicting LN malignancy from EUS images. Methods: This multicenter study included EUS images from 9 centers and 4 countries. Lesions were classified as malignant or benign and delimited with bounding boxes. Definitive diagnoses were based on positive fine needle aspiration/biopsy or surgical specimen and, if negative, a minimum 6-month clinical follow-up was required. A CNN was developed including both detection and characterization modules. Performance metrics included detection rate, sensitivity, specificity, accuracy, and AUC were calculated. Results: Overall, 59,988 images from 108 patients were used to develop the model. The CNN distinguished malignant from benign LNs with a sensitivity of 93.7%, a specificity of 98.3% and an overall accuracy of 95.8%. The AUC was 0.95. Conclusion: To our knowledge, this study is 1 of the first to evaluate the performance of deep learning systems for LN assessment using EUS imaging, highlighting CNNs' potential to improve diagnostic accuracy and support clinical decisions. Preoperative LN status assessment is crucial for adjusting treatment plans based on metastatic involvement. This artificial intelligence-powered imaging model with diverse demographic data showed excellent detecting and classification capabilities, highlighting its potential to provide a valuable tool to refine LN assessment with EUS, ultimately supporting more tailored and efficient patient care.