Abstract
Accurate prediction of lymph node metastasis (LNM) in T1 esophageal squamous cell cancer is critical for guiding treatment decisions after endoscopic submucosal dissection (ESD). We developed a deep learning-based artificial intelligence model using whole slide images (WSIs) to predict LNM and reduce overtreatment. The model was trained, validated, and internally tested on 160 surgically resected cases (72 LNM+, 88 LNM-) from 374 patients without prior ESD, achieving an AUC of 0.949 (95% CI: 0.912-0.986) on internal test. Further validation was performed on an external ESD cohort comprising clinically high-risk cases with invasion depths from MM to SM2. The model attained an accuracy of 90.1%, sensitivity of 81.8%, specificity of 91.4%, and an F1-score of 69.2%. It correctly classified 90.1% of samples, with a negative predictive value (NPV) of 96.9%. The high NPV and specificity underscore the model's utility in minimizing overtreatment while preserving diagnostic accuracy in high-risk T1 esophageal cancer.