Integration of clinical, pathological, and endoscopic data improves prediction for lymph node metastasis in early gastric cancer: a retrospective multicenter study

Background: Accurate prediction of lymph node metastasis (LNM) is crucial for treatment decisions in early gastric cancer (EGC). Current preoperative methods for LNM prediction are often insufficient. Patients and methods: This multicenter retrospective study enrolled 605 EGC patients who underwent endoscopic submucosal dissection (ESD) and/or surgery across five institutions between 2013 and 2023. We developed a deep learning model, LNMate, to predict LNM before ESD and evaluated its impact on endoscopists’ diagnostic performance. We also conducted immunohistochemical (IHC) analysis on tissue samples from 32 patients. Additionally, we constructed a deep learning-based endoscopic nomogram (DLEN) to estimate LNM risk after ESD, validated through comparison with the eCura system. Results: LNMate showed high predictive performance with area under the curves (AUCs) ranging from 0.843 (95% confidence interval CI, 0.782–0.904) to 0.875 (95% CI, 0.814–0.936). With LNMate assistance, endoscopists improved diagnostic accuracy by 13.8%, particularly in specificity (mean increase of 0.17, P = 0.03). IHC analysis showed associations between CD8+ and CD20+ cell enrichment with non-LNM predictions, and CD68+ macrophage infiltration with LNM. The DLEN outperformed the eCura system (AUC, 0.91 95% CI, 0.86–0.97 vs. 0.71 95% CI, 0.58–0.83, P < 0.01), reducing oversurgery by 24.2%, with no false negatives. Conclusions: The deep learning-based intelligent system for the entire treatment process of EGC showed excellent performance in predicting LNM, offering valuable decision support for both endoscopists in pre-ESD treatment planning and surgeons in post-ESD surgical decision-making.