LMRNet: a biopsy-derived deep learning model for predicting lymph node metastasis in early gastric cancer with pathological interpretability

Early gastric cancer (EGC) typically carries a favorable prognosis, and many patients can achieve curative outcomes through local resection when lymph node metastasis (LNM) is absent. Therefore, reliable preoperative evaluation of perigastric LNM is essential for guiding individualized treatment decisions. This study aimed to establish an accurate deep learning model for predicting LNM in EGC using biopsy slides and to investigate the pathological and microenvironmental features that underlie the model’s predictive behavior. T1-stage gastric cancer patients from multiple centers with pathologically confirmed LNM status after D2 lymphadenectomy were included. H&E-stained biopsy slides were used to train deep learning models, and three high-performing models (ViTamin, ConvNeXt V2, Swin Transformer V2) were integrated into an ensemble model, LMRNet. Its performance was evaluated using multicenter cohorts. Additionally, a dual Swin Transformer classifier and HoVer-Net were employed to characterize histopathological subtypes and tumor microenvironment (TME) features, allowing interpretation of the histopathological and microenvironmental patterns associated with model predictions. A total of 660 patients were enrolled. LMRNet achieved AUCs of 0.947 in the internal biopsy cohort and 0.900 in the external biopsy cohort, demonstrating high accuracy. In multicenter surgical cohorts, performance remained stable with AUCs ranging from 0.768 to 0.894. The subtype classifier showed strong agreement with TCGA annotations and revealed that diffuse-type regions were associated with higher LNM risk. TME analysis demonstrated that patches enriched with inflammatory cells, particularly those in close spatial proximity to tumor cells, received significantly higher LNM risk. Regions where stromal cells formed structural barriers between tumor and inflammatory cells were assigned lower risk. These findings highlight specific morphological and cellular interaction patterns captured by LMRNet. LMRNet provides accurate and generalizable LNM prediction based on biopsy slides. By integrating subtype and TME analyses, the study uncovers morphological correlates of predicted metastatic risk in EGC.