image: Spatial distribution of LNM risk in the WSI tumor area for patients with positive lymph nodes (A) and negative lymph nodes (B); (C) Gradient-weighted class activation map based on the CrossFormer model.
Credit: ©Medicine Plus
Gastric cancer ranks as the fifth most common malignancy worldwide and the fifth leading cause of cancer-related mortality. In recent years, endoscopic surgery has expanded the indications for endoscopic treatment of early gastric cancer. Previous studies have demonstrated that the incidence of lymph node metastasis (LNM) in T1b gastric cancer ranges from 22.1% to 27.3%, with the majority of patients not developing LNM. Accurate identification of high-risk features for LNM in T1b gastric cancer would facilitate broadening the indications for endoscopic therapy.
This large-scale, multicenter, retrospective study developed a predictive model for LNM status based on 1,023 patients with T1b gastric cancer. The developed DL-T1 nomogram demonstrated robust predictive performance and consistency across participating centers. The DL-T1b score, developed and validated using hematoxylin-eosin (HE)-stained digital pathology, was employed to predict LNM status in T1b gastric cancer, achieving an area under the curve (AUC) of 0.910 (95% CI: 0.869, 0.945) in the testing cohort. Integrating six baseline risk factors with the DL-T1b score into a nomogram further improved predictive accuracy, offering a precise decision-making tool for personalized treatment of T1b gastric cancer.
Notably, current weakly supervised deep learning models directly predict LNM from whole-slide images (HE-WSIs) with generally low accuracy. In contrast, this study employed the CrossFormer model to calculate the LNM probability for each image patch within the tumor region, and defined the DL-T1b score as the proportion of image patches with an LNM probability of 1.00 relative to the total number of image patches across the entire tumor area, achieving an AUC of 0.910 for predicting LNM in T1b gastric cancer. Gradient-weighted class activation mapping (Grad-CAM) further revealed that tumor cells and their microarchitectural patterns potentially contribute to LNM prediction (Figure 1).
This study provides the first evidence that predicting LNM status in patients with T1b gastric cancer using deep learning and pathomics features is feasible and highly accurate, facilitating the identification of potential beneficiaries of endoscopic surgery for early gastric cancer. Furthermore, it establishes a theoretical foundation for constructing LNM prediction models based on biopsy pathology in larger-scale early gastric cancer cohorts.