Automated pancreatic cancer pathology image segmentation using deep learning to quantify lymphocyte stroma ratio

Pancreatic ductal adenocarcinoma (PDAC) exhibits profound tumor microenvironment heterogeneity, and conventional prognostic tools often fail to adequately quantify critical features like the lymphocyte-stroma ratio (LSR). Manual assessment of LSR is prone to variability and inefficiency, which limits its clinical utility. We developed a Vision Transformer-based model using whole-slide images (WSIs) from multiple centers for training and validation. Our model achieved high accuracy in tissue classification and demonstrated that higher LSR is significantly correlated with improved overall survival. Kaplan–Meier survival curves indicated a significantly higher risk for the LSR-low group in both development ( P = 0.002) and multicenter cohorts ( P < 0.001). We also built an online platform for automated WSI segmentation and quantification, providing both visualization and quantification of tissue categories. This study establishes a robust, automated LSR quantification system validated across multicenter PDAC cohorts, offering a scalable solution for precision oncology by linking computational biomarkers to therapeutic outcomes.