ABSTRACT Histopathological hematoxylin and eosin (H&E) slides contain valuable prognostic information for pancreatic ductal adenocarcinoma (PDAC), yet systematic feature extraction remains challenging. This multi‐center study developed and validated an automated prognostic model using deep learning on digitized whole‐slide images from 873 PDAC patients with surgical resection across three academic centers. The CrossFormer architecture achieved superior performance in external validation (area under the curve AUC = 0.774), significantly outperforming ResNet‐18 (AUC = 0.716), ResNet‐50 (AUC = 0.737), and DenseNet‐121 (AUC = 0.729). Gradient‐weighted Class Activation Mapping identified key prognostic features including desmoplastic stroma, high nuclear‐to‐cytoplasmic ratio, tumor necrosis, and immune cell infiltration. The pathomics signature effectively stratified patients into low‐risk and high‐risk groups with significant survival differences ( p < 0.001). Critically, carbohydrate antigen 19‐9 (CA19‐9) retained prognostic value only in low‐risk patients (hazard ratio HR = 2.70, p < 0.001) but not in high‐risk patients (HR = 0.998, p = 0.990). High‐risk patients derived substantial benefit from adjuvant chemotherapy (HR = 0.56, p = 0.038), whereas low‐risk patients showed no significant benefit (HR = 0.83, p = 0.562). These findings provide actionable clinical insights: treatment intensification for high‐risk patients and CA19‐9‐guided monitoring for low‐risk patients. This validated, interpretable model transforms routine H&E slides into quantitative prognostic tools, enabling personalized treatment strategies without additional testing costs.
Chen et al. (Tue,) studied this question.