Abstract
<title>Abstract</title> <p> Background Tumor–stroma ratio (TSR) and histological growth patterns are recognized prognostic factors in pancreatic ductal adenocarcinoma (PDAC); however, conventional assessment remains subjective and limited in scalability. This study aimed to develop a high-precision digital pathology pipeline and to investigate the prognostic significance of TSR and ductal morphology in PDAC. Methods A ResNet-50–based deep learning model was applied to 740 hematoxylin–eosin (H&E)–stained whole-slide images from 320 PDAC patients to automatically identify and quantify nine histological components, including tumor epithelium and stroma. Quantitative metrics—TSR, tumor burden, and stromal content—were correlated with clinicopathological variables and survival outcomes. Prognostic modeling was conducted using five machine learning algorithms (Random Forest, XGBoost, k-nearest neighbors, CatBoost, and logistic regression), with hyperparameter optimization performed using Optuna. Major ductal growth patterns, namely large-duct predominant (LDP) and small-duct predominant (SDP) morphologies, were evaluated as complementary histological features. Results The ResNet-50 model achieved high classification performance (accuracy of 0.94 in external validation and 0.97 in the TCGA cohort) and enabled reliable quantification of tumor and stromal components. A high TSR was significantly associated with shorter progression-free survival (PFS; <italic>p</italic> = 0.0002) and overall survival (OS; <italic>p</italic> = 0.026), as well as higher histological grade, presence of vascular tumor thrombi, and increased Ki-67 index. Further analyses demonstrated that the adverse prognostic impact of TSR was primarily driven by increased tumor burden, whereas stromal content alone showed no significant association with survival. Tumors exhibiting LDP morphology were characterized by lower TSR, less aggressive histopathological features, and significantly improved OS and PFS compared with SDP tumors. Among the machine learning models, the optimized Random Forest model identified tumor burden as the strongest predictor of poor clinical outcomes. Conclusions The ResNet-50–based digital pathology framework enables precise and automated quantification of PDAC histological components and TSR. A high TSR—predominantly reflecting increased tumor burden—is strongly associated with unfavorable prognosis, while LDP morphology delineates a subset of PDAC with less aggressive biological behavior. </p>