AI-assisted spatial profiling of the pancreatic ductal adenocarcinoma ecosystem identifies prognostic tissue-immune features

Digital pathology and AI are increasingly applied in immuno-oncology, enabling quantitative and spatial analysis of histopathological images beyond human visual assessment. Pancreatic ductal adenocarcinoma (PDAC) is among the deadliest cancer worldwide, characterized by a heterogeneous tumor microenvironment and limited biomarkers for patient stratification and therapy response evaluation. This study aimed to develop an AI-assisted spatial profiling framework integrating machine and deep learning approaches to characterize the PDAC tumor-immune ecosystem and identify features with potential clinical relevance. Whole-slide images from 53 PDAC patients, including neoadjuvant chemotherapy-treated and untreated cases, were analyzed using H&E, picrosirius red, and CD68 immunohistochemistry. A deep learning tissue classifier pretrained on colorectal cancer quantified tumor and stromal areas and computed Spatial Entropy. QuPath-based pixel classifiers segmented CD68-positive regions and picrosirius red-positive areas, quantifying macrophage and fibrosis abundance, and their spatial aggregation using the Morisita Index and Entropy. CTransPath-based slide embeddings were projected onto UMAPs to explore clustering patterns. The tissue classifier achieved a global F1-score of 0.79 on PDAC slides. Stromal and fibrotic spatial arrangement emerged as independent prognosticators, with combined Fibrosis Entropy and macrophage abundance providing the strongest survival stratification. Neoadjuvant chemotherapy induced tissue remodelling, characterized by increased fibrosis abundance and reduced Fibrosis Entropy and macrophage infiltration. Stroma Entropy was prognostic specifically in treated patients, suggesting its potential utility as a biomarker of therapy-induced tissue response. Slide embeddings clustered according to treatment regimen. This AI-driven framework enables quantitative profiling of the PDAC ecosystem, revealing interpretable features with prognostic value and supporting therapy response grading.