AI-powered analysis of pancreatic ductal adenocarcinoma tissues to study the tumor immune ecosystem and identify novel classifiers

Background : Digital pathology and artificial intelligence (AI) are emerging as powerful tools in immuno-oncology, enabling enhanced diagnostic and prognostic workflows. The current trend regards the application of AI on histopathological images to extract relevant features beyond human visual perception. Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers worldwide, for which lack of efficient biomarkers to stratify patients and grade the response to therapy may play a significant impact. Aim : This project presents a computational pipeline integrating machine and deep learning approaches to characterize the PDAC tumor immune ecosystem and extract features with potential clinical relevance. Methods : Whole-slide images from 53 PDAC patients, including those treated and untreated with neoadjuvant chemotherapy (NAT), were analyzed. Slides were stained with H&E, picrosirius red to detect fibrosis, and a CD68 antibody to detect macrophages. On H&E slides, a deep learning tissue classifier pretrained on colorectal cancer quantified Tumor and Stroma areas and their Spatial Entropy. QuPath-based pixel classifiers segmented CD68+ regions to compute macrophage abundance and spatial aggregation measured by the Morisita Index, and quantified picrosirius red+ areas to assess fibrosis and collagen maturity. Tissue and immune metrics were correlated, and slide-level embeddings extracted with the foundation model CTransPath were projected onto UMAP to explore clustering patterns. Results : The deep learning tissue classifier achieved a global F1-score of 0.79 on PDAC slides. High macrophage abundance and Stroma Entropy were associated with worse overall survival. Integrating immune with tumor or stromal features improved patient stratification. NAT-treated patients showed increased fibrosis and reduced macrophage infiltration, with Stroma Entropy being predictive only in this group. Moreover, CTransPath-derived embeddings of CD68 and picrosirius red-stained slides clustered according to NAT regimen. Conclusions : This AI-driven pipeline enables quantitative spatial profiling of the PDAC immune microenvironment, revealing interpretable features with prognostic and predictive value. Our findings highlight the potential of computational pathology to derive clinically meaningful biomarkers and support therapy response evaluation.