Learning from histopathology images: AI-driven biomarkers for pancreatic ductal adenocarcinoma

P. Venditelli

• Promotor: G. Litjens and J. van der Laak
• Copromotor: J. Hermans
• Graduation year: 2026
• Radboud University

Abstract

The overarching aim of this thesis is to develop and validate artificial intelligence (AI)-based methodologies to improve prognostic assessment in pancreatic cancer, with a specific focus on histopathological analysis. In addition to enhancing predictive accuracy, a key objective is to gain a deeper understanding of the tumor's morpho-functional behavior—namely, how its structural and cellular characteristics relate to clinical outcomes. To achieve this, the work begins with the development of an automated model for tumor segmentation in hematoxylin and eosin (H&E)-stained whole-slide images. Building upon this, the thesis explores the automatic quantification of the tumor-stroma ratio (TSR), a promising morphological biomarker, and investigates its association with patient survival. Subsequently, the study expands to include a broader set of morphometric features, evaluating their combined prognostic value about overall survival. In parallel, different feature extraction strategies—ranging from handcrafted descriptors to deep learning embeddings—are systematically compared to assess their utility in survival modeling.