Artificial intelligence for pancreatic cancer detection

Download thesis

N. Alves

  • Promotor: H. Huisman
  • Copromotor: J. Hermans
  • Graduation year: 2025
  • Radboud University

Abstract

This thesis explores the application of AI for the detection of pancreatic cancer in contrast-enhanced CT imaging, with a particular emphasis on the translational path toward real-world clinical implementation. The central hypothesis across all chapters is that AI can augment radiologists by enhancing diagnostic accuracy and efficiency, while upholding rigorous standards of patient safety. The work presented here aims to address key gaps in the current research landscape:

Fragmented literature and lack of a unified research agenda for image-based AI in pancreatic cancer: Chapter 3 provides a comprehensive scoping review of existing AI applications in this domain, mapping the landscape across the patient pathway. This identifies critical challenges and opportunities, and proposes a structured research agenda to guide future efforts. Chapter 4 builds on this by demonstrating how this agenda can be operationalized to support clinically impactful AI integration across different stages of the care pathway.

Insufficient anatomical context and localization in AI systems for cancer detection: Chapter 2 introduces the first end-to-end AI system designed to detect and localize pancreatic cancer, with a focus on integrating surrounding anatomical structures to improve the detection small, early-stage tumors.

Lack of benchmarking between AI systems and radiologists: Chapter 4 presents the first large-scale observer study benchmarking AI performance in an international Grand Challenge, and directly comparing AI with radiologists for pancreatic cancer detection, offering rigorous insights into potential clinical benefits and limitations.

Absence of robust safety mechanisms in clinical cancer detection AI: Chapter 6 proposes an uncertainty quantification metric tailored for safe deployment of AI across diverse diseases, imaging modalities, and datasets. Chapter 7 further explores the utility of uncertainty estimates in guiding AI training, particularly in leveraging large-scale unlabelled data to improve model robustness and reliability.