Towards safe implementation of AI-based malignancy probability estimation of pulmonary nodules on thoracic CT

D. Peeters

• Promotor: C. Jacobs, W. Prokop and R. Vliegenthart
• Graduation year: 2026
• Radboud University

Abstract

The central aim of this thesis was to advance the reliability and safety of AI for pulmonary nodule malignancy probability estimation in both lung cancer screening and routine clinical care. Rather than focusing solely on improving predictive performance, this work addressed some key limitations that can slow down the clinical deployment of AI systems. In Chapter 2, we demonstrated that estimating AI uncertainty can reliably identify cases in which the performance of AI starts to degrade, enabling more cautious and informed use of model predictions. In Chapter 3, we showed that out-of-distribution detection can identify previously unseen data scenarios (dataset shifts) that might negatively affect AI reliability when deployed in clinical practice. In Chapter 4, we established a standardized benchmarking framework through the LUNA25 challenge and demonstrated that an AI system can outperform the average radiologist in malignancy probability estimation of indeterminate nodules on lung cancer screening CT. Finally, in Chapter 5, we showed that using semi-automated nodule detection strategies for the development of AI can substantially reduce radiologist workload with compromising AI model performance. Together, these chapters form a complementary framework that supports responsible, safe and reliable development and deployment of AI for lung nodule malignancy probability estimation.