Uncertainty estimation in digital pathology: Towards applying artificial intelligence in an uncertain clinical world

J. Linmans

• Promotor: G. Litjens and J. van der Laak
• Graduation year: 2025
• Radboud University

Abstract

The main focus of this work is to evaluate different methods of uncertainty estimation and develop novel methodologies to help improve the reliability of machine learning model deployment within the context of digital pathology. This research can be divided into three key blocks:

The investigation and evaluation of prevalent and effective uncertainty-aware models that can distinguish out-of-distribution from in-distribution data based on the uncertainty of predictions. Translating the pixel-level uncertainty into whole-slide level uncertainty and doing so efficiently enough to accompany the large-scale nature of digital pathology (Chapter 2);

The development and validation of a stochastic classification framework capable of modeling uncertainty within the context of inter-observer variability of cancer grading tasks (Chapter 3);

The investigation of the potential of unsupervised learning methods, particularly the recently introduced diffusion-based models, to be used as universal anomaly detectors when trained solely on benign tissue (Chapter 4).