Advancing probabilistic and causal deep learning in medical image analysis

[thesis] [cite]

Abstract

The power and flexibility of deep learning have made it an indispensable tool for tackling modern machine learning problems. However, this flexibility comes at the cost of robustness and interpretability, which can lead to undesirable or even harmful outcomes. Deep learning models often fail to generalise to real-world conditions and produce unforeseen errors that hinder wide adoption in safety-critical critical domains such as healthcare. This thesis presents multiple works that address the reliability problems of deep learning in safety-critical domains by being aware of its vulnerabilities and incorporating more domain knowledge when designing and evaluating our algorithms. We start by showing how close collaboration with domain experts is necessary to achieve good results in a real-world clinical task - the multiclass semantic segmentation of traumatic brain injuries (TBI) lesions in head CT. We continue by proposing an algorithm that models spatially coherent aleatoric uncertainty in segmentation tasks by considering the dependencies between pixels. The lack of proper uncertainty quantification is a robustness issue which is ubiquitous in deep learning. Tackling this issue is of the utmost importance if we want to deploy these systems in the real world. Lastly, we present a general framework for evaluating image counterfactual inference models in the absence of ground-truth counterfactuals. Counterfactuals are extremely useful to reason about models and data and to probe models for explanations or mistakes. As a result, their evaluation is critical for improving the interpretability of deep learning models.