Towards robust and reliable disease classification in medical imaging

In healthcare, machine learning systems have the potential to improve clinical workflows, leading to better patient outcomes, shortened waiting times, and reduced health inequalities. Such systems have already achieved human-level performance in numerous challenging image-based disease detection tasks. Nonetheless, no model is error-free. Hence, for the safe and reliable use of machine learning in clinical practice, there is a critical need to establish safeguards for error detection and to ensure models work robustly across different environments. This is the purpose of this thesis. In the first part, we focus on improving the reliability of image-based disease classification models. Specifically, we study the detection of misclassified samples, and introduce novel methods for automated performance estimation and dataset shift identification. This research is key for developing systems that alert clinicians when machine learning models fail. In the second part, we address model robustness, focusing on developing novel methods to ensure predictions remain accurate regardless of changes in image acquisition protocols. On the one hand, we propose an unsupervised recalibration method to automatically correct clinical metric drifts induced by acquisition shifts at test-time, for already trained models. On the other hand, we propose counterfactual contrastive learning, a novel framework leveraging causal generative modelling to enhance the robustness of contrastively-learned image representations. Overall, by developing new methods to (i) improve reliability and performance monitoring of disease classification models and (ii) enhance model robustness against image acquisition shifts, this thesis takes important steps towards enabling the safe deployment of machine learning systems in medical imaging.