Probabilistic and causal reasoning in deep learning for imaging

• Main Author: Pawlowski, Nick
• Format: Dissertation
• Language: English
• Published: Imperial College London 2022
• DOI: 10.25560/96806

Typical machine learning research in the imaging domain occurs in clearly defined environments on clean datasets without considering realistic deployment scenarios. However, applied machine learning systems are exposed to unexpected distribution shifts and still need to produce reliable predictions without relying on spurious correlations. Similarly, such systems encounter ambiguous or unseen inputs and need to communicate their uncertainty. Often, AI systems support a human operator and should provide interpretable explanations of their decisions. This thesis argues for a probabilistic and causal approach to machine learning that is robust to spurious correlations, improves interpretability, and communicates uncertainty. First, we investigate the learning abilities of neural networks that are constrained to extracting information from image patches. We show that careful network design can prevent shortcut learning and that restricting the receptive field can improve the interpretability of predictions. We tackle uncertainty estimation by introducing a Bayesian deep learning method to approximate the posterior distribution of the weights of a neural network using an implicit distribution. We verify that our method is capable of solving predictive tasks while providing reliable uncertainty estimates. Moving on, we frame various medical prediction tasks within the framework of outlier detection. We apply deep generative modelling to brain MR and CT images as well as histopathology images and show that it is possible to detect pathologies as outliers under a normative model of healthy samples. Next, we propose deep structural causal models as a framework capable of capturing causal relationships between imaging and non-imaging data. Our experiments provide evidence that this framework is capable of all rungs of the causal hierarchy. Finally, with further thoughts on applications of uncertainty estimation, robust causal estimation, and fairness we conclude that the safe and reliable deployment of AI systems to real-world scenarios requires the integration of probabilistic and causal reasoning.