PhD defence G. (Gijs) van Tulder

Dissertation in short:

Machine learning is used to automate many tasks in medical image analysis, such as segmentation and classification of MRI and CT images. However, machine learning models do not always generalize well to new domains, because their predictions rely on subtle image characteristics that may vary between sources. For example, a model that is optimized for images from one type of scanner might work less well for images from a different type. Similarly, there may be differences between scans collected in different hospitals, or with different scanning protocols, that make it difficult to apply existing models to new datasets. Domain adaptation methods address this generalization problem by reducing the differences between images from different sources, or by making the models less sensitive to them.

In this thesis, we use domain adaptation methods based on deep learning and representation learning. These methods map images from different sources to a shared feature space, where they can be analyzed by a common classification or segmentation model that works similarly for all domains. We evaluate supervised and unsupervised learning for several medical imaging tasks, such as lung tissue classification and brain tumor segmentation, using several domain adaptation approaches, such as optimizing representation similarity, cross-domain reconstruction, and domain adversarial learning. We explore how these methods learn common representations, which assumptions they make about the data, and when they do and do not work. In the final chapter, we use these findings to discuss how domain adaptation can be used in future medical imaging applications.