Robust Imaging Biomarkers for Brain Tumors

The goal of this PhD project is to develop robust imaging biomarkers for buildingprediction models of brain tumors. Brain tumor biomarkers are used for analysis of the disease, e.g. with respect to tumor grade, tumor recurrence and survival. We focus on magnetic resonance (MR) imaging, which is the most widely used imaging method for studying the brain and its disorders. Various scanning sequences exist for MR imaging that each give different contrasts and are used to visualize different biological properties of the tissues of interest. To study and diagnose brain tumors, several different MR sequences are typically used. However, an MR scan in one clinical center may not look the same as in another, even if the same subject and type of sequence is used. There are many factors that influence the resulting image, which makes development of computational diagnostic tools difficult. In this thesis we emphasize developing methods that overcome these issues to facilitate clinical adoption. We build generative models of tumor shape and apply them to both prediction and segmentation of brain tumors. We develop interpretable and robust imaging biomarkers based on whole-brain segmentations of MR images, and apply them to survival prediction of glioblastoma. These biomarkers measure the deformation of brain structures surrounding the tumor, and are computed fully automatically, while being sequence-adaptive. Using these biomarkers, we show improvement in survival prediction over models that only consider conventional non-imaging biomarkers.