Computational Modeling of Medical Images of Brain Tumor Patients for Optimized Radiation Therapy Planning

In brain tumor radiation therapy, the aim is to maximize the delivered radiation dose to the targeted tumor and at the same time minimize the dose to sensitive healthy structures – so-called organs-at-risk (OARs). When planning a radiation therapy session, the tumor and the OARs therefore need to be delineated on medical images of the patient’s head, to be able to optimize a radiation dose plan. In clinical practice, the delineation is performed manually with limited assistance from automatic procedures, which is both time-consuming and typically suffers from poor reproducibility. There is, therefore, a need for automated methods that can segment both brain tumors and OARs. However, there is a noticeable lack in the literature of methods that simultaneously segment both types of structures.

To automatically segment medical images of brain tumor patients is difficult because brain tumors vary greatly in size, shape, appearance and location within the brain. Furthermore, healthy structures surrounding a tumor are pushed and deformed by the so-called mass effect of the tumor. Moreover, medical imaging techniques often result in imaging artifacts and varying intensity across imaging centers.

The goal of this PhD-project was to develop automated segmentation methods that can handle both brain tumors and OARs. In the first part of the project, we developed a model for tumor shape and used it to develop a fully automated generative method specifically for brain tumor segmentation. This method performed favorably compared to other state-of-the-art methods. In the second part of the project, we used a probabilistic atlas-based model capable of detailed modeling of the spatial organization in a healthy brain, and extended it to handle various OARs. We incorporated this model into the previously used modeling framework. In experiments, we showed that the resulting model was capable of simultaneous segmentation of brain tumors and OARs, while also being capable of adapting to varying image sequences and images from different imaging centers.