Fully Automated Medical Image Analysis Facilitating Subsequent User Analysis

In a clinical setting, accuracy is paramount for medical image analysis tasks such as segmentation and registration. Since it is often required that results be manually verified by a human expert, computational techniques designed to aid clinicians in these image analysis tasks are usually interactive, requiring user input. However, these techniques cannot take advantage of the time between when an image is acquired and when a clinician is available to provide input. In this thesis, we will present novel techniques for automatically processing medical images, with the goal of facilitating later analysis by a human expert. These techniques fall into two classes. The first class involves leveraging prior anatomical information to automatically generate results that are robust and independent of initialization. The second class involves precomputing data that is used to greatly increase the speed and responsiveness of subsequent interactive techniques, saving clinicians valuable time. Each of the techniques presented focus on encoding meaningful uncertainty information, which can guide human experts to potential errors or pathologies.