Prior Knowledge for Targeted Object Segmentation in Medical Images

Medical image segmentation, the task of partitioning an image into meaningful parts, is an important step toward automating medical image analysis and is at the crux of a variety of medical imaging applications, such as computer aided diagnosis, therapy planning and delivery, and computer aided interventions. However, existence of noise, low contrast and objects' complexity in medical images preclude ideal segmentation. Incorporating prior knowledge into image segmentation algorithms has proven useful for obtaining more accurate and plausible results on targeted objects segmentation. In this thesis, we develop novel techniques to augment optimization-based segmentation frameworks with different types of prior knowledge to identify and delineate only those objects (targeted objects) that conform to specific geometrical, topological and appearance priors. These techniques include employing prior knowledge to segment multi-part objects with part-configuration constraints and encoding priors based on images acquired from different imaging equipment and of differing dimensions. Our objective is to satisfy two important aspects in optimization-based image segmentation: (1) fidelity-optimizability trade-off, and (2) space and time complexity.Particularly, in our first contribution, we adopt several prior information to build a faithful objective function unconcerned about its convexity to segment potentially overlapping cells with complex topology. In our second contribution, we improve the space and time complexity and augment the level sets framework with the ability to handle geometric constraints between boundaries of multi-region objects. In our first two contributions we opt for ensuring the objective function is flexible enough (even if it is non-convex) to accurately capture the intricacies of the segmentation problem. In our third contribution, we focus on optimizability. We propose a convex formulation to augment the popular Mumford-Shah model and develop a new regularization term to incorporate similar geometrical and distance prior as our second contribution while maintaining global optimality. Lastly, we efficiently incorporate different types of priors based on images acquired from different imaging equipment (different modalities) and of dissimilar dimensions to segment multiple objects in intraoperative multi-view endoscopic videos. We show how our technique allows for the inclusion of laparoscopic camera motion model to stabilize the segmentation.