Image registration methods, which aim at aligning two or more images into one coordinate system, are among the oldest and most widely used algorithms in computer vision. Registration methods serve to establish correspondence relationships among images (captured at different times, from different sensors or from different viewpoints) which are not obvious for the human eye. A particular type of registration algorithm, known as graph-based deformable registration methods, has become popular during the last decade given its robustness, scalability, efficiency and theoretical simplicity. The range of problems to which it can be adapted is particularly broad. In this thesis, we propose several extensions to the graph-based deformable registration theory, by exploring new application scenarios and developing novel methodological contributions.Our first contribution is an extension of the graph-based deformable registration framework, dealing with the challenging slice-to-volume registration problem. Slice-to-volume registration aims at registering a 2D image within a 3D volume, i.e. we seek a mapping function which optimally maps a tomographic slice to the 3D coordinate space of a given volume. We introduce a scalable, modular and flexible formulation accommodating low-rank and high order terms, which simultaneously selects the plane and estimates the in-plane deformation through a single shot optimization approach. The proposed framework is instantiated into different variants based on different graph topology, label space definition and energy construction. Simulated and real-data in the context of ultrasound and magnetic resonance registration (where both framework instantiations as well as different optimization strategies are considered) demonstrate the potentials of our method.The other two contributions included in this thesis are related to how semantic information can be encompassed within the registration process (independently of the dimensionality of the images). Currently, most of the methods rely on a single metric function explaining the similarity between the source and target images. We argue that incorporating semantic information to guide the registration process will further improve the accuracy of the results, particularly in the presence of semantic labels making the registration a domain specific problem.We consider a first scenario where we are given a classifier inferring probability maps for different anatomical structures in the input images. Our method seeks to simultaneously register and segment a set of input images, incorporating this information within the energy formulation. The main idea is to use these estimated maps of semantic labels (provided by an arbitrary classifier) as a surrogate for unlabeled data, and combine them with population deformable registration to improve both alignment and segmentation.Our last contribution also aims at incorporating semantic information to the registration process, but in a different scenario. In this case, instead of supposing that we have pre-trained arbitrary classifiers at our disposal, we are given a set of accurate ground truth annotations for a variety of anatomical structures. We present a methodological contribution that aims at learning context specific matching criteria as an aggregation of standard similarity measures from the aforementioned annotated data, using an adapted version of the latent structured support vector machine (LSSVM) framework.
展开▼
