Improving Deformable Multi-Grid B-spline Registration using Heuristics and Deep Learning

The purpose of deformable image registration is to recover acceptable spatial transformations that align two images acquired at different periods of time into the same coordinate system. Due to the growing importance of modern image-guided therapies, the techniques continue to be widely investigated. B-spline based models are commonly used in modeling this type of transformations because of the robustness of modeling continuous dense deformations. However, shortcomings exist in state-of-art B-spline based models, such as the incapability of modeling discontinuous motions and the difficulty in clinical application due to requiring the estimation of an intractable number of free parameters. This thesis develops adaptive B-spline registration models for 3D image volumes and investigates the application of machine learning techniques to improve the accuracy and efficiency in several aspects. The specific contributions are: Reduce the computational burden in parameter optimization for both single grid and pyramid B-spline models, which makes it possible for registration of large-scale 3D image volumes. On one hand, the reduction is accomplished by introducing a parallel framework that partitions each image volume into independent tiles and rapidly solves an analytic formulation of the B-spline algorithm. On the other hand, the degree of freedom is dynamically defined by inspecting local regional features and thus dramatically reduces the number of parameters to optimize. Automatically learn B-spline parameter configuration that best suitable for expressing the deformation underlying registration. The use of multiple grids of varying spatial resolution has been thoroughly investigated due to the sensitivity of the optimization process to local minima while attempting to accurately recover complex local deformations. An octree structured B-spline pyramid construction pipeline, as well as a convolutional neural networks (CNN) based sparsity learning technique, are proposed to infer B-spline grids prior optimization to accurately model anatomical deformations. An approach that improves and evaluates registration quality with fully automatic landmark detection and matching on lung CT images. The respiratory deformation between images acquired from different phases is learned using CNNs to interpolate the correspondence of candidate landmarks on images. The final set of landmarks are integrated into registration to algorithmically improve the registration quality by providing higher-level knowledge of the anatomical form of the images.