Learning task-optimal image registration with applications in localizing structure and function in the cerebral cortex

Abstract

In medical image analysis, registration is necessary to establish spatial correspondences across two or more images. Registration is rarely the end-goal, but instead, the results of image registration are used in other tasks, such as voxel-based morphometry, functional group analysis, image segmentation and tracking. In this thesis, we argue that the quality of image registration should be evaluated in the context of the application. Consequently, we develop a framework for learning registration cost functions optimized for specific tasks. We demonstrate that by taking into account the application, we not only achieve better registration, but also potentially resolve certain ambiguities and ill-posed nature of image registration. We first develop a generative model for joint registration and segmentation of images. By jointly modeling registration and the application of image segmentation, we demonstrate improvements in parcellation of the cerebral cortex into different structural units. In this thesis, we work with spherical representations of the human cerebral cortex. Consequently, we develop a fast algorithm for registering spherical images. Application to the cortex shows that our algorithm achieves state-of-the-art accuracy, while being an order of magnitude faster than competing diffeomorphic, landmark-free algorithms. Finally, we consider the problem of automatically determining the "free" parameters of registration cost functions.

(cont.) Registration is usually formulated as an optimization problem with multiple tunable parameters that are manually set. By introducing a second layer of optimization over and above the usual registration, this thesis provides the first effective approach to optimizing thousands of registration parameters to improve alignment of a new image as measured by an application-specific performance measure. Much previous work has been devoted to developing generic registration algorithms, which are then specialized to particular imaging modalities (e.g., MR), particular imaging targets (e.g., cardiac) and particular post- registration analyses (e.g., segmentation). Our framework provides a principled method for adapting generic algorithms to specific applications. For example, we estimate the optimal weights or cortical folding template of the generic weighted Sum of Squared Differences dissimilarity measure for localizing underlying cytoarchitecture and functional regions of the cerebral cortex. The generality of the framework suggests potential applications to other problems in science and engineering formulated as optimization problems.