Tools for Deformable Image Registration

Medical imaging sensors can be used to noninvasively probe tissue morphology and monitor material deformations associated with growth, disease, or normal physiology. Deformable image registration provides a framework for extracting and quantifying this information. Image registration is, however, an inherently ill-posed inverse problem. The research in this dissertation investigates techniques to regularize image registration problems. Preprocessing, integral and post-processing tools were developed to complement existing image registration techniques. These tools included techniques for dynamic (sequential) spatial filtering, histogram-matching, sequential image tracking, rezoning, and other techniques that allow the solution of difficult practical registration problems. These tools have been incorporated into the continuum mechanics based deformable template registration method known as Warping. Warping was then used to solve a variety of typical registration problems, utilizing regularization tools when needed. Additionally, two quantitative techniques for assessing image registration results have been developed and implemented. The singular value decomposition (SVD) topology tracking technique hierarchically assesses the match in topology between image data sets. The technique also provides a systematic method of tracking the evolving topology of the deformed template data during registration and comparing it with the topology of the target image data. In order to apply the technique to volumetric image data, a three-dimensional (3D) extension of the SVD was formulated. The technique was applied to registration problems using both simulated and real image data. Results indicated that spikes in the inner product map of the template may correspond to potential difficulties in the registration. The second technique utilized the variances of the registration potential energy to evaluate the relative influence of image data and registration method on image registration solutions. An image influence parameter was defined, which provided a regional measure of the relative influences of the image energy and registration method over the domain of the template model. The technique was applied to several representative image registration problems. Results indicated that choice of image registration method played a significant role in registration results in image regions of conflicting or inadequate image texture.