Information-Theortestic Approach to Deformation Image Transformations with Application to Brain Image Segmentation

摘要

This thesis presents a systematic study of deformable image transformations for nonrigidly aligning a template to an image. The study concentrates on an information theoretic similarity measure, fluid deformation, and prior shape constraints. The similarity between a target image and a template is measured based on their mutual information. Suitability of the mutual information measure for non-rigid-body image registration is systematically investigated. A mutual information bound is derived, and a gradient calculation, which scales linearly with the volume size, is presented. Modification of a fluid model proposed elsewhere is shown to retain the desirable properties of the deformation while allowing more efficient numerical implementation. Shape information is learned by performing eigenshape analysis on a training set of correct deformations of a single template to several typical segmentations. The most likely deformations are then promoted according to the learned shape information. The shape modeling technique does not require a prelabeling or ordering of points in the training set and can handle multiple shapes simultaneously. Based on these results, a new method is developed for nonrigidly aligning a template to a study image. The approach is robust to a wide variety of contrast variations and supports large, curved geometric variations. This method has been experimentally validated using synthetic, magnetic resonance, and cryosection images. It is also incorporated into a brain image segmentation method under development at the University of Illinois. Potential applications include image segmentation, functional brain mapping, and automatic target recognition.