Multimodal brain volume registration using information theory
This thesis is an investigation of techniques used to register medical image volumes of different modalities. A detailed survey of existing image registration technologies is performed. An unstructured, automated technique employing the maximization of mutual information is found to be superior. Using this technology, a registration routine is developed to register three-dimensional brain volumes. The performance of the routine is evaluated using translated and noise-corrupted magnetic resonance and computed tomography brain volumes. The routine is found to achieve sub-voxel accuracy. Two novel approaches are presented to improve the speed of the registration procedure: a seed generating algorithm to reduce the optimizational complexity of the problem, and a selective sampling schedule based on information content to reduce the computational requirements of the cost function. An evaluation of these modifications finds them to reduce overall registration time by two thirds.