Source: Dissertation Abstracts International, Volume: 61-05, Section: B, page: 2669.

Director: James Scott Duncan.

This thesis develops an information theoretic registration framework where the segmentation and registration of dual anterior-posterior and left lateral portal images to a treatment planning three-dimensional computed tomography (CT) image is carried out simultaneously and iteratively. The proposed registration framework is termed the minimax entropy registration framework as it has two steps, the max step and the min step. Appropriate entropies are evaluated in each step in order to segment the portal images (the max step) and to estimate the registration parameters (the min step). The registration framework is based on the intuition that if some structure can be segmented in the portal image, the segmented structure, in addition to the gray-scale pixel intensity information, can be used to better estimate the registration parameters. On the other hand, given an estimate of the registration parameters, information from the high resolution 3D CT image dataset can be used to guide segmentation of the portal images. Performance analysis and comparisons to other registration methods demonstrates the robustness and accuracy of the proposed registration framework.

To further improve the estimated segmentation of the portal images and the accuracy of the estimated registration parameters, correlation among the image pixel intensities is modeled using a one-dimensional Markov random process. Line processes are incorporated in the Markov random process model which estimate the edges between the segmented regions. As a future research direction, we propose to incorporate the estimated edges in the min step to further improve the registration. The proposed framework is independent of the image dataset and hence, in general, can be straightforwardly extended to register any low resolution, low contrast image to a high resolution, high contrast image.