Medical Image Registration Guided by Application-Specific Geometry

Abstract

Image registration is an important task in medical image processing. Among its applications are inter-patient registration to perform segmentation of organs, registration of follow-up scans to propagate the in tissue accumulated radiation dose of a radiotherapy, and registration to perform deformation analysis over time or within a population. Generally, depending on the type of images and the nature of the spatial variation between the images, application specific registration settings need to be chosen, such as the image similarity metric, the type of optimizer and the number of degrees of freedom for the transformation model. For various sorts of applications the options available in a general registration algorithm are limited to obtain good registration results since they do not exploit application specific geometry knowledge. Specific applications can benefit from prior shape knowledge of a population, geometric properties of structures in the image, or knowledge about discontinuities in the deformation field. In this thesis various extensions to a general registration algorithm are proposed to tailor the algorithm to the issues in the applications involved. The method proposed in Chapter 2 deals with inter-patient registration of patients with cervical cancer. Between patients are large variability in organ shape and position is observed that requires large and complex deformations. To guide registration in finding these deformations a statistical shape, trained on the shape of the segmentations of the population, is incorporated as penalty term in the optimization process. In Chapter 3 intra-patient registration of the images acquired for external beam radiation therapy and brachytherapy is performed. The missing volume of the applicator, as used in cervical brachytherapy, is modeled as a surface mesh and its volume is minimized during registration. For registration of images with sliding organs, in Chapter 4 we propose a new transformation model that is accompanied by a geometric penalty term. The proposed method is applied on inhale-exhale CT scans in which the lungs slide along the thoracic cage. Additionally, examples of registration of synthetic images and a registration of a patient with cervical cancer are given. Chapter 5 proposes a improved similarity metric for groupwise registration. The computation of the groupwise metric that consists of the normalized cross-correlation between all image pairs in the group is reduced to a linear complexity. In the last Chapter, Chapter 6, a summary and a general discussion is given.