Statistical analysis of medical images
Abstract
To compare brain images of the same subject; taken at slightly different orientations or of different subjects altogether, the first step is to align the images so that corresponding pixels in the two images represent homologous biological points. In this thesis, an approach to such image alignment using thin-plate splines is presented. The pixels are assumed to be noisy samples of underlying smooth functions on continuous domains, and the aim is to map the coordinates of one image domain onto those of the other. No pixel imputation is performed in the estimation of the mapping as is done in other commonly used alignment algorithms that work on discrete domains, and it is not necessary to identify landmarks in the images. The approach is nonparametric, with smoothness constraints enforced by placing a certain roughness penalty on the estimate of the mapping rather than by assuming some smooth parametric form. Among issues addressed are the isotropic invariance of the approach, the computation, and the smoothing parameter selection by generalized cross-validation. Simulated brain image data examples are used to illustrate the technique.
Degree
Ph.D.
Advisors
Gu, Purdue University.
Subject Area
Statistics|Biomedical research
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.