Member of Technical Staff, Statistics and Learning Research DepartmentBell Laboratories, Alcatel-Lucent
A Deformation Model for Images
Large quantities of medical images are acquired daily at nearly every medical center in the United States, but statistical models and associated softwares that would facilitate the automated analyses of these images are lacking. The goal of this research is to develop methodology that will make such automated image analysis possible. The methodology proposed is based on an atlas-based deformation model which finds a one-on-one mapping from the atlas image to a target image in the same image class. The deformation is carried out by generalized landmarks called ``facets''. Knowledge about the atlas image is thus transferred to the target image through facets. A large number of facets are placed in the volume of the atlas (often on a lattice). Each of them is then located in the target image. The model for the new location has two components: a Markov random field prior with pair-wise difference on a nearest neighborhood system, and a likelihood component based on the agreement of features in the atlas and target image. A new measure of feature difference was introduced that is robust under a variety of conditions. The iterative conditional modes (ICM) algorithm is used to obtain the maximum a posterior estimate of facet locations. The model was used to automatically segment magnetic resonance mouse brain images. It was also applied to inter-subject registration of human brain images. Both qualitative and quantitative evaluation of the results are presented.