Physics-Based Nonrigid Registration for Medical Image Analysis

In this paper we describe a new algorithm for nonrigid registration based on an elastically deformable model. The use of registration methods has become an important tool for computer-assisted diagnosis and therapy. Our goal was to improve analysis in various applications of neurology and neurosurgery by improving nonrigid registration. A local gray level similarity measure is used to make an initial sparse displacement field estimate. The field is initially estimated at locations determined by local features after which a linear elastic model is used to infer the volumetric deformation across the image. The associated partial differential equation is solved by a finite element approach. A model of empirically observed variability of the brain was created from a dataset of 155 young adults. Both homogeneous and inhomogeneous elasticity models were compared. Although a two and three dimensional version of our method was implemented, we focus on the description of the three dimensional case. For computational efficiency a parallel implementation of each part of the algorithm was developed. The algorithm has been used for various medical applications including interpatient registration of MR scans of the brain, intraoperative images of neurosurgery showing brain shift, a study of gait and balance disorder and preliminary results for a template driven segmentation approach for the caudate. ii Zusammenfassun