Symmetric Data Attachment Terms for Large
Mirza Faisal Beg, ALI RAZA KHAN · 2007
Nonrigid medical image registration between images that are linked by an invertible transformation is an inherently symmetric problem. The transformation that registers the image pair should ideally be the inverse of the transformation that regis- ters the pair with the order of images interchanged. This property is referred to as symmetry in registration or inverse consistent reg- istration. However, in practical estimation, the available registra- tion algorithms have tended to produce inverse inconsistent trans- formations when the template and target images are interchanged. In this paper, we propose two novel cost functions in the large de- formation diffeomorphic framework that are inverse consistent. These cost functions have symmetric data-attachment terms; in the first, the matching error is measured at all points along the flow between template and target, and in the second, matching is en- forced only at the midpoint of the flow between the template and target. We have implemented these cost functions and present ex- perimental results to validate their inverse consistent property and registration accuracy.