Fluid registration of medical images using Jensen-Rényi Divergence reveals 3D profile of brain atrophy in HIV/AIDS

We propose a novel fluid image registration strategy based on an information-theoretic measure, the Jensen-Rényi Divergence (JRD) of two images. We modified the definition of JRD, which is based on the joint histogram of two images, to develop a variational approach in which driving forces are applied throughout the deforming image to maximize the JRD between it and the target image. A viscous fluid regularizer was applied to guarantee diffeomorphic (i.e., smooth, one-to-one) deformation mappings. The resulting partial differential equation (PDE) was solved iteratively by convolving the applied force field with the Green's function of the linear differential operator [1]. The fluid JRD method provided accurate, robust correspondences for registrations requiring large deformations in 2D and 3D. Finally, we applied our algorithm to tensor-based morphometry (i.e. shape analysis) of 3D brain MRIs from 26 HIV/AIDS patients and 14 matched healthy control subjects, showing that the algorithm can help identify subtle and clinically significant differences in brain structure. Detected white matter changes were correlated with cognitive impairment in AIDS. These techniques may help measure and visualize disease burden in drug trials and in morphometric studies of degenerative brain disease. (see www.loni.ucla.edu/~mchiang for colored figures).