An information theoretic approach for non-rigid image registration using voxel class probabilities

摘要: We propose two information theoretic similarity measures that allow to incorporate tissue class in non-rigid image registration. The first measure assumes probabilities have been assigned each of the images be registered by prior segmentation both them. One is then non-rigidly deformed match other such fuzzy overlap corresponding voxel object labels becomes similar ideal case whereby probability maps are identical. Image assessed during registration divergence between and actual joint distributions images. A second proposed applies a available for only one images, instance an atlas matched study guide thereof. Intensities minimizing conditional entropy intensities given image. derive analytic expressions gradient with respect individual displacements force field drives process, which regularized viscous fluid model. performance class-based evaluated context inter-subject atlas-based MR brain compared maximization mutual using intensity information. Our results demonstrate incorporation significantly improves classes after matching. methods here open new perspectives integrating single output used other.