A kernel-based graphical model for diffusion tensor registration

In this paper, we propose a novel method for the spatial normalization of diffusion tensor images. The proposed method takes advantage of both the diffusion information and the spatial location of tensor in order to define an appropriate metric in a probabilistic framework. A registration energy is defined in a Reproducing Kernel Hilbert Space (RKHS), encoding the image dissimilarity and the regularity of the deformation field in both the translation and the rotation space. The problem is reformulated as a graphical model where the latent variables are the rotation and the translation that should be applied to every tensor and the observed variables are the tensors themselves. Efficient linear programming is used to minimize the resulting energy. Quantitative and qualitative results on a manually annotated dataset of diffusion tensor images demonstrate the potential of the proposed method.