Diffusion Tensor Registration Using Probability Kernels and Discrete Optimization

In this report, we propose a novel diffusion tensor registration algorithm based on a discrete optimization approach in a Reproducing Kernel Hilbert Space (RKHS) setting. Our approach encodes both the diffusion information and the spatial localization of tensors in a probabilistic framework. The diffusion probabilities are mapped to a RKHS, where we define a registration energy that accounts both for target matching and deformation regularity in both translation and rotation spaces. The six-dimensional deformation space is quantized and discrete energy minimization is performed using efficient linear programming. We show that the algorithm allows for tensor reorientation directly in the optimization framework. Experimental results on a manually annotated dataset of diffusion tensor images of the calf muscle demonstrate the potential of the proposed approach