Regularization of diffusion tensor images

We present a regularization scheme for diffusion tensor images, that respects the geometrical structure of diffusion ellipsoids and does not introduce artifacts such as anisotropy drops. The method can be stated as a variational problem and solved by means of a gradient flow. The main ingredient is the notion of a distance between two ellipsoids that considers differences in shape as well as differences in orientation. The method is specialized to the case of cylindrically-symmetric ellipsoids and implemented in terms of ordinary vector manipulations such as cross and dot products. The regularization algorithm is tested using a synthetic tensor field and a dataset acquired from a diffusion phantom. In both cases the algorithm was able to reduce the noise from the tensor field.