This paper presents a kernel-based method of correspondence detection in diffusion tensor images (DTI), a key step towards their deformable registration. The proposed method is driven by a few focus points chosen in white matter, characterized by a unique morphological signature which incorporates the anisotropy, orientation and the anatomic context by using "oriented" Gabor filters and several candidate matches for each focal point, defined using tensorial similarity of these orientation specific signatures. The final focal point correspondences are defined via minimization of a function that seeks to satisfy three criteria : sparsity, which enforces unique correspondence, tensorial similarity of Gabor features, and smoothness, and are obtained by solving a constrained non-linear optimization problem with inequality bound constraints, using an optimization solver that employs primal-dual interior point algorithms and which ensures global convergence. The solution of the optimization problem produces the best correspondences for the focal points and uses these correspondences to obtain the optimal kernelized interpolation parameters for non-focal points. Experimental results on human brain data in which datasets with tumor are matched with normal brains, demonstrates the ability of the method in determining very good correspondences in the white matter, and its applicability to datasets with large mass effect as in tumors.