Registration of diffusion tensor magnetic resonance images and its application to the quantitative analysis of human brain white matter

This thesis concerns with the registration of the diffusion tensor (DT) magnetic resonance images which store a 3-D rank-2 Cartesian tensor at each voxel. This problem is of technical interest because registering tensor-valued images has to face a number of challenges unique to the handling of tensor values. In particular, tensors are not rotation-invariant and warping of tensor-valued images is more complex than that of scalar-valued images. This research is also driven by an important clinical application: the study of human brain white matter (WM) differences across populations using DT images. A prerequisite to such studies is the image spatial normalization that maps individual subject images onto a common atlas to remove shape differences between individuals, which in turn requires the registration of DT images. In this thesis, we describe a novel algorithm for this purpose. The algorithm is formulated in the standard optimization framework for image registration. Our contribution is in devising an objective function that handles the complexity of tensor image warping analytically. The resulting algorithm can optimize tensor orientations explicitly and can assess image matching with metrics that compare DTs as a whole. It captures nonlinear shape differences between individuals by incrementally estimating the solution transformation and at each increment we estimate a piecewise affine transformation. A novel similarity metric for DTs is also presented and a computationally efficient implementation of the algorithm is demonstrated using the conjugate gradient optimizer with analytic derivatives. As an application of the registration algorithm, we describe an iterative method that constructs a WM atlas from a set of input DT images and simultaneously establishes the correspondences between the input images and the atlas. The strength of the approach lies in its ability to better align WM structures by utilizing full information available to DTs, which is a result of using the proposed algorithm for registration. We demonstrate the advantage of this approach over alternatives that register images of rotation-invariant scalar values derived from DTs using image data from a clinical study. We show that the proposed approach improves the statistical significance of the detected WM differences.