Hierarchical Diffusion Tensor Image Registration Based on Tensor Regional Distributions

INTRODUCTION: Diffusion tensor imaging (DTI) is capable of non-invasively measuring water diffusion in vivo. While DTI has been widely em delineate potential white matter abnormality in different neurological diseases, registration of diffusion tensor images across different subjects i prerequisite for detailed statistical analysis on voxel-by-voxel basis. However, spatial normalization of diffusion tensor images is challen technically and computationally given that tensor data representation is high dimensional in nature, and thus it requires not only to spatially also to appropriately reorient the tensor at each voxel. Conventional DTI registration methods generally extract tensor scalar features from ea as such to construct scalar maps. Subsequently, regional integration and other operations such as edge detection can be performed to ex features to guide the registration. There are, however, two major limitations with these approaches. First, the computed regional features reflect the actual regional tensor distributions. Second, by the same token, gradient maps calculated from the tensor-derived scalar feature m not represent the actual tissue tensor boundaries. To overcome the two major limitations associated with the currently available approach approach is proposed by performing regional computation and edge detection directly on the tensors. Regional tensor distribution information mean and variance, is computed in a multiscale fashion directly from the tensors by taking into account voxel neighborhood of different sizes, a capturing tensor information at different scales, which in turn can be used to hierarchically guide the registration. Such multiscale scheme alleviate the problems of trapping in a local minimum. Regional information is also more robust to noise since one can better determine the properties of each tensor by taking into account the properties of its surrounding, mitigating the effect of noise. Also incorporated in our metho information extracted directly from tensors, which is crucial to facilitate registration of tissue boundaries. Detailed descriptions of this new app provided below. METHOD AND RESULTS: Tensor Regional Distributions: For each voxel in diffusion tensor images, multi-scale regional tensor distribution in was extracted from its multi-scale neighborhoods. Specifically, for a given tensor D(x, y, z), the regional tensor distribution information was extracted from its neighborhood {D(u, v,w)|(u, v, w) N(x, y, z)}. By varying the size of N(x, y, z) or by scaling the image itself, a rich set of multi-scale tensor regional distribution information could be obtained and employed to drive the registration hierarchically. To avoid tensor computation in a curved space, we took the matrix logarithm of all the original tensors, i.e., log(D(x, y, z)), for all (x, y, z). The tensor regional mean and variance for each voxel were then computed in the log-space. Tensor Edges: To better extract tissue boundaries, we extended Canny edge detector to work directly on diffusion tensor images instead of their scalar maps. Canny edge detector can be used to extract image gradient boundaries, and is robust to noise due to the employment of Gaussian filter to smooth out the noise prior to edge detection. For fast edge detection, 3D Gaussian-based image filtering was implemented using three subsequent steps of 1D Gaussian filtering along the x, y, and z directions independently. This was then followed by gradient map computation and nonmaximum suppression in the 3D space. Note that edge detection was performed in the log-space. Hie Deformable Registration: For each voxel, features obtained above were grouped into an attribute v shown in Fig. 1, these attribute vectors are rich enough to permit discrimination of different brain a structures. For computation efficiency, a subset of voxels with distinctive attribute vectors was selec set of driving voxels, providing temporary landmarks for correspondence matching. The initial driving voxels were employed to obtain a mo starting registration, facilitating relatively less distinctive driving voxels during later stages of the registration. The brain volume was deformed linear hierarchical fashion similar to that in [1]. Simulated Deformation Fields: We generated 20 simulated deformation fields, which serv ground truth, using the statistical model of deformation (SMD) proposed in [2]. Using a set of brain images as the template, 20 simulated brain images can be constructed, which were then registered back onto the template using the proposed method. Registration accuracy can then be evaluated by comparing the deformation fields with the ground truth. The average Euclidean distance deformation field error given by the proposed method is 0.67 voxel (std: 0.43), indicating a subvoxel registration accuracy. In comparison, Yang et al.’s method which was tested using a similar dataset [3], reported a mean error of 0.86 voxel; the proposed method yields more than 20% improvement. Fiber Trac Using FACT [4], fiber bundles passing through two ROIs (see Fig. 2) were tracked, extracted, and com quantifying registration accuracy in these specific regions. Using a measure similar to that referred to as the m closest distances in [5], errors given by the proposed method are 0.62mm and 0.77mm (stds: 0.16 & 0.07) for and splenium fibers, respectively. The rather small errors with our approach signify good registration accuracy Simulation: We introduced atrophy on the simulated brains, by modifying the tensors in the selected region to isotropic (Fig. 3). After registering them onto the template, we tested whether the atrophic region was still detect paired t-test. The results are shown in Fig. 3. The average t values for affine registration and our method are 4.26 and 9.91, respectively. Rea brain images of real subjects were registered onto a randomly selected template and an average image was generated from the registered im FA edge map of the average image, superimposed on the FA map of the template image (Fig. 4), indicates good consistency between their FA REFERENCES: [1] D. Shen et al., IEEE TMI, 21(11), 1421–1439, 2002. [2] Z. Xue et al., Neuroimage, 33(3), 855–866, 2006. [3] J. Yang et al., Imaging’08, 2008. [4] S. Mori et al., A. Neurology, 47(2), 265–269, 1999. [5] G. Gerig et al., IEEE EMBS, 4421–4424, 2004. Fig. 4 FA edge map of the group-averaged tensor image superimposed on template FA map. (a) Point of interest (b) Similarity Ma