Simultaneous Consideration of Spatial Deformation and Tensor Orientation in Diffusion Tensor Image Registration Using Local Fast Marching Patterns

Diffusion tensor imaging (DTI) plays increasingly important roles in surgical planning, neurological disease diagnosis, and follow-up studies in recent years. In order to compare the tractography obtained from different subjects or the same Subject at different timepoints, a key step is to spatially align DTI images. Different front scalar or multi-channel image registration, tensor orientation should be considered in DTI registration. Several DTI registration methods have been proposed before, and some of them are based on first extracting the orientation-invariant features and then registering images using traditional scalar or multi-channel registration techniques followed by tensor reorientation. They essentially do not fully use the tensor information. Other methods Such as the piece-wise affine trans formation and the diffeomorphic non-linear registration algorithms use analytical gradients of the registration objective functions by considering the reorientation of tensor during the registration. However, only local tensor information Such as voxel tensor similarity is utilized in these algorithms, which can be regarded as it counterpart of the traditional intensity similarity-based image registration in the DTI case. This paper proposes a novel DTI image registration algorithm, called fast marching-based simultaneous registration. It not only considers the orientation of tensors but also utilizes the neighborhood tensor information of each voxel, which is extracted from a local fast marching algorithm around voxels of interest. Compared to the voxel-wise tensor similarity-based registration, richer and more distinctive tensor features are used in this algorithm to better define correspondences between DTI images. Thus, more robust and accurate registration results can be obtained. In the experiments, comparative results using the real DTI data show the advantages of the proposed algorithm.