Multichannel large deformation diffeomorphic metric mapping and registration of diffusion tensor images

Image registration is an important process in medical image analysis and computational anatomy. Images of different subjects that are acquired from different scanners at different times will be in different coordinate systems. Therefore, registration is an indispensable pre-processing step for the analysis and visualization of images in a common coordinate system.;In this dissertation, we concentrate on the registration problem of diffusion tensor magnetic resonance images (DTMRI or DTI). We utilize the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework for the solution of this problem. In its original form, the LDDMM algorithm calculates a nonlinear transformation between scalar valued images. As our main contribution, we extend LDDMM to be a multichannel image matching algorithm. We seed scalar valued images obtained from symmetric second order tensor field of DTI data as channels to the multichannel LDDMM algorithm. Then, we calculate a transformation between tensor images by optimally matching the channel images together. The resulting transformation is used to reorient and transform the tensor images, hence performing the desired registration.;The multichannel LDDMM is a nonlinear registration algorithm between vector valued (multichannel) images. This method can be used to calculate diffeomorphisms between two subject images when more than one modality image (such as T1-weighted, T2-weighted magnetic resonance image or computed tomography image) are available for the subjects. The multichannel LDDMM can also be used to register binary region of interest (ROI) label images of substructures in the brain. For example, in two brain images if different substructures are segmented and labeled, each label can be used as a channel to the algorithm. Then, a transformation can be calculated between these brain images.;In this thesis, we also use the multichannel LDDMM algorithm to estimate templates for DTI datasets. Using the well established expectation-maximization algorithm and a series of multichannel LDDMM calculations, we estimate DTI templates that are better representations of data populations.