Research On Key Techniques In Medical Image Registration

With the development of medical imaging,an increasing number of clinical applications require comparison and analysis of images acquired from different subjects,from the same subject at different imaging times,or even from different imaging modalities.Medical image registration is a crucial procedure in image analysis and computational anatomy,it is widely used in clinical practice for disease diagnosis,surgical navigation,brain atlas and various medical evaluations.Due to the multimodality,complexity and noncontinuity of medical images,medical registration is a challenging task.Structures of medical images are usually nonlinear or highly nonlinear,such as diffusion tensor images(DTI)lying on nonlinear manifold.However,no matter mono-or multi-modal,rigid body or non-rigid body,parametric or non-parametric,most existing registrations either neglect the nonlinear geometry of manifold-valued data,directly work in Euclidean space which is a linear space,or could not take into full consideration for the topological structure of the feature space and spatial information,having great influence for preserving topological structure.In this dissertation,we systematically study a few of medical imaging techniques,especially for magnetic resonance imaging(MRI)and diffusion tensor imaging(DTI).Mathematically,topology,differential geometry and geometrical algebra are utilized as spatial analysis tools.Topologically structure of nonlinear medical data and spatial relationships among data are intensively studied.We concentrate on registration algorithm on accuracy,robustness and topological homeomorphism.The major contributions of this dissertation are as follows.(1)Most non-parametric based diffeomorphic algorithms assume the grey value constancy and ignore the plenty and topological peoperties of nonlinear structure of high-dimensional data,which can physically preserve reasonableness.Based on diffeomorphic Demons algorithm,a locally adaptive topology preservation for diffeomorphic Demons(LATPDD)registration in Magnetic Resonance Imaging is propose.To caputer richer spatial information and geometric structures,firstly,we construct symmetric positive definite(SPD)matrices and form a high-dimentional Lie group manifold under certain constraints.Next,manifold learning is used to adaptively select appropriate neighborhoods for each sample voxel,leading to the more exact local linear approximation.Therefore,linearization degree of tangent space is improved.In the result,local nonlinearization structure of manifold is preserved.The topological structure of the feature space under diffeomorphic transformation is physically plausible.(2)Due to the orientation feature of diffusion tensor images,tensors need to be reoriented during affine registration.However,traditional schemes either limit its application on rigid deformation or suffer from computation complexity since iteration.A canonical form-based affine registration of diffusion tensor images,named as CFARD,is proposed in order to overcome these limitions.We transform voxel sets into canonical forms,matching voxel sets under affine transformation is considered equivalent to find matching canonical forms of two sets under rotation transformation.Affine registration problem is simplified as rigid registration.In addition,the effect of non-rigid components is preserved in the reorientation process leading to a more biologically plausible transformation.Our model outperforms conventional reorientation-based models in accuracy,the main reson is that conventional models only extract the rigid rotation component of affine transformation,but discards the deformation component,that is to say,no contribute comes from translation,scaling,shearing of transformation to fiber orientation.Furthermore,to improve computation load caused by tensor reorientation,we use a rotation group--Lie group SO(3)to describe mathematical structure of diffusion tensor.For an efficient optimition,SO(3)is converted to the quaternion rotation form to seek a closed-form solution of absolute orientation,in which no iteration is required.Therefore,computational complexity is reduced.(3)Mutual information based multimodal registration fails to consider the image structure information and spatial relationship among pixels,and assumes that there exists a global statistical relationship between anatomic individuals.A machine learning based algorithm is proposed for multimodal image registration of medical images.Firstly,takes advantage of a linear dynamic model(LDM)to describe high-dimensional spatial nonlinear structure of the image,where feature space contains more spatial information.Then to parameterize LDM and constitute Lie group elements,which form Riemannian manifold.Secondly,embeds Riemannian manifold to a high-dimensional Hilbert space.Finally,incorporates kernel functionss into Hibert feature space to seek an optimal similarity measure.Kernel tricks map the originally nonlinear feature space into a implicit and high-dimensional vector space.The algorithm can efficiently works on both rigid and affine registrations...