Research On Robust Registration Methods For Multimodality Medical Images

According to the physical principles for imaging, multimodality medical imagesare usually divided into two types: structural and functional images. Structural imagesprovide mainly high-resolution images with geometric and anatomical information.Functional images provide metabolic or neurochemical changes characteristic but withcoarser resolution. In order to fuse complementary characteristics of the underlyinganatomy and tissue microstructure from medical images of different modalities, thegeometric registration is a preliminary and crucial step and has a paramount influenceon the performance of image fusion. Medical image registration is defined as theprocess that determines the best structural and physiological correspondence betweentwo or more medical images.In multimodality medical imaging, medical images are sometimes blurringbecause the adverse influence of the radioactive ray, tracer and high magnetic field forhuman body must be reduced. Moreover, there are the physical limitations of theimaging process itself, such as noise, limited resolution, insufficient contrast orinhomogeneity. Therefore, robust registration of multimodality medical images is avery challenging problem. The main achievements of this dissertation are as follows:(1). To improve the robustness and precision of the maximization of mutualinformation (MI) similarity, a principal ordinal feature and hybrid entropy basedregistration method is presented. A principal ordinal feature (OF) is defined and used torepresent the spatial information between the neighboring pixels and the properties ofthe specified micro-structure in medical images. Integrating with pixel intensities, asimilarity measure based on hybrid entropy is defined to register multimodality images.The proposed method is demonstrated using several pairs of multimodality medicalimages and the experimental results show that the noise of images can be effectivelysuppressed and compared with some existing methods, the proposed registrationalgorithm is of higher precision and better robustness.(2). Due to the nonlinear relationships of high-dimensional OFs in nature, a locallinear embedding (LLE) and hybrid entropy based registration method is proposed. Forhigh dimensional OFs, the LLE algorithm is used to dimensionality reduction and theinverse mapping of LLE is used to fuse complementary information of OFs together.Then a novel similarity measure based on hybrid entropy which integrates intensity with OF is defined to register multimodality images. Through quantitative evaluations,the proposed measure is proven to provide improved robustness with accuracy.(3). Non-rigid registration of medical images has become a challenging task inmedical image processing and applications. A LLE and improved L-BFGSoptimization based registration method is proposed. A hierarchical transformationmodel of medical images is developed. The global motion is modeled by an affinetransformation while the local motion is described by a free-form deformation (FFD)based on B-splines. Then the proposed LEE and hybrid entropy based similaritymeasure is chosen as the registration function. Finally an improved L-BFGS algorithmis used to search the optimal registration parameters. We evaluate the effectiveness ofthe proposed approach by applying it to the simulated brain image data. Theexperimental results show that the proposed registration algorithm is of higherprecision and faster speed. However, it cannot be applied to image registration withlarge anatomical variation in noisy environment.(4). Based on a new physical model of image registration, a locality preservingprojection (LPP) and particle swarm optimization (PSO) based registration method isproposed. A large deformation between the reference image and the floating image isdecomposed into a series of small deformations and the result of every deformation istreated as a frame in the registration video. Therefore, the registration process isequivalent to a smooth and continuous video and the registration space is on a certainnonlinear manifold of lower dimensionality. However, the registration manifold isusually too complicated to build the correspondence with numerous deformationparameters. Hence, medical images are projected into manifold space by the LPPalgorithm and the optimal registration parameters are searched by the PSO algorithm.Though the proposed method is not as good in precision as what we expect because ofdimension reduction, it can be widely used to many applications.