医学图像配准的关键技术研究
发布时间：2018-08-12 08:59
【摘要】：随着医学成像技术的发展,越来越多的临床应用要求对来自不同主体,不同时期或不同成像设备的医学图像进行比较和分析。医学图像配准是医学图像分析和计算解剖的一个关键步骤,被广泛应用在疾病诊断、手术导航、人脑图谱和各种医学评价等方面。然而,医学图像的多样性、复杂性和非连续性等特征使得医学图像配准技术具有很大的挑战性。医学图像数据(如扩散张量图像)通常都是非线性结构,存在于非线性的流形上。已有的配准技术,无论单模态还是多模态,刚体还是非刚体,基于参数还是非参数,要么忽略流形的非线性几何,直接在线性的欧式空间下进行研究,要么对非线性数据结构所包含的丰富的空间信息考虑不够。然而这些信息对空间变换下图像拓扑结构的保持具有重要意义。本文较系统地对几种医学成像技术进行了研究,尤其是磁共振成像技术和扩散张量成像技术。以拓扑学、微分几何和几何代数作为空间分析的数学工具,对医学数据的特征拓扑结构以及数据间的空间关系进行深入探讨,围绕配准算法的鲁棒性,精度和拓扑保持性对非线性医学成像数据的高维空间分布关系进行研究。本文的主要贡献如下。(1)传统的非参数微分同胚配准算法只是基于像素灰度恒定的假设,忽略了高维空间变换中数据流形的非线性结构的丰富性和拓扑性对保持合理物理结构的影响。本文在微分同胚Demons算法的基础上,提出了一种局部自适应拓扑保持的MR图像配准。为了获得更丰富的空间信息和几何结构,首先构造正定对称矩阵,并在一定条件下形成高维非线性的李群流形,然后利用流形学习方法进行自适应的邻域选择,从而更精确地逼近流形的线性切空间,保持流形的非线性结构,使图像特征空间的拓扑结构在非线性的微分同胚变换中更好地保持物理合理性。(2)针对传统的DT图像中张量的重定向策略只适合于刚体配准或因迭代产生的计算代价,本文把张量集转换成一种点集的规范式,提出一种规范式下的DT图像仿射配准。在这种规范式下,仿射变换下两个张量集之间的配准就可以转换为旋转变换下规范式之间的配准,但仍然保持了仿射变换中的非刚体形变分量对重定向的影响,使得形变在解剖结构上更合理。传统的基于重定向的DT图像仿射配准算法只提取刚体旋转部分进行空间变化,忽略平移,缩放和切变等形变分量对仿射变换的影响。所以比较传统方法,本文算法用于仿射变换能获得更好地精度。进一步,为了改善由于重定向引起的计算代价,使优化过程更有效率,利用旋转变换群---李群SO(3)描述弥散张量的特殊数学结构,用单位四元数旋转替代三维旋转矩阵,从绝对定向中找到最优闭式解,能大大减少计算代价。(3)针对传统多模态配准方法忽视图像的空间结构和像素间的空间关系和假定灰度全局一致,提出一种基于学习理论的多模态图像配准算法。本文利用自回归线性动态模型描述图像的局部高维非线性空间结构,这使得特征空间包含了更多空间信息。然后通过参数化动态模型构造出具有李群结构的群元素,并形成黎曼流形,接下来把黎曼流形嵌入到更高维的再生核希尔伯特空间,再在核空间引入核函数寻找最优的相似性测度,这种核技巧可以把非线性数据映射到一个隐式的高维中进行处理。算法不仅对刚体配准,而且对仿射配准也适用。
[Abstract]:With the development of medical imaging technology, more and more clinical applications require the comparison and analysis of medical images from different subjects, different periods or different imaging equipments. However, the diversity, complexity and discontinuity of medical images make medical image registration challenging. Medical image data (such as diffusion tensor images) are usually nonlinear structures and exist on nonlinear manifolds. Rigid or non-rigid, based on parameters or non-parameters, or ignoring the nonlinear geometry of manifolds, directly study in linear Euclidean space, or the rich spatial information contained in the non-linear data structure is not considered enough. Several medical imaging techniques, especially magnetic resonance imaging and diffusion tensor imaging, are systematically studied. Topology, differential geometry and geometric algebra are used as mathematical tools for spatial analysis. The characteristic topological structure of medical data and the spatial relationship between medical data are discussed in detail. The robustness of registration algorithm is focused on. The main contributions of this paper are as follows. (1) The traditional non-parametric differential homeomorphism registration algorithm is based on the assumption that the pixel gray level is constant, ignoring the richness and topological pairs of the nonlinear structure of the data manifold in the high-dimensional space transformation. Based on the differential homeomorphism Demons algorithm, a locally adaptive topology preserving MR image registration algorithm is proposed in this paper. In order to obtain more spatial information and geometric structure, a positive definite symmetric matrix is constructed, and a high-dimensional nonlinear Lie group manifold is formed under certain conditions, and then a manifold is used. Adaptive neighborhood selection is used to approximate the linear tangent space of the manifold more accurately and keep the nonlinear structure of the manifold, so that the topological structure of the image feature space can keep the physical rationality better in the nonlinear differential homeomorphism transformation. (2) The traditional DT image tensor redirection strategy is only suitable for rigid bodies. In this paper, tensor set is transformed into a normal form of point set, and a new affine registration method for DT images is proposed. In this form, the registration between two tensor sets under affine transformation can be transformed into the registration between the normalizations under rotational transformation, but the affine transformation is still preserved. The influence of non-rigid deformation components on redirection makes the deformation more reasonable in anatomical structure. The traditional affine registration algorithm based on redirection only extracts the rotational part of rigid body for spatial change, ignoring the effects of translation, scaling and shear on affine transformation. Further, in order to improve the computational cost caused by redirection and make the optimization process more efficient, the special mathematical structure of the dispersion tensor is described by the rotational transformation group-Lie group SO(3). The unit quaternion rotation is used instead of the three-dimensional rotation matrix to find the optimal closed-form solution from the absolute orientation, which can be greatly reduced. (3) Aiming at the neglect of spatial structure and spatial relationship between pixels and the assumption that gray level is globally consistent in traditional multi-modal registration methods, a multi-modal image registration algorithm based on learning theory is proposed. Spaces contain more spatial information. Then, group elements with Lie group structure are constructed by parameterized dynamic model, and Riemannian manifolds are formed. Next, Riemannian manifolds are embedded into higher dimensional reproducing kernel Hilbert space, and kernel functions are introduced into kernel space to find the optimal similarity measure. The algorithm is not only suitable for rigid registration, but also for affine registration.
【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2178575
[Abstract]:With the development of medical imaging technology, more and more clinical applications require the comparison and analysis of medical images from different subjects, different periods or different imaging equipments. However, the diversity, complexity and discontinuity of medical images make medical image registration challenging. Medical image data (such as diffusion tensor images) are usually nonlinear structures and exist on nonlinear manifolds. Rigid or non-rigid, based on parameters or non-parameters, or ignoring the nonlinear geometry of manifolds, directly study in linear Euclidean space, or the rich spatial information contained in the non-linear data structure is not considered enough. Several medical imaging techniques, especially magnetic resonance imaging and diffusion tensor imaging, are systematically studied. Topology, differential geometry and geometric algebra are used as mathematical tools for spatial analysis. The characteristic topological structure of medical data and the spatial relationship between medical data are discussed in detail. The robustness of registration algorithm is focused on. The main contributions of this paper are as follows. (1) The traditional non-parametric differential homeomorphism registration algorithm is based on the assumption that the pixel gray level is constant, ignoring the richness and topological pairs of the nonlinear structure of the data manifold in the high-dimensional space transformation. Based on the differential homeomorphism Demons algorithm, a locally adaptive topology preserving MR image registration algorithm is proposed in this paper. In order to obtain more spatial information and geometric structure, a positive definite symmetric matrix is constructed, and a high-dimensional nonlinear Lie group manifold is formed under certain conditions, and then a manifold is used. Adaptive neighborhood selection is used to approximate the linear tangent space of the manifold more accurately and keep the nonlinear structure of the manifold, so that the topological structure of the image feature space can keep the physical rationality better in the nonlinear differential homeomorphism transformation. (2) The traditional DT image tensor redirection strategy is only suitable for rigid bodies. In this paper, tensor set is transformed into a normal form of point set, and a new affine registration method for DT images is proposed. In this form, the registration between two tensor sets under affine transformation can be transformed into the registration between the normalizations under rotational transformation, but the affine transformation is still preserved. The influence of non-rigid deformation components on redirection makes the deformation more reasonable in anatomical structure. The traditional affine registration algorithm based on redirection only extracts the rotational part of rigid body for spatial change, ignoring the effects of translation, scaling and shear on affine transformation. Further, in order to improve the computational cost caused by redirection and make the optimization process more efficient, the special mathematical structure of the dispersion tensor is described by the rotational transformation group-Lie group SO(3). The unit quaternion rotation is used instead of the three-dimensional rotation matrix to find the optimal closed-form solution from the absolute orientation, which can be greatly reduced. (3) Aiming at the neglect of spatial structure and spatial relationship between pixels and the assumption that gray level is globally consistent in traditional multi-modal registration methods, a multi-modal image registration algorithm based on learning theory is proposed. Spaces contain more spatial information. Then, group elements with Lie group structure are constructed by parameterized dynamic model, and Riemannian manifolds are formed. Next, Riemannian manifolds are embedded into higher dimensional reproducing kernel Hilbert space, and kernel functions are introduced into kernel space to find the optimal similarity measure. The algorithm is not only suitable for rigid registration, but also for affine registration.
【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前2条
1 闫德勤;刘彩凤;刘胜蓝;刘德山;;大形变微分同胚图像配准快速算法[J];自动化学报;2015年08期
2 许鸿奎;江铭炎;杨明强;;基于改进光流场模型的脑部多模医学图像配准[J];电子学报;2012年03期
相关博士学位论文 前1条
1 周志勇;医学图像非刚性配准方法研究[D];中国科学院研究生院（长春光学精密机械与物理研究所）;2013年
，本文编号：2178575
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2178575.html
