图像分割的水平集模型及其在医学图像分割中的应用研究

发布时间：2018-02-25 06:01

本文关键词： 医学图像分割聚类分割水平集模型几何主动轮廓模型曲线演化理论　出处：《兰州大学》2016年博士论文　论文类型：学位论文

【摘要】：图像分割理论的发展中,水平集模型的出现极大地推动了图像分割领域的研究。水平集模型结合曲线演化理论,应用图像梯度矢量流,有效克服参数模型的不足,大大的扩展了主动轮廓模型的研究。本文在研究水平集分割模型的基础上,对水平集图像分割中的问题进行了探索,从算法模型的结构上、能量函数的优化方面、参数计算的复杂性及函数的最小化优化等方面对医学图像数据的分割方法进行了研究。除此之外,还针对处理不同图像体数据的适应性算法进行了研究,分别提出基于变分计算,曲率理论,先验知识,聚类分析,形态特征和图论优化的算法模型。本文的主要研究内容如下:(1)提出了变分能量拟合水平集算法(GVLS),算法采用全局和局部图像结构信息的思想,引导水平集模型捕捉噪声以外的微小细节,能够较完整的检测图像目标的结构信息。引入PM各向异性滤波算法,进一步应用差分计算使图像特征突显,并结合局部、全局最小化算法模型的特点,改善了水平集模型分割强度有变化、背景复杂及有强烈噪声干扰图像目标的能力。(2)提出了梯度向量流水平集算法(GVFLS),算法中用一种新的能量项来计算水平集模型的计算复杂度与相关约束性的最小化过程。描述过程是基于正则化和曲线曲率的优化思想,权衡了计算的复杂度,用曲率估算达到目标轮廓的精度要求,使水平集算法分割曲线收敛于期望的目标边缘。(3)提出了形态特征先验知识几何主动轮廓算法(MCS),为了在初始分割时能更好的定位分割目标,引入中点圆Hough算法确定左心室图像的圆形结构,算法将基于CV模型的几何主动轮廓模型的图像目标区域信息和先验形状信息表示成速度场,嵌入到主动轮廓的迭代方程中,在先验信息模型的引导下,将图像背景与目标进行初始定位与标记,驱使曲线演化于最终目标边界。(4)提出了基于聚类分割的几何主动轮廓算法(Km GAC),算法模拟了分割曲线的内部和外部区域,将能量函数通过聚类算法进行迭代最小化,将图像的前景和背景进行学习,使演化曲线能量函数降到最小值。聚类算法计算图像目标的前景和背景的分段值,同时修订和发现前景、背景值,经过对算法的优化,算法具有很好的目标边界捕捉能力,演化速度迅速。(5)提出了基于图优化的多相水平集分割算法(MLS),经过改进的水平集方法既不需要求解欧拉方程,也不需要计算任何的偏微分方程,模型的最小化用图分割的思想进行优化。算法对目标的初始化要求低,参数选择自由,曲线演化稳定,收敛迅速。最后,为了对本文中五种算法的性进行比较说明,将五种算法在同一医学脑部数据上进行了比较,并分析了各算法所面对的数据侧重点,对分割效果进行了分析,达到了研究目的。
[Abstract]:In the development of image segmentation theory, the appearance of level set model has greatly promoted the research of image segmentation field. Combining with curve evolution theory, the level set model applies image gradient vector flow to overcome the deficiency of parameter model. Based on the research of the level set segmentation model, this paper explores the problems in the horizontal set image segmentation, from the structure of the algorithm model, the optimization of energy function. In addition to the complexity of parameter calculation and the minimization optimization of function, the segmentation method of medical image data is studied. In addition, the adaptive algorithm for processing different image volume data is also studied. Based on variational calculation, curvature theory, prior knowledge, clustering analysis, The main research contents of this paper are as follows: (1) A variational energy fitting level set algorithm (GVLSU) is proposed, which adopts the idea of global and local image structure information. Guiding the level set model to capture the small details other than noise can detect the structure information of the image target completely. The PM anisotropic filtering algorithm is introduced to make the image feature prominent by using differential calculation and combining with the local image. The characteristics of the global minimization algorithm model improve the variation of the segmentation intensity of the level set model. A gradient vector flow level set algorithm (GVFLSS) is proposed, in which a new energy term is used to calculate the computational complexity and the minimization of correlation constraints of the level set model. The description process is based on the idea of regularization and curve curvature optimization, Considering the complexity of the calculation, the curvature estimation can meet the precision requirement of the target contour. The level set algorithm converges to the desired edge of the target. (3) A priori knowledge geometric active contour algorithm with morphological features is proposed in order to better locate the segmentation target in the initial segmentation. The circular structure of the left ventricular image is determined by the midpoint circle Hough algorithm. The image target region information and the prior shape information of the geometric active contour model based on CV model are expressed into the velocity field and embedded into the iterative equation of the active contour. Under the guidance of the prior information model, the image background and target are initially located and marked. Driven the curve to evolve to the final target boundary. (4) A geometric active contour algorithm based on clustering segmentation is proposed. The algorithm simulates the internal and external regions of the segmentation curve and minimizes the energy function iteratively through the clustering algorithm. By learning the foreground and background of the image, the energy function of the evolutionary curve is reduced to the minimum. The clustering algorithm calculates the segmentation value of the foreground and background of the image object, at the same time, the foreground and background value are revised and discovered, and the algorithm is optimized. The algorithm has a good ability to capture the target boundary, and the evolution speed is rapid. (5) A multi-phase level set segmentation algorithm based on graph optimization is proposed. The improved level set method does not need to solve the Euler equation. There is no need to calculate any partial differential equations, the minimization of the model is optimized by the idea of graph partitioning. The algorithm has low requirements for object initialization, free choice of parameters, stable evolution of curves, and rapid convergence. In order to compare the characteristics of the five algorithms in this paper, the five algorithms are compared on the same medical brain data, and the data focus of each algorithm is analyzed, and the segmentation effect is analyzed to achieve the purpose of the research.
【学位授予单位】：兰州大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】