【作者】 方林；

【导师】 张建伟；

【作者基本信息】 南京信息工程大学 ， 应用数学， 2012， 硕士

【摘要】 目前,心血管疾病已经成为我国仅次于恶性肿瘤的第二号杀手,临床诊断和医学研究表明,借助成熟计算机技术对医学影像进行分析对患有心血管疾病人群的规范诊疗和控制病情进展具有重要作用。然而由于成像机制等因素影响,多数医学图像带有噪声、弱边界、灰度不均匀性(偏移场现象)以及复杂背景等特点,使得早期图象分割方法如图像特征空间分类、基于区域方法以及图论方法等难以得到较好分割结果。基于水平集的活动轮廓模型具有严谨数学理论框架,能够融合解剖组织先验信息,采用光滑轮廓曲线来描述目标边界等优点,逐渐成为医学图像分割领域学者研究重点。活动轮廓模型图象分割模型可分为基于边界模型和基于区域模型。基于传统的边界模型,由于使用梯度信息驱动力使得模型易受到图像噪声或弱边界影响,从而陷入局部最优或造成边界泄,并且对初始曲线设置条件要求较高。基于区域的模型,如小邻域统计信息的模型,虽较好刻画图像小邻域的灰度信息,能够降低偏移场影响,但是邻域窗口大小和邻域各向同性对模型影响较大,使得模型易陷入局部极值。为此,本文对活动轮廓模型应用医学特定目标分割进行深入研究,引入图像统计信息来改进传统边界模型,构造出基于灰度统计信息驱动力,较好反映目标灰度信息,并使得初始化自由度较高。本文主要研究内容和创新点包括以下几个方面：(1).提出了一种基于区域自适应水平集图像分割模型。该算法引入基于区域信息的符号压力函数来控制曲线扩张(或收缩)至目标边界,改进了传统边界模型中人为设置初始曲线,降低了模型参数对图象分割影响。(2).提出了一种基于局部信息的测地线活动轮廓模型。该模型的外力由图像的邻域均值信息构造,从而有效地克服了灰度不均匀对分割结果的影响,有效地分割灰度不均匀的医学图像。(3).提出了一种小邻域统计信息测地线活动轮廓模型。该模型带有原始测地线活动轮廓模型方向性分割特点,模型外力由小邻域统计信息符号压力函数构造,拟合图像局部区域灰度分布,可以快速有效地分割左心室图像和脑部肿瘤图像。(4).提出了一种基于图像局部与全局特征的活动轮廓模型图像分割算法。该算法融合了图像局部信息和全局信息。局部信息结合了图像局部均值和方差信息,以克服图像灰度不均匀的影响。全局信息则可以较好地提高模型处理图像弱边界的能力,并防止模型陷入局部最优。实验结果表明,改进算法分割出较为精确的心脏左心室核磁共振图像。更多还原

【Abstract】 Currently, cardiovascular disease has become the second death cause except malignant tumors in China, the clinical diagnosis and medical studies show that it is important for cardiovascular disease people to diagnose and control disease progression with the mature computer technology. Medical segmentation for the area of targets of interest is the important step for image analysis and understanding. Due to the imaging mechanism and other factors, the majority of medical images contain noise, weak border, intensity inhomogeneity (bias field phenomena) as well as complex background. For early image segmentation methods, such as classification methods based on image feature space, the regional approaches and the methods based on graph theory, they are difficult to get better segmentation results. Compared to the early image segmentation methods, active contour models based on level set, which has the framework of the rigorous mathematical theory, merges prior information of the anatomical organization, could use smooth contours to describe the target boundary, they gradually become the focus of medical image segmentation scholars.The active contour models based on image segmentation could be divided into the models based on edge and the models based on region.Segmentation models based on edge, which use gradient information as driving force, have limited capacity of dealing with image the weak boundary or noise, thus fall into local optimums or cause the border to vent. and require higher initial curve set conditions.The models based on region such as local statistical models. The models use small neighborhood statistical information as driving force to overcome the bias field, but they are sensitive to the size of a small neighborhood window parameter, isotropy of the neighborhood, and quality of image background. These factors are easy to make the models fall into local optimums.We deeply study active contour models in order to segment the area of targets of interest. Introducing statistical information of image to improve the models based on edge, we study small neighborhood statistical information to construct a new driving force which could overcome the bias field, reflect the target gray information, and improve the freedom of the initial contour. Main study content and innovations for this paper include the following aspects:(1). Image segmentation model based on region adaptive level set evolution was presented. Signed pressure force function based on regional information was introduced so that the contour shrieked when outside the object or expanded when inside the object, improved the shortage of artificial setting initial contour, overcame the influence of parameters of the weighted area.(2). A new model of GAC (ggeodesic aactive ccontour) based on local region was presented. Information of the local mean was used to make pressure force to segment images with intensity inhomogeneity. Experimental results with different medical images show that the new model can get the better results in efficient way.(3). A new model of GAC based on local statistical was presented, which has original GAC model with directional segment feature in order to reduce image background of the impact of segmentation results. The local statistics signed pressure force function is used to external force in order to describe the local image intensity distribution. Experimental results with left ventricular and brain tumor images show that the new model can get the better results in efficient way.(4). We proposes image segmentation active contour model based on local and global features. The local fitting term based on means and variances of local intensities to cope with intensity inhomogeneity. The global intensity fitting term, which conquers the unexpected local minimum stemming from image local intensity, deals with weak. The experiment results demonstrate that this algorithm is effective for segmenting the left ventricle MR images.更多还原