【作者】 周頔；

【导师】 须文波；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2011， 博士

【摘要】 高精度、高速度的医学图像配准及图像分割技术,因其在临床上能辅助医生作出更准确、更快速的诊断而受到广泛的关注。但又因其处理对象的特殊性和实现的难度,一直就是研究界的重点攻克项目。本论文的目的是提高现有医学图像处理工作站中的两个重要功能模块:图像配准及图像分割的性能。准确的图像配准对综合利用多次成像的信息或者多种成像设备的信息,达到信息互补,进而提高诊断效率和正确性十分重要。但困难是,一般情况下配准操作中使用到的相似性测度函数都是具有局部极值的,而且医学图像,尤其是三维体数据图像往往都数据量非常庞大。图像分割是图像处理领域和计算机视觉中的一个基本而关键的问题,目前在工程上应用最广泛的仍是交互式半自动分割技术。交互式分割方法中最常用的有两种:Snake分割算法以及Live-wire分割算法。但在应用中发现,他们各自都存在一些缺陷:无论是经典Snake算法还是后来推广范围最大的GVF Snake算法,在分割带噪声的图像时,都很容易受到噪声点的干扰而无法正确逼近目标物体边界;而传统的Live-wire算法则因为其自身计算最小代价路径的方式,而无法一次性分割物体的大曲率边缘。有鉴于此,本论文的主要内容是结合医学图像处理工作站中经常配备的三个具体功能项:数字减影CT颅脑血管造影、CT肺实质分割、DSA心脏轮廓提取,着重在精度更高、速度更快的图像配准方法上,以及交互式图像分割技术上,进行理论及应用的研究。首先,本论文在医学图像配准上的研究及成果,主要集中于以下几点。设计了一种准确、快速的医学图像配准方法,并且进而开发出以该配准方法为基础的数字减影CT颅脑血管造影技术。该方面的工作围绕以下两点展开:通过提高优化搜索算法的性能来应对因相似性测度函数而造成的局部极值问题,以及采用多分辨率的全局优化算法与局部优化算法相结合的方法在保证精度的前提下加快对三维图像数据的处理速度。具体为:在提高优化搜索算法的性能方面,通过分析SPSO算法及QPSO算法各自的特性及存在的缺陷,提出了一种引入混合概率分布和多样性控制策略的改进型QPSO算法(DRQPSO)。经实验手段确定出使DRQPSO算法具有最优性能的三个参数。与其他优化算法的对比实验证明,提出的DRQPSO算法在搜索精度及收敛速度方面皆具有优越的性能,尤其适合多峰函数的优化任务。在将DRQPSO算法引入图像配准过程之后,对于含有噪声的图像配准精度符合亚像素要求的概率上升至91.67%。在加快三维图像配准速度方面,本论文采用多分辨率策略对原始大尺寸三维图像进行分层处理,并将全局优化算法DRQPSO与局部优化算法POWELL结合,在不同层级的图像上进行最优化参数搜索,使含有噪声的三维医学图像的配准在保证精度的前提下符合了实时要求。并且基于该种多分辨率混合优化配准算法完成了数字减影CT颅脑血管造影任务。其次,本论文在医学图像分割的研究及成果上,主要集中于以下两点。提出了一种新的Snake外力场---GCF,该外力场是通过边缘映射图的卷积方式扩展形成的。这是为了克服经典Snake算法以及GVF Snake算法容易受到图像噪声点的干扰而无法正确逼近目标物体边界的问题。实验证明,基于这种GCF外力场的Snake算法在图像包含大量噪声的情况下,也能在保留边缘信息的前提中排除掉噪声带来的不良影响,正确收敛到目标物体的真实边界上,而且在深凹部位的收敛速度远远快于GVF Snake。该算法应用于CT肺实质分割中,能符合临床精确度要求。提出了一种基于可控平均代价路径的新型Live-wire算法。这是为了改善传统Live-wire算法无法一次性分割大曲率边缘的缺陷。该算法从代价函数的构造以及最优路径的产生这两点上进行改进,使路径的代价与这条路径的长度脱离关系,从而可以一次性分割曲率较大的目标物体边界,减少了人工设定的节点的个数,进而减少了整个分割过程的耗时。该算法应用到包括DSA心脏轮廓提取在内的多种图像分割中,取得了满意的效果。论文最后对所做的工作与主要研究成果进行了总结,并提出了进一步的研究方向。更多还原

【Abstract】 Accurate and robust medical image registration and segmentation play a crucial role in clinical diagnosis by providing fast and vivid method to observe organs. However, since the particularity and the complexity of the processing object, image registration and segmentation are widely taken change to regarded.The aim of our research is to improve the two maily function modules of medial grapics workstation: image registration and image segmentation.Precise registration can increase treatment efficiency and minimize neurological damage by providing complementary information which acquired from the same patient but at different times or with different devices. Otherwise, the precise and quick registration has been proved to be a difficult task due the the tendency of trapping into the local optima when using mutual information as the similarity metrics, and the dramatic large size of medical image data.Image segmentation is the key problem both in image processing and machine vision fields. And interactive segmentation methods including Snake and Live-wire, are the most widely used segmentation technologies in clinical applications. But they both have drawbacks: traditional Snake and the GVF Snake algorithms are both too sensitive to the noise, and traditional Live-wire can not segment the edge with large curvature disposably.Therefore, the contents of this paper mainly involves the research of the accurate and quick image registration and segmentation methods, as well as the application in CT head digital subtraction angiography, CT lung segmentation and the extraction of the DSA heart contour.First, the accurate and quick image registration research focuses on the following aspects.Improve the performance of the optimization method in registration. Present a multi-distribution-based and diversity-controlled novel QPSO algorithm (DRQPSO). Determine the best suitable parameters which makes the DRQPSO most efficient through experiments. The results of the DRQPSO on several benchmarks show that it may have better performance than QPSO and PSO, especially in solving the multimodal functions. The semi-pixel satisfaction rate of noisy image registration raises to 91.67% when using the DRQPSO as the optimization method.Accelerate the registration process by the use of multiresolution strategy and the hybrid optimization algorithm combined DRQPSO and Powell, which makes the 3D image registration meet the real-time acquirement under the precondition of semi-pixel precision. Furthermore, apply this accurate and quick registration method to the CT head digital subtraction angiography successfully.Second, the research on the image segmentation involves the following two aspects.Propose a new external foce for Snake algorithm, called gradient convolution field (GCF). GCF is calculated by convolving the edge map generated from the image with the user-defined convolution kernel, with the aim to overcome the noise sensitivity of the traditional Snake algorithm. Experiments and comparisons with GCF are presented to show the advantages of this innovation, including superior noise robustness, reduced computational cost, and the flexibility of tailoring the force field. When applied this GCF Snake to CT lung segmentation, the results meet the clinical precision acquirement.Develop a improved Live-wire algorithm based on the controllable average cost path, overcome the disadvantage of tradition Live-wire which is unable to segment the the edge with large curvature disposably. The improved Live-wire algorithm change the cost terms function and redefine the optimal path as the minimum average cost path between two seed points, which frees the optimal path from the restriction of the straight line. Experiments conducted on a variety of image types including DSA heart images, have shown that this improved Live-wire algorithm requires less seed points than the traditional algorithm when delineating the same boundary and as a result, reduces the time required to complete the calculation.The main contributions in this work are summaried at the end and future research directions are also put forward.更多还原