【作者】 王利；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 核磁共振成像(Magnetic Resonance Imaging, MRI)技术已经成为脑疾病临床诊断的重要辅助手段。准确地分割脑MR图像对脑部解剖、脑疾病诸如阿尔茨海默氏病、帕金森氏病以及精神分裂症的分析与研究等具有重要的指导意义。然而在实际应用中,由于射频场的不均匀性等因素,导致脑MR图像的灰度均匀性变差,其表现为同一组织的像素灰度沿空间呈缓慢平滑的变化,直接导致传统的基于全局灰度值的分割模型分割失败；其次,脑MR图像成像过程中由于受仪器设备等物理原因影响,使得图像中经常含有噪声,影响分割精度。除此之外,图像中存在Partial Volume(PV)效应,即各个软组织之间边界比较模糊,不明确和不连续。这些都给脑MR图像的准确分割造成了很大困难。和成人脑MR图像相比,新生儿脑MR图像分辨率更低,PV效应更加显著；由于新生儿的大脑处在迅速发育中,导致图像中的灰度不均匀性的成因更加复杂；白质／灰质对比度的逆转,使得处在脑脊液和灰质交界处的像素具有和白质相似的灰度,使得分割更加困难。所有这些使得脑MR图像的分割仍然是一个充满挑战的课题。大部分脑MR图像分割方法是基于单个像素的,是离散的,容易受到初始化的影响。为此,本文对成人脑MR图像和新生儿脑MR图像分割进行了深入研究,首先采用连续的方法,提出了一些基于活动轮廓的分割模型来克服脑MR图像中存在的灰度不均匀性；然后,进一步将这些分割模型统一到一个凸优化框架中,同时分割脑MR图像和估计偏移场,特别地,由于采用了凸优化技术,可以很好地克服初始化的影响；最后我们进一步将该凸优化框架模型和活动轮廓模型应用到新生儿脑MR图像分割中,取得了较好的效果。所做的主要工作和研究成果如下：(1)为了克服灰度不均匀性,提出了一种基于图像局部均值的分割模型。利用多相位水平集方式来拟合局部灰度,并采用一种简单而有效的初始化方法,达到更快更准地分割脑MR图像。由于采用了多相位水平集方法,该模型可以同时得到白质、灰质和脑脊液。另外水平集方法很好地保证了分割结果的光滑性,可以得到光滑的边界／曲面。(2)为了克服对初始化的敏感性,提出了一种基于局部均值和全局均值的活动轮廓模型。该模型分别定义了一个局部拟合力和全局拟合力。其中,局部拟合力用于驱动曲线向目标靠近并在目标边界处停止,全局拟合力用于驱动远离目标边界的曲线向目标靠拢。在这两个拟合力的作用下,该模型对初始曲线较不敏感同时克服脑MR图像中存在的灰度不均匀性。(3)提出了一种基于局部高斯概率的活动轮廓分割模型。通过分析每个像素邻域的灰度分布,利用窗口函数定义了一个局部高斯拟合能量。局部高斯概率中的两个参数,局部均值和局部方差是随着空间变化而变化的,因此本文方法可以很好地克服脑图像中存在的灰度不均匀性。值得一提的是,本文方法的局部均值和局部方差不再是显式定义的,而是可以通过变分法原理严格推导出来的。3T和7T的脑MR图像分割实验证明本文方法不仅可以有效地克服灰度不均匀性,还可以克服低对比度。(4)提出了一种新颖的基于凸优化的脑MR图像分割模型,同时进行图像分割和估计偏移场,因此可以克服灰度不均匀性的影响。特别地,由于该模型关于图像分割变量是凸的,因此对初始化具有很强的鲁棒性。另外该模型也可以克服噪声的影响。和传统的基于梯度下降法求解相比,我们采用Split Bregman方法对该模型进行快速求解,使得对一幅2D的图像处理时间一般不超过1秒,3D的数据处理时间一般不超过100秒。大量的二维和三维实验证明本文方法无论在时间效率上还是准确度上都取得了满意的结果。而且,我们还进一步给出一个基于凸优化的脑MR图像分割框架模型,该框架综合利用了图像的局部统计信息。比较实验也证明了本文提出的框架模型是有效的。(5)提出基于凸优化和耦合水平集方法的新生儿脑MR图像分割框架。耦合水平集方法综合利用了图像的局部灰度统计信息、基于Atlas的先验信息、大脑皮层的厚度信息。为了克服初始化对分割结果的影响,我们采用凸模型进行分割,然后进行必要的校正,从而为耦合水平集方法提供一个可靠的初始化。该新生儿脑MR图像分割框架已经在10个新生儿数据上取得了满意的结果,和目前流行的新生儿MR图像分割方法相比,本文方法取得了较高的分割精度。更多还原

【Abstract】 Accurate segmentation of magnetic resonance (MR) images of the brain is undoubtedly of great interest for the study and the treatment of various pathologies such as Alzheimer dis-ease, Parkinson or Parkinson related syndrome. In fact, intensity inhomogeneity often occurs in MR images due to radio frequency (RF) coils or acquisition sequences. The intensity inho-mogeneity in MR images often appears as an intensity variation across the image. Thus the resultant intensities of the same tissue vary with the locations in the image. Although usually hardly noticeable to a human observer, such a bias can cause serious misclassifications when intensity-based segmentation algorithms are used. In addition, images are often corrupted by various noises that challenge the segmentation. Besides these difficulties, the boundaries be-tween tissues are quite fuzzy due to the partial volume (PV) effect. Compared with the adult brain MR images, contrast in neonatal MR images is much lower than that of adult because the majority of white matter is as-yet unmyelinated and has a water content closer to that of gray matter than in adults and adolescents. Besides the image contrast, the intensities of tissues are significantly affected by intensity inhomogeneity due to not only RF inhomogeneity but also biological properties of the developing tissue, which leads to a large overlap in their intensity distributions. The inversion of contrast between GM and WM, compared to adult MRI, is also a difficulty given the limited resolution of neonate MRI. Due to this inverted GM/WM con-trast, many voxels between CSF and GM can be incorrectly classified as WM by conventional intensity-based segmentation approaches. In a word, it remains challenging to segment brain MR images, especially for neonatal brain images. In this paper, we first propose a few segmen-tation methods based on active contour model to deal with the intensity inhomogeneity. These proposed active contour models are then intergraded into a convex framework to simultaneously segment the brain MR images and correct bias fields. The convexity of the framework makes the model robust to initialization. We further extend the convex framework and active con- tour model proposed previously into neonatal brain MR image segmentation. Our work mainly includes the following parts:(1) To overcome the difficulty caused by intensity inhomogeneity in the segmentation of magnetic resonance (MR) images, this paper presents a new multiphase level set method for segmentation of brain MR images. The proposed model utilizes local image intensities, which enables it to cope with intensity inhomogeneity. With the multiphase level set framework, the model can extract brain white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) simultaneously and provide a smooth contour/surface. Comparisons of 2D and 3D segmentation prove that the proposed model is effective.(2) We propose an improved region-based active contour model in a variational level set formulation. We define an energy functional with a local intensity fitting term, which induces a local force to attract the contour and stops it at object boundaries, and an auxiliary global intensity fitting term, which drives the motion of the contour far away from object boundaries. Therefore, the combination of these two forces allows for flexible initialization of the contours. The proposed model is first presented as a two-phase level set formulation and then extended to a multi-phase formulation. The proposed method has been applied to brain MR image segmen-tation with desirable results.(3) We propose a new region-based active contour model in a variational level set for-mulation for image segmentation. In our model, the local image intensities are described by Gaussian distributions with different means and variances. We define a local Gaussian distri-bution fitting energy with a level set function and local means and variances as variables. The means and variances of local intensities are considered as spatially varying functions to handle intensity inhomogeneities. It is worth noting that the local intensity means and variances, which are two variables of the proposed energy functional, are strictly derived from a variational prin-ciple, instead of being defined empirically. Comparisons with classic methods demonstrate the advantages of the proposed method in terms of accuracy. (4) We propose a novel method to simultaneously segment the brain MR images and cor-rect the bias fields. The proposed energy is multiphase and convex with respect to its partition variables. Therefore, our method is very robust to the initializations. We also propose a fast and accurate minimization algorithm based on the Split Bregman method for our energy. The flexibility of the method is shown with 2D and 3D segmentation examples of brain MR images. Moreover, we propose a novel convex framework utilizing local image statistical information for simultaneously segmentation of the brain MR images and correction of the bias fields. Many models can be seen as the special cases of the proposed framework. Comparisons also demon-strate the advantages of the proposed convex framework.(5) We present a novel surface-based method for neonatal brain segmentation. Our method effectively utilizes local image information, atlas prior knowledge, and cortical thickness con-straint for guiding the segmentation, by integrating them into a coupled level set method. We also provide a robust initialization method using convex optimization for this coupled level set method. Our proposed method has been validated on 10 subjects with promising results.更多还原