【作者】 陈志彬；

【导师】 邱天爽；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2010， 博士

【摘要】 本论文的目标是开发自动分割算法,将脑MRI图像分割成不同的组织,即脑白质、脑灰质和脑脊液,为脑发育与人的衰老、有关脑疾病的诊断和治疗、脑外科手术规划和导航等应用提供定量脑测度信息。在本论文中,提出了几个非参数化变形模型和统计信息或模糊信息相结合的算法,从不同模态的MRI图像中分割出白质、灰质、脑脊液等脑组织。对提出的算法采用实验的方法进行验证和评估,实验图像包括T1-加权、T2-加权和PD-加权的仿真和真实的脑MRI图像。论文首先提出了一个基于直方图分析的非参数变形模型算法。算法中,用混合高斯模型(mixture Gaussian Model, MGM)来对图像的灰度直方图进行建模,利用期望最大化(Expectation Maximization, EM)算法来估计混合高斯模型中各分量的参数。获得的参数用于产生新的约束项,以指导水平集曲线的进化,并最终完成脑MRI图像的组织分割。用仿真和真实的MRI图像对算法进行了评估,论文中提供了MRI图像的分割结果,并对分割结果进行了定量评价。其次,对基于区域的几何活动轮廓(Region-based Geometric Active Contour, RGAC)模型进行了研究。在稳定性分析基础上,采用新的区域压力项对RGAC算法进行了改进。新算法解决了原算法存在的稳定性问题。与原算法相比,改进算法的迭代次数明显减少,分割结果对参数的敏感度较低,因而有更好的韧性。算法能够从T1-加权、T2-加权和PD-加权等不同模态的MRI图像中分割白质、灰质、脑脊液等脑组织。用10个MRI仿真图像和5个真实MRI数据集对新算法进行评估,并与其它算法的分割结果进行了比较,验证了算法改进的可行性和有效性。算法中采用的模糊区域指示子函数被推广应用到模糊自适应水平集算法中,它是几何轮廓模型的一个改进算法。在曲线进化过程中,算法能够自适应地调整曲线进化的方向,达到快速收敛的目的。同时也克服了经典活动轮廓模型算法对图像的梯度信息过度依赖和因高斯平滑造成的边界定位精度下降的问题。文中通过仿真和真实的MRI图像对算法进行了评估。论文最后提出了一个模糊C-均值和水平集方法相结合的多类算法。算法由一组常微分方程组成,一个组织类别由一个水平集函数表示。算法由各向异性扩散滤波、模糊聚类分析以及水平集分割方法等三个主要阶段组成。算法用人工合成图像、20幅仿真MRI图像和10个真实MRI图像集进行了评估。与多相算法相比,多类算法降低了计算复杂度,能够以更快的速度收敛。与其它算法相比,多类算法有更好的分割性能和噪声鲁棒性。算法中采用的模糊逻辑,考虑了MRI图像中脑组织的模糊性和不确定性,与硬分割算法性比,能够包含更丰富的信息。与水平集方法相结合,有利于提高算法的性能和韧性。更多还原

【Abstract】 The research goal in this dissertation is to develop an automatic segmentation method to segment brain MRI images into different tissue classes (gray matter, white matter, and cerebrospinal fluid), to provide quantitative brain measurements to the study of brain development and human aging, disease diagnosis and treatment, surgical planning and navigation, and other applications. In this dissertation, we develop several algorithms which integrate the non-parametric deformable models with statistical information or fuzzy information of images to segment the brain MRI images. These algorithms are assessed and validated with the experiments on multi-modalities of MRI images:T1-weighted, T2-weighted and PD-weighted.We firstly present a histogram-analysis based non-parametric deformable model, where the intensity histogram of the MRI image is modeled via the mixture Gaussian model (MGM). The parameters of each component in MGM are estimated via the Expectation Maximization (EM) algorithm. Then the estimated parameters are used to generate the constraint term to guide the evolution of the level set curves to achieve the brain tissues segmentation. The algorithm is evaluated with the simulated and real MRI images. The segmentation results and quantitative analyses are provided.We then explore the region-based geometric active contour (RGAC) of Suri. Based on the stability analysis, we propose the improved algorithm of the RGAC with new regional force terms. The new algorithm solves the underlying stability problem associated with the original algorithm. Compared with the original algorithm, the improved algorithm achieves convergence with less iteration number, and its segmentation results are less sensitive to some parameters. The algorithm can segment brain tissues into gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF) classes from different modalities of MRI images (T1-weighted, T2-weighted, and PD-weighted). The algorithm is evaluated with ten simulated MRI images and five real MRI datasets, and compared quantitatively with other algorithms. The proposed fuzzy region indicator is used in an adaptive level set method. The method adaptively adjusts the directions of fronts during the curves evolution and achieves the final segmentation with fast convergence rate. The algorithm overcomes the limitations, associated with geometric active contour, of overly relying on the gradient information of images and reducing in accuracy of boundary locating caused by Gaussian smoothing. The algorithm is evaluated with the experiments on simulated and real MRI images.Finally, we present a multiclass algorithm by integrating fuzzy clustering analysis with the level set methods. The algorithm uses a set of ordinary differential equations; each of them represents a class to be segmented. The algorithm consists of the anisotropic diffusion filtering, fuzzy clustering analysis and the level set refining segmentation steps. The algorithm is also evaluated with synthetic images,20 simulated MRI images and 10 real MRI datasets. Compared with the multiphase algorithm, the multiclass algorithm reduces the computational complexity, can achieve faster convergence. The comparison with other algorithms indicates the better segmentation performance and good robustness to noises.The adopted fuzzy logic framework in the proposed algorithms allows for the ambiguity and uncertainty of the brain tissues in MRI images, can retain more information than the crisp approaches. The combination of fuzzy C-means (FCM) with level set methods allows to improve the segmentation performance and robustness of the algorithms.更多还原