【作者】 林亚忠；

【导师】 陈武凡；

【作者基本信息】 第一军医大学 ， 生物医学工程， 2004， 博士

【摘要】 图像分割是指将图像划分成一系列彼此互不交叠的匀质区域。作为一项最基本技术，它在图像分析、图像压缩等图像处理领域发挥极其重要作用，尤其是精确的医学图像分割在临床诊断中是必不可少的。 基于吉伯斯随机场的先验模型通常被用于解决退化图像病态逆问题正则化求解，并通过提供良好的空间上下文约束信息，在贝叶斯医学图像分割中广泛运用；然而，在临床分割中，由于复杂的医学结构和图像的退化现象，导致了该模型在正则化过程中，需要适当改进以适应临床的不同需求。因此，本文针对该模型开展了深入系统地研究，并提出一系列相应的解决算法。 首先，本文针对吉伯斯随机场在分割中参数估计难的问题，通过统计与训练，提出联合的最大似然与最大后验估计方法，在迭代中完成参数估计并实现对图像的吉伯斯贝叶斯分割； 其次，本文针对引入高阶邻域空间约束信息在医学图像分割中所面临的尴尬问题，通过扩展单一分辨率的马尔科夫模型到多分辨率领域，提出一种混合金字塔随机场模型，只需考虑二阶邻域就能实现传统单一分辨率下只有引入高阶邻域才能更好解决的分割精度和效率问题； 其三，本文针对医学图像多类模糊分割所面临的瓶颈问题，通过建立一种新颖的广义模糊吉伯斯随机场模型，分别从先验模型和似然模型入手，提出一套适合医学图像多类模糊分割的理论和技术方法； 另外，本文针对水平集在解决复杂组织结构和形状拓扑关系改变分割过程中遇到的边界泄漏问题，设计出一种自适应的广义模糊速度场，通过提供更鲁棒的边界信息和更可靠的运动停止力，解决了传统以梯度图为边界信息的边界泄漏问题。 本文通过大量的实验验证了所提模型与其改进方法的有效性。更多还原

【Abstract】 Image segmentation is to separate an image into a lot of un-overlapped and homogeneous regions. As a fundamental technique, image segmentation has being played most key role in the image processing field, such as image analysis, image compression and so on. Especially, precisely segmenting the regions for some medical images is essential to clinical diagnosis.The Gibbs random prior model is often used to solve the ill-posed inverse problems in regularization for degraded image, and also to medical Bayesian segmentation due to providing an excellent spatial contextual constraints information. However, the classical GRFs model must be revised in the process of regularization to meet the clinical needs because of the complicated structure and degraded phenomenon in medical image. In the paper, some researches about the model have been developped deeply and systematically, and a series of approaches have been proposed to address them correspondingly.Firstly, in order to perform the parameters estimation about segmentation based on GRFs, a method fusion of maximum likelihood with maximum a posterior has been introduced after training data of an image and getting image statistic to solve the problems of parameters estimation and Bayesian segmentation based on GRFs during the iteration.Secondly, a hybrid pyramid Gibbs random model is provided, by extending a single MRFs to a multi one, to overcome the embarrassment derived from high neighborhood system used to describing the spatial contextual constraints. By using the proposed model, second order neighborhood system is enough to solve the problems on segmentation precise and its efficiency which are performed well only by a high one for a single MRFs.Thirdly, a novel generalized fuzzy Gibbs random model is constructed to overcome the bottleneck brought by multi-class fuzzy segmentation in medical images. Moreover, a series of theories and techniques about fuzzy segmentation are derived from the models of prior and likelihood.Additionally, an adaptive speed term based on generalized fuzzy operator is proposed to replace the traditional gradient-based edge map applied in level set segmentation in order to solve the problem of boundary leakage, which is expected to provide more robust edge estimation and more reliable information used as stopping criteria for curve evolution in dealing with the topology changing of the shape and the complexity of medical structures.A lot of experiments are also provided to prove the validity of the models and their corresponding approaches mentioned in the paper.更多还原