【作者】 马守明；

【导师】 程显毅；

【作者基本信息】 江苏大学 ， 计算机应用技术， 2007， 硕士

【摘要】 在临床诊断和病理学研究中,为了准确地分辨医学图像中的正常组织结构和异常病变,需要对医学图像进行分割。由于医学图像对比度较低,加之组织特征的可变性、不同组织之间或者组织和病灶之间边界的模糊性以及微细结构(如血管、神经)分布的复杂性,使得医学图像分割成为一个难点。论文在分析了国内外有关医学图像分割方法相关文献的基础上,基于Agent原理与技术,提出了一种新的基于MAS(Multi-Agent System)的医学图像分割方法。论文的主要工作及创新点是:(1)对国内外典型医学图像分割方法进行了深入地分析,综述了图像分割的基本方法,并针对医学图像的特点,阐述了基于MAS的医学图像分割方法的可行性和必要性。(2)对MAS协作求解机制和Agent群体强化学习进行了详细介绍。讨论了几种典型的MAS体系结构,指出MAS不仅在结构上存在相关性,而且在行为上也存在相关性,这些相关性正是MAS协作的起因。将传统的强化学习技术引入到MAS中,形成Agent群体强化学习机制。(3)在上述工作基础上,通过对典型MAS组织结构优缺点的分析,提出一个基于Agent图(Agent Graph)的三级MAS协作组织模型,该模型较好地兼顾了系统对通信开销、效率、可靠性等方面的要求。以个体强化学习和Agent组(Agent Team)的概念为基础,设计了一种引入先验知识的强化函数,提出了基于Agent组的群体强化学习算法。(4)针对医学图像分割,提出一种图像分割的高层模型,对该模型的各个部分进行了详细介绍,指出了模型设计中的几个关键问题。根据该模型,给出了基于MAS的医学图像分割技术的流程和关键算法。(5)利用论文的分割算法对图像进行了分割实验,人脑模拟图像组织的分割和FCM方法的分割视觉效果对比,人脑冠状面MR图像的MAS分割和ML分割及CGS分割的比较以及异常脑组织的分割实验结果表明,论文的基于MAS的自适应分割算法较好地完成了图像的分割,视觉效果良好,具有很好的适应性,分割效果与领域专家的意见基本一致,在医学应用上具有实际意义。(6)基于集对分析SPA(Set Pair Analysis),给出了分割方法评价模型,并用该方法对论文的实验工作进行了初步的分析比较。更多还原

【Abstract】 It is required to segment the medical images in order to distinguish normal tissues and abnormal pathological changes in the clinic diagnose and pathology research. As the usually lower contrast of medical image, the changes of tissue character, the fuzzy character of different tissue or between tissue and focus and the distributing complexity of imperceptibility structure, such as vas and nerve, the medical image segmentation becomes a difficult problem.After a detailed investigation about the development of medical image segmentation in domestic and foreign related fields, the paper brings forward a new method of medical image adaptive segmentation based on MAS (Multi-Agent System). Its main contents and innovations include:(1) This paper provides an elaborately analysis about typical medical image segmentation methods of domestic and foreign related fields, shows a more comprehensive overview of the image segmentation technology. The feasibility and necessity of medical image segmentation based on MAS was expounded.(2) This paper carries out a detailed introduction of the mechanisms of MAS collaboration and the reinforcement learning of Agent groups. After the discussion of several typical MAS architecture, this paper points out the MAS collaboration is aroused by the relevance which existing in the structure and behavior of MAS. The introduction of traditional reinforcement learning techniques into MAS forms the Agent group reinforcement learning mechanism.(3) Based on the above works, a three-tier MAS collaboration organization model was proposed based on Agent graph after an analysis of the advantages and disadvantages of the typical MAS organizational structure. The model has better balance between the communication costs, efficiency, reliability, and so on. Using the individual reinforcement learning and the concept of Agent team, this paper designs a reinforcement function which introduced to priori knowledge, brings forward a group reinforcement learning algorithm based on Agent group.(4) For the purposes of medical image segmentation, the paper gives a high-level modal of image segmentation and a detailed introduction about the model, and points out several key issues in the design process. According to the model, we give the flow and key algorithm of the image segmentation technology. (5) Image segmentation experiments were conducted using our algorithm. The vision effect contrast of different tissues of brain simulation image by MAS and FCM techniques, the human brain coronal MR image segmentation by MAS, ML and CGS segmentation and the abnormal brain tissue segmentation experimental validate this method is effective to realize the segmentation. The result of experiment indicates the method can implement the automatic segmentation and its vision effect is excellent with good adaptability. The segmentation effect is basically consistent with the expert of medical image segmentation field and significative in medical applications.(6) This paper presented a succinct segmentation technology evaluation model which based on Set Pair Analysis. Our experimental results were analyzed and compared using this method.更多还原