【作者】 邹刚；

【导师】 孙即祥；

【作者基本信息】 国防科学技术大学 ， 电子科学与技术， 2010， 博士

【摘要】 医学显微图像自动识别与分析一直是生物医学工程领域的研究热点。论文围绕鼻咽癌显微细胞图像中有形成分提取过程和基于协同模式分类识别方法的关键技术展开研究,在细胞图像的有形成分提取方面主要研究了细胞图像的滤波、分割技术、细胞边缘的提取技术、重叠细胞的分割及重叠区域的修复技术;在鼻咽癌细胞的协同模式分类识别方法方面重点研究了协同模式识别的学习方法、协同神经网络的优化方法和协同模式识别的不变性等关键技术,并将协同模式识别应用到了鼻咽癌细胞的智能分类识别上,并对各种算法都进行了实验和分析。在鼻咽癌细胞的有形成分提取方面,首先分析了细胞图像的主要降质因素及噪声模型,采用了基于梯度与各向异性扩散方程的自适应去噪方法,在选择性过滤图像背景噪声和脉冲噪声的同时,保留了图像中目标边缘细节信息。针对于细胞图像分割,提出了一种基于粒子群优化模糊C均值与Markov随机场的耦合聚类分割算法,以分割细胞和组织,利用模糊C均值算法局部搜索的特点,将粒子群优化聚类结果作为后续FCM算法的初始值,同时采用Markov随机场与模糊聚类的耦合策略计算适应度函数,在考虑灰度信息的同时,利用像素的空间信息对分割的影响以提高分割的效果。在细胞边缘提取方面,改进了多尺度形态边缘检测算法,用不同尺度的结构元素分别检测出图像的不同尺寸的边缘信息,然后采用证据加权的融合方法对不同尺寸的边缘图像进行融合,通过细胞的边缘信息获取单个细胞。针对显微病理图像中常有的细胞重叠现象,提出了一种组合细胞散点图和改进Snake模型的重叠细胞分割方法及重叠区域的修复方法;在获取边界重叠掩膜图像的基础上,针对因重叠导致重叠区域图像信息变化的情况,采用了一种自适应迭代卷积的快速图像修复方法。协同模式识别学习算法是协同模式分类识别主要研究的内容,本文从四个角度出发对协同模式识别的学习方法进行了改进,首先改进了协同原型模式的修正方法,利用粒子群的全局优化能力控制修正力度,以减少样本“过修正”现象,获取最优协同原型模式。接着,从原型模式和伴随模式同时学习的角度出发,提出了一种基于记忆梯度法的协同模式学习算法,利用协同势能函数的进化过程,将记忆梯度法引入到协同进化的动力学过程,将求解原型模式和伴随模式归结为求解非线性最优化问题,以同时获得最优原型模式和伴随模式。然后,从降低实验模式的相关性以提高协同模式识别性能的角度出发,提出了一种基于非下抽样Contourlet变换的稀疏协同模式学习方法,将Contourlet变换和协同模式识别进行结合,采用非下抽样Contourlet变换获得训练样本的低频和高颇变换系数,并根据其特点通过融合以获得模式的最优稀疏表示,消除样本冗余信息的影响,以生成最优原型模式。最后,将传统的细胞特征参数提取和协同模式分类方法相结合,提出了一种基于粗糙集约简特征维数的鼻咽癌细胞协同原型模式选择方法,在已有的细胞特征参数基础上,通过粗糙集区分矩阵约简的方法生成不同维数的约简集,然后通过协同分类方法检验生成的约简集,根据识别结果,从不同约简集中选择最适合鼻咽癌细胞协同分类的约简集以生成协同原型模式。在协同神经网络的优化方面,从序参量变换的角度,提出了一种正交逼近的协同神经重构网络模型,针对这种模型,采用了一种协同序参量的正交逼近的变换方法,利用正交多项式函数的逼近能力,将其应用到协同神经网络中以获取序参量变换参数,并在训练过程中引入快速的权值确定法以提高重构速度,加入序参量变换层以提高协同神经网路的自学习能力。从协同神经网络参数优化的角度,提出了一种基于差分进化的协同神经网络约简参数优化方法,对约简的参数模型利用差分进化的方法搜索最优协同进化参数,同时采用了均方适应度方差的机制自适应调整搜索速度和搜索精度,克服差分进化算法参数调整困难的不足,以提高差分算法的寻优能力,从而提高整个协同神经网络的性能。在协同模式识别不变性方面,采用了一种新的基于共轭梯度优化仿射参数估计的协同不变性算法,将协同神经网络中的匹配问题转化为函数优化问题,通过势能函数的进化,采用试验模式和仿射参数交替迭代的优化策略估计最优参数,通过协同神经网络中测试模式和原型模式同化等效的推论,然后由序参量进化方程得到正确的识别模式。更多还原

【Abstract】 The automatic recognition and analysis of medicine microscopic image is researched hotspots in biomedical engineering domain. Extraction process of nasopharyngeal carcinoma cell images’corporeal components and key technology of synergetic pattern recognition method are researched in this paper. In the field of extration process to cell images’formed element, intelligent extraction key technology for corporeal components in cell image was researched in this dissertation, which mainly involves cell image filtering, image segmentation, cell edge extraction, overlapped cells segmentation and inpainting technology of overlapped region. Synergetic pattern recognition key technology for nasopharyngeal carcinoma cell images was also rearched in this dissertation, which mainly involves synergetic pattern recognition learning method, synergetic neural network optimizing method and synergetic pattern recognition invariant property.Image filtering includes two ways such as filtering noise and enhancing the edge.The dissertation firstly discussed the factors of lowering quality and analyzed the noise model. The anisotropic diffusion equation self-adapting filter algorithm based on the gradient is proposed. The image background noise and the pulse noise have been seperately filtered by the ameliorative anisotropy fliter and the adaptive median filter, at the same time the detail information were reserved. Secondly, fuzzy clustering method (FCM) optimized by particle swarm optimization (PSO) and coupled with markov random field is discussed, which taking the clustering result of PSO as the initialized value of the FCM. By adopting the couple method of Markov random field and fuzzy clustering to calculate the fitness function, and apply the algorithm to the representative image segmentation to get the center of clustering. Both image guidance information and spatial information imposed by Gibbs smoothness prior to the pixel labels is used to effectively in segmenting the cell images. A multi-scale morphological edge detection algorithm based on evidence syncretic fusion is proposed. Edges of different size were detected by using different scale operator, and cell edge images were combined with the way of evidence syncretic fusion. To the overlapping cells segmentation often appears in the medical microscopic images. A segmentation algorithm for overlapped cell images based on cell scatter plot and modified Snake was proposed. In order to get the mask image based on the overlapped region, a fast image inpainting algorithm was presented based on adaptive iterative convolution to solve the information deterioration resulted from overlapping.Four improved algorithm was presented on synergetic pattern recognition form different point of view. Firstly, A synergetic prototype modify method with particle swarm optimization algorithm is applied to avoided information saturation. The method could get the optimal prototype by the global optimize ability of particle swarm optimization. Secondly, a synergetic training algorithm based on potential energy function optimized is applied form the view of meanwhile learning. The studying of potential energy function dynamics process can train prototype vector and adjoint vector meanwhile. The nonlinearing optimization approach is introduced to synergetic dynamics evolution process, using the memory gradient algorithm instead of the steepest gradient algorithm to optimize the potential energy function. Thirdly, a synergetic classification algorithm based on prototype vectors fusion with sparse decomposition is applied from the view of reduction experiment mode relativity and redundant information. It’s a trend of the recognition way research in synergetic vectors that the character value is used as prototype vectors instead of image pixel. Contourlet transform is new image representation scheme which have directionality and anisotropy. In this paper, the characteristic of contourlet transform is analyzed combined with synergetic pattern recognition. A new fusion method based on contourlet transform for prototype vectors generation is proposed. The coefficients structure and the framework’s fusion procedure are given in detail to get the prototype vectors. Lastly, a synergetic classification algorithm based on prototype modify with rough set methods is presented from the view of traditionary cell characteristic parameter combined with synergetic pattern recognition. Which is focused on prototype modify from eigenvalue. The essence of Rough set theory is a mathematic tool describing imperfection and uncertainty, can effectively analyze and deal with those imprecise. The division matrix of rough set can get the best reduce result, and furthermore dynamic rough set method is applied and optimal non-linear features are got as prototype vectors. The optimal prototype vector which fit mostly nasopharyngeal carcinoma cell images recognition could be selected from the experiment result on different reduce result by the way of synergetic pattern recognition method.The synergetic recognition of nasopharyngeal carcinoma cell images is also focused on optimize of synergetic neural network. The order parameter is the determinant of synergetic systems-ordering. The transform method of order parameter can use the neural network self-learning ability to improve recognition performance of systems. A model of order parameters transform based on orthogonal polynomial approximating was presented, which can figure out a group of linear transformation parameters for order parameters using self-learning power of synergetic neural networks. A weights-direct-determination method is proposed which could immediately determine the neural-network weights in the training process. Experiment shows that the new algorithm can effectively search the reconstruction parameters and the recognition ability of system is improved. The recognition performance of synergetic neural network could improve by adjusting parameters in the neural network system. The parameters of synergetic neural network are optimized to improve the recognition effect by sufficiently using the self-learning abilities of synergetic neural network. Differential evolution is an effective searching algorithm for global approximate optimal solution, which has the characteristics of convergence fast to better solution. An algorithm of parameters optimization based on differential evolution was proposed. This new algorithm is used to search the global optimum attention parameters of SNN in the corresponding parameter space. Fitness mean square variance is adopted to modify searching speed and searching precision in the adaptive manner, because the parameters of differential evolution algorithm are hard to adopt dynamically, and the way of fitness mean square variance could helps improving the optimizing abilities of the algorithm. The new algorithm has better parameter searching abilities, both globally and locally, and can hardly been trapped into local extreme. A reduction parameter model is applied in this algorithm which improves the recognition ratio of the synergetic neural network system effectively.Invariance method is an important aspect of Synergetic Pattern Recognition research. Usually there are deformation between test pattern and prototype pattern. A synergetic invariance algorithm is proposed in this paper, which is based on alternant iterative match. The question of match is converted to question of function optimization in synergetic neural network. A potential energy function optimization algorithm which based on conjugate gradient method is proposed, and the optimum parameters of test pattern and affine transform are gotten by the way of alternant iteration. The nationalization of test pattern is equivalent to nationalization of prototype pattern in synergetic neural network. The right pattern can be gotten by the dynamic evolvement of order parameter.更多还原