【作者】 范金坪；

【导师】 张永林；

【作者基本信息】 暨南大学 ， 生物医学信息技术， 2010， 博士

【摘要】 利用计算机技术和宫颈细胞病理学诊断技术对宫颈细胞图像进行定量分析和自动识别在宫颈癌和癌前病变的筛查及诊断中具有重要的实用价值和应用前景。由于宫颈细胞涂片制片和染色方式的差异性、背景的复杂性、细胞形态的多样性和不规则性、细胞重叠等使得宫颈细胞图像的计算机处理及识别难度较大。国内对于宫颈细胞图像的自动识别研究开展的较少,亟待突破和提高。本文在前人研究的基础上,结合宫颈细胞病理学基础知识,运用图像分析技术和模式识别技术,系统的研究了宫颈细胞图像的分割方法、特征参数的计算和选择以及宫颈细胞分类识别技术。内容主要涉及以下几个方面：利用宫颈细胞结构的特点,结合基于区域内一致性及区域间差异性的C-V模型,提出构建两个独立的水平集函数来逼近细胞组成区域的边界。根据细胞核、细胞质及背景区域间的灰度和面积差异性对水平集函数的演化方程进行定义及改进,将一个三目标分割的问题转化为两个两目标分割的问题。由于ROI区域中可能存在多个连通细胞区域,提出一种提取主细胞连通区域的分割方法。对各类宫颈细胞图像的分割实验表明,通过调整两个独立水平集的加权系数就能够实现任意弱边界细胞图像组成区域的分割。在成功实现灰度宫颈细胞图像分割的基础上,结合基于矢量量化的C-V模型进行彩色宫颈细胞图像的分割。针对宫颈细胞图像中存在的细胞粘连及重叠的问题,结合极限腐蚀和凹区检测的方法对重叠细胞进行重叠判断和分离。已经将该方法成功的用于宫颈重叠细胞(核)图像的分离。在对宫颈细胞组成区域进行精确分割的基础上,提取出能对宫颈细胞进行分类识别的特征参数,主要包括形态、色度、光密度、纹理等共87个宫颈细胞特征参数。全面考虑特征选择的要求,利用遗传算法从原始特征集中选择特征,以特征的可靠性和可区分性原则定义适应度函数,自适应生成变异概率,采用两点交叉和最优保存策略来实现遗传算法。考虑到遗传算法初始种群生成的随机性,提出一个特征选择的评价准则。提取出满足特征评价准则的特征组合得到特征子集用于宫颈细胞图像的分类识别。使用具有一个单隐层的BP神经网络对遗传算法的特征选择结果进行验证,结果表明使用遗传算法选择出来的特征子集的识别率远高于原始特征集的识别率,证明使用遗传算法进行特征选择是有效的。鉴于单个神经网络的泛化性能不高,采用神经网络集成进行宫颈细胞图像的分类识别,使用Bagging算法生成个体神经网络。为了降低将癌变细胞识别为非癌变细胞的误识率及总体误识率,使用级联的方式将两个神经网络集成组合起来,组成一个两级神经网络集成。结果表明使用两级神经网络集成进行宫颈细胞图像的分类识别,不仅使总体误识率大大降低,而且最重要的是能降低将癌变细胞判别为正常细胞的误识率。本文对宫颈细胞图像的分割、特征提取和选择以及宫颈细胞识别等算法进行了系统的研究和改进。实验结果表明,本文提出的方法能够较好的完成宫颈细胞图像的定量分析和识别任务,为宫颈细胞图像自动分析系统的开发及实现奠定了理论基础。更多还原

【Abstract】 The quantitative analysis and automatic recognition of cervical cell image utilizing computer technology and cervical cytological diagnosis have significant practical value and application prospect on the screening and diagnosis of cervical precancerous lesion and cervical cancer. Due to the cervical smears slice-making and staining technique differences, the background complexity, the cell form diversity and irregularity and the cell overlap making it difficult to process and recognize the cervical cell. image. The research on the automatic cervical cell image recognition has developed not so many and urgent need breakthrough and improvement.Based on the previous research, we make a systematical study on the techniques of cervical cell image segmentation, characteristic parameters computation and selection and cell image recognition by applying the depth research on image analysis, pattern recognition technique and cytological pathology knowledge. The main contents of this thesis can be summarized as follows:According to the structure of the cell, two independent level set functions have been constructed to approach the borders of the cell composition based on the Chan-Vese model with intra-region coherence and inter-region diversity properties. Through defining and improving the evolution equation of level set function on the basis of gray and area differences among nuclei, cytoplasm and background region, we can convert an three-region segmentation problem to two two-region segmentation problems. While there may be more than one connected cell regions in region of interesting (ROI), a method of main connected cell region extraction has been introduced. The segmentation experiment on each types of cervical cell image shows that by means of adjustment of weight values of two level set functions, any cell image with weak edges can be segmented precisely. Based on the success gray cell image segmentation, the color cell image segmentation has been proposed combined with the vector-valued Chan-Vese model. Focus on the problem that there exist cell adhesion and overlap in the image, a method combined with corrosion limit and concave area detection has been introduced to judge and separate the overlap cell image. The method has been operated on the overlapped cervical cell or nuclei images separation successfully. On the basis of precise segmentation of cell compositions, characteristic parameters that can be used in cell recognition are extracted, including morphology, color, optical density and texture features and 87 features are extracted in all. An approach is proposed to perform feature selection based on genetic algorithm. Define the fitness function on the principle of high reliability and distinction of feature subsets, generate the mutation probability adaptively, use two-point crossover and maintain optimal strategy, the genetic algorithm can be carried out to select the optimal features. In view of the randomness of original pop set generation, en evaluation criteria has been introduced to extract the qualified features to make up the optimal feature sets which can be used on the cervical cell image recognition.BP neural network with one hidden layer has been used to testify the efficiency of feature selection with genetic algorithm, the result shows that the recognition rate using the features selected by genetic algorithm is higher than the original features and it is effective to select the features by applying genetic algorithm. Considering that the generalization performance of single neural network is not high enough, we propose using neural network ensembles to recognize the cell image and generate the individual network by bagging algorithm. In order to decrease the error recognition rate of identifying malignant cells to normal ones and the overall error recognition rate, we quote a two-layer ensemble by cascading two neural network ensembles together and to finish the recognition task. The experimental results show that the overall error recognition rate of the two-layer neural network ensembles has been decreased significantly, the more important is that the error recognition of diagnosing malignant cells to normal ones has been decreased greatly.This thesis has making systematic research and improvement on the cervical cell image segmentation, characteristic parameters computation and selection and cell image recognition. The experiment results show that the methods we proposed in this thesis are effective to achieve the mission of quantitative analysis and automatic recognition of cervical cell image, and has establish a foundation on cervical cell image automatic analysis system.更多还原