真菌隐球酵母菌(Cryptococcus neoformans )属于担子菌类族(和菌类相关)，能引起脑膜炎和肺部感染等。它能引起主要发生于免疫系统缺陷的人感染。例如：艾滋病患者，器官移植者和白血病患者。真菌隐球酵母菌，最主要的病毒因素是在细胞周围出现了一个聚多糖的包裹的外壳(荚膜)。生物学家已经验证不同的免疫器官疾病的导致的真菌隐球酵母菌的荚膜的厚度有变化。因此量性地判定荚膜形态，可以量性地判定酵母细胞对人体的病理程度，同时能够快速可靠的测量荚膜的大小形状对于医学和生物学研究都是非常有用和必须的。 随着数字图像处理的方法发展，尤其是显微镜数字图像的广泛采用，可以利用生物数字图像处理技术量性地测量细荚膜的大小和形状。并且通过分析真菌隐球酵母菌荚膜的形态学的变化，来分析它的人体的病理状态。当真菌隐球酵母菌的外面出现厚的细荚膜，可以认为它呈现一种病理状态。这就需要生物学家和计算机专家研究出有效的方法和工具，能够正确良好的分割这种显微镜图像，并且能够自动从分割结果中测试出有效的参数，表示不同的状态。 目前研究存在的主要问题是： ·合理有效图像特征描述：这其间的核心问题是到底什么参数能够表示病理状态？或者更确切地能够用这些参数，能表示不同程度的病理状态吗？具体涉及到参数是图像的纹理、颜色、位置、形状等等，同时确定哪些参数更能够有针对性地解决问题 ·图像的准备：另外一个比较棘手的问题就是怎样准备比较好的、适用的图像。由于酵母细胞的生物结构的特殊性，对其进行图像准备也很大的难度。同时在各种类型的细胞图像中，哪种图像更易于获取及方便后面的处理？彩色图像、灰度值图像、还有通过负染色获取的荧光图像、负染色及正染色图像等等。目前采用的负染色的Indian ink图像存在着可复制性差的问题。同时通过这种方法准备的图像，还存在着目标对象边缘比不高，图像的噪声大、分辨率不高、细胞的荚膜和细胞核的形态差异太小等问题。并且细胞的亮度也有的不高。 ·图像的处理：对于已经获取的图像中，还存在大量的粘在一起的细胞簇。对于胞图像，有的是由于正在“出芽生殖”，有的是酵母细胞的游动偶然碰到一起而引起的。以前的工作对上述的问题既没有做出判断，并且也没有处理这些细胞簇图像。另外尽管需要一个快速、自动的图像处理模型，但是由于酵母细胞图像呈现的特殊性，所以采用传统的图像处理算法，来完成前面获取的图像“低层次”、“中层次”的处理效果并不好。 本文将主要围绕上述的问题，采用数据挖掘和图像处理的高层次处理相结合的方法，将生物图像作为数据挖掘的对象，同时利用数据挖掘中的知识学习的各种方法，进行图像的高层次处理，自动抽取有意义的语义信息。也就是说本文紧紧围绕图像挖掘、图像处理技术开展研究工作。 本文的主要研究成果与创新点如下： ①建立了一个基于功能的真菌隐球酵母菌的图像挖掘系统。提出采用神络网络、遗传算法、决策树、决策表等多种机器学习的方法，对真菌隐球酵母菌进行数据挖掘，建立初步预测模型，能够实现有监督和无监督的图像分类。在比较了不同分类算法的运算结果的基础上，取得较优的预测模型。进而通过较优的预测模型，对酵母细胞的病理进行判断。 ②不变矩被提出采用的作为形状描述特征，通过数据驱动的参数选取方法，确定对后的参数选取方案。并且分析和验证为什么这些参数会更有效。提出小波编码方法对参数进行编码压缩。数据驱动的参数选取方案，将原来的参数由77×135，减少到77×10个。小波编码方法能够将参数由77×135减少到77×36个。 ③提出噪声放大为前景目标的混合方法对图像进行预处理，提取细胞的二值和灰度值模板。 ④提出一个改进快速的分水岭算法，对出芽生殖的图像进行抛弃，对因游动而粘在一起的图像进行有效分割。提出了基于局域的模糊模板驱动的图像分割算法，能够对高噪声背景细胞图像进行较佳分割。 本论文的部分结果已经被荷兰皇家科学院，利用于真菌隐球酵母细胞的病理分析。 关键词：图像挖掘，矩特征，局域模糊模板驱动的图像分割法，数据驱动参数选取

 Cryptococcus neoformans belongs to the group of basidiomycetes and can cause meningitis. It can cause infection of the people who has the immune deficiencies, i.e. HIV, organ transplant patients and leukemia patients. The most significant factor of virulence is the presence of polysaccharides around the yeast cell. A measurement of the quantity of the capsule is very important to measure the severity of the pathogen of yeast cell. Thus it is very important to measure the morphology quickly reliable is very useful and necessary. With the development of image processing, especially the adoption of microscope image, pathogen can be analyzed through analyzing the morphology of the capsule using the image processing techniques. So it is very necessary for computer scientist and biologists to segment the image accurately and calculate the attributes to represent different conditions. This paper will utilize the combination of data mining and high level image processing using biology image as the input object for mining. Different Dataa Mining methods have been used to build the prediction model such as Neural Network, Genetic Evolution Algorithm, Decision Tree and Decision Table, etc. From data mining, high level image processing will be used to mine out the semantic information. The result of the dissertation has been used for Fungal Diversity of Royal Academic and Science of The Netherlands. The problems now exist: ·A reasonable quick image mining system: A quick and stable image mining system is the key to problem. It can complement high level image processing effectively to mine out useful knowledge. ·Reasonable and effective image descriptor and selection: The center of the problem is what can represent the pathogen condition? Or accurately to say, can we use these features to represent different pathogenic conditions, involving image texture, color, position and shape, etc. Which features have more powerful predictable ability? ·Image preparation: Another trouble is how to prepare a more suitable image. Indian ink images lack the reproductive ability and at the same time, the images we got via this way, have low sharpen, low contrast, great noise and low resolution. ·Image processing: For the prepared images, there are clumped cell images. For these images, there are no judgments whether the forming reasons are because of budding condition or floating problems. Previously no clumped cells have been included. Although we need to build up an automated model, but the processing result of tradition method for low/mid level is not good enough. According to the problems above, image mining and image processing are focused. The contributions of this paper are: ·Image mining: Multi methods of machine learning such as Neural Network, Genetic Algorithm, Decision Table and Decision Tree, will be utilized to do the supervised and unsupervised classification. After the comparison of the results, the better result will be selected to build the prediction models, and further on to diagnose the pathogenic conditions. ·Image descriptors and selections: Features such as Moments and so on, are utilized to represent the morphologies, and at the same time through data driven attributes selection, we can make sure the final features set. At last we prove the result on theory of biologist. ·Image processing: A quick and relative watershed function has been proposed by us. Budding yeast cells were deleted. Floating yeast cells were separated. Local regional fuzzy mask based segmentation methods were proposed by us to process the image. ·Image preparation: Indian ink images are used for supervised classification model and Nigrosine images are utilized for unsupervised model. Keywords: image mining, moments, local regional fuzzy mask, data driven