【作者】 张赞超；

【导师】 夏顺仁；

【作者基本信息】 浙江大学 ， 生物医学工程， 2008， 博士

【摘要】 尿沉渣检查是临床检查中心检验操作规程中的一项基本要求,根据尿液有形成分的形状、纹理等特征来识别尿液中的红细胞、白细胞、管型、上皮细胞、结晶等各种生理或病理成分,从而为医务人员对泌尿系统疾病做出正确的诊断提供帮助。常规体格检查或化学实验中不能发现的异常变化,常可通过尿沉渣有形成分的定性以及定量分析而确定。由此可见,尿沉渣的镜检具有非常重要的临床意义。为了提高结果的客观性,减少人为因素对结果的影响,提高尿沉渣检查的标准化,降低工作人员的劳动强度,研制全自动的尿沉渣分析仪具有重要的意义。论文研究了尿液有形成分的检测方法,利用数字图像处理技术和模式识别对尿沉渣图像的分割和有形成分的分类进行了深入研究。从图像特征出发,将多种处理算法有机的结合,通过计算和反复实验,提出了一套完整有效的对尿液有形成分进行分析和分类的处理方案。论文的主要任务是在全自动尿沉渣分析仪硬件系统的基础上,设计并实现其控制操作和管理软件,设计并实现对尿沉渣图像的自动识别算法。为了实现对尿液有形成分的准确分析,关键是对图像进行有效的分割。图像的质量是图像分割的基础,论文提出了基于主分量分析的自动聚焦算法,有效提高了获取的图像的质量。另外,分割前对图像的预处理可以有效降低分割的难度,提高分割的准确性。论文提出了基于局部均值拉伸的图像增强和光照不均匀消除及阴影抑制算法;然后采用细胞神经网络对图像进行分割,并结合形态学操作和基于对测点的区域二次分割方法,得到最终的有形成分分割结果;再提取有形成分区域的多项形状、统计和纹理参数,对各个区域进行描述;最后利用多BP神经网络组合得到的分级分类器对区域进行分类,从而得到识别的最终结果。论文工作主要有六个部分:第一部分对系统的硬件、软件系统和控制流程进行介绍。第二部分为对尿沉渣图像的预处理算法。论文提出了基于局部均值拉伸的图像增强算法,通过拉伸图像中各个点与其某邻域内灰度均值的差异达到增强图像对比度的目的;同时可以通过对邻域内均值的变换,消除图像中光照的不均匀和阴影。第三部分为系统中的自动聚焦算法。论文提出了基于主分量分析(Principal component analysis,PCA)的自动聚焦算法,算法通过对同一视野的对幅图像分别计算多个清晰度评价指标,然后通过PCA算法对这些指标进行融合,得到最终的清晰度评价。第四部分为尿沉渣图像的分割算法,论文利用细胞神经网络,设计合适的模板对图像进行分割;然后结合形态学操作,实现对图像的初步分割;最后论文提出了基于对侧点的区域二次分割算法,通过对区域边缘形状的分析,确定二次分割的位置,从而实现分离粘连细胞、消除区域毛刺的目的。第五部分为图像区域特征的提取算法,通过对图像特征及其分布曲线的分析,确定若干形状参数、统计参数和纹理参数作为对区域的描述,为分类作好准备。第六部分为有形成分的分类,介绍了BP神经网络和支持向量机的基础知识,通过对测试结果的初步分析,提出了分级的分类器组合方法,然后在每级中采用多个BP网络进行投票以提高分类的准确度,从而最终实现全自动的尿沉渣检查。更多还原

【Abstract】 The examination of urinary sediment,which is a basic requirement of the operation regulations in the Center for Clinical Laboratories,can be used to distinguish the red blood cell[RBC],white blood cell[WBC],cast,epithelia, crystal,and various physiological or pathological particles in the urine according to morphological,textural,and other features of urinary sediment, and to help the doctors to diagnose diseases of urinary systems accurately. The abnormal changes of urinary systems may not be detected by conventional medical examination or chemical experiments,however,these changes can be distinguished by analyzing the urinary sedimentary components qualitatively and quantitatively.This shows that the examination of urinary sediment has an important clinical application.For the sake of improving the objectivity of the results,enhancing the standardization of the examination,and reducing the labor intensity of staff, it is of great significance to develop a fully automated urine analyzer.The detection methods of urinary sediment visible components are discussed and analyzed in the dissertation.The segmentation and classification of the urinary sediment imaging are discussed deeply by using digital image processing and pattern recognition methods.In terms of the characteristics of the urinary sediment images,a set of complete effective processing scheme for urinary sediment visible components is proposed by combing multi-processing arithmetic with amount of computing and trials.The major tasks of this dissertation are to design and implement the controlling operation and managing software of the urine analyzer,and to perform the algorithms for processing the urinary image and classifying the urinary visible components.In order to accurately analyze urinary visible component,the effective imaging segmentation play a key role.The quality of the images is the foundation of the imaging segmentation.In this dissertation, auto-focus method based on Principle Component Analysis[PCA]is firstly proposed,which can improve the images quality effectively.On the other hand,image preprocessing can decrease the difficulty and increase the accuracy of image segmentation.This dissertation proposes a preprocess algorithm which enhances the edges of objects by stretching the difference between every pixel and the local gray mean value and eliminates the disequilibrium of illumination by nonlinear transform of the local gray mean value.And then,combined with morphological operations and post-process of regions,the cellular neural network[CNN]is employed to get the final segmentation results.By analyzing the initial test results,a layered classifier with multi-BP networks combined is proposed to classify the objects which are described by certain morphological,statistical and textural features.This dissertation is composed by the following six parts:PartⅠ:An introduction to hardware system,software system and control flow of the analyzer.PartⅡ:The pre-processing algorithm to urinary sediment images.An image-enhancement algorithm is proposed,which can be used to enhance the contrast of images by stretching the difference between each point’s gray value and the local gray mean value.Furthermore,the uneven illumination and shadow in the image can be eliminated by taking certain nonlinear transform to the gray mean values of the neighborhoods.PartⅢ:The auto-focus algorithm based on Principle Component Analysis is firstly proposed.Several indices of definition of every image in the same visual field are calculated in advance,and then are integrated by PCA algorithm to evaluate the definition of each image.PartⅣ:Methods of urinary images segmentation.By designing suitable CNN templates and combining with morphological operations,CNN is employed to complete the initial segmentation.And then,post-process linking the opposite points which are conformed by analyzing the shape of the edges is carried out to separate conglutinate cells and eliminate burrs of them.PartⅤ:The algorithm to extract the features of the regions.By analyzing the image features and their corresponding distribution curves,certain morphological,statistical and textural features are calculated to describe the regions,and therefore prepare for further classification. PartⅥ:The classification of the urinary sediment components.The basic knowledge of BP Neural Network and Support Vector Machines(SVM)are introduced at first.By analyzing the initial test results,a layered classifiers combination method is proposed.Finally,several BP networks are combined by voting in each layer to improve the accuracy.By combining these methods presented previously,the fully automatic urinary sediment examination is completely implemented.更多还原