【作者】 杨雪琴；

【导师】 房斌；

【作者基本信息】 重庆大学 ， 计算机应用技术， 2010， 硕士

【摘要】 尿沉渣有形成分的自动识别对临床尿检具有重要意义。人工镜检方式的传统尿沉渣检查法不但劳动强度大、受主观因素影响,而且主要集中于有形成分的定性检查,不利于快速、准确的定量分析。随着数字图像处理和模式识别技术的发展,计算机辅助尿沉渣显微图像的分析成为可能。尿沉渣自动分析仪的研发除了能极大的提高临床检验的效率、降低检验医生的劳动强度外,还有利于医院的信息化和疾病诊断判别的标准化,也给医疗资源共享、远程会诊提供了便利。市场上已经出现了多种尿沉渣自动分析仪,尿沉渣图像的自动处理通常可分为有形成分的分割和识别与计数。对识别而言,一般从形态和纹理两方面提取特征并采用分类器进行分类。但目前对尿沉渣有形成分识别的研究中,通常只提取面积、周长、圆度等简单的形状特征,而在纹理特征方面几乎都是在空间域进行纹理特征的提取。本文在前人研究的基础上,围绕尿沉渣有形成分的识别,针对各种有形成分特有的形态特征,提出一些新的形态表示法。而在纹理特征的研究中,利用小波域高频系数在纹理特征表示中的优势,提出基于小波域统计纹理特征的纹理识别法。在图像分割方面:介绍了图像分割的常用方法,以及一种基于灰度差分的双阈值尿沉渣图像分割法和基于分水岭算法的粘连细胞分割法,用分割结果的二值图膨胀后的边界作为初始轮廓,采用Snake模型提取细胞封闭轮廓。在形态特征研究与应用方面:介绍了常用的形态特征描述法,以及中轴提取算法。改进基于距离变换的中轴提取算法,以适应管型形态描述的需要,提取管型单像素宽、连通且无分枝的中轴,基于中轴提出一种描述弯曲管型形态的方法,采用决策树分类器,结合其它形态描述法提出一种管型形态识别方法。研究Hough变换的理论及实现,将基于Hough变换的直线检测法用于尿沉渣中结晶的识别,而将基于Hough变换的圆检测法用于白细胞团与上皮细胞的区分以及白细胞团的分割与计数。将基于Hessian矩阵的血管增强算法应用于精子图像的增强,结合Otsu和区域生长等算法,提出一种定位头部、追踪尾部的精子识别方法。在纹理特征研究与应用方面:介绍了统计纹理特征提取的常用方法,包括矩特征、空间自相关函数、灰度共生矩阵等;借助基于小波变换域的多尺度纹理图像分割思想,提出一种基于小波域统计纹理特征的纹理分类方法,并应用于尿沉渣图像中有形成分的纹理识别中。更多还原

【Abstract】 The automatic recognition of urinary visible components is of great significance in clinical examination of urinary sediment. The traditional manual microscopic examination method is not only labor intensive, sensitive to subjective factors, but also make against with rapid and accurate quantitative diagnosis as it is mainly centralize on the qualitative examination of visible components. With the development of digital image processing and pattern recognition technology, the computer-aided analysis of urinary sediment microscopic images has become possible. The invention of automatic urinary sediment analyzer can not only greatly improve the efficiency of clinical examination, reduce the labor intensity of physicians, but also provide help on hospital’s informationization, standardization of disease diagnosis, and facilitate the sharing of health care resources and remote consultation.Several automatic urinary sediment analyzers have emerged in markets. The automatic process of urinary sediment images is commonly divided into segmentation, recognition and counting. For recognition part, the commonly used method is extracting shape and texture features and use classifier for classification. At present, the shape features used for recognition are usually some simple features, such as area, perimeter and circular degree and so on, and the used texture features are usually extracted in spatial domain. Based on the previous research, some new shape description methods are proposed according to the special shape features of some kinds of visible components in urinary sediment microscopic images. And for the research of texture features, as the high frequency wavelet coefficients have special advantages for texture feature representation, a texture recognition method based on the wavelet domain statistical texture features is proposed.For image segmentation: the common used image segmentation methods are introduced, and a gray variance based bi-thresholding urinary sediment image segmentation method and a watershed based overlapped cells’segmentation method are introduced. Using the snake model for cells’boundary location, and using the edge of binary image after dilation as it’s initial contour.For shape features and their applications: the common used shape description methods and centerline extraction methods are introduced. The distance transform based centerline extraction method is improved for the extraction of single pixel wide, connected, and no branch centerline of casts, and based on the extracted centerline, a tube-like shape description method for casts recognition is proposed. Combined with other methods and using decision tree as classifier, a shape recognition method of casts is proposed. The theory of Hough transform and its implementation are studied, using the Hough transform based line detection method for the recognition of crystal, and using the Hough transform based circle finding method for distinguishing between white blood cell clusters and epithelial cells, and for the segmentation and counting of white blood cells. The Hessian matrix based vessel enhancement method is applied to the enhancement of sperm image, combined with Otsu binarization and region growing method, a head locate and tail tracing method for sperm recognition is proposed.For texture features and their applications: The common used statistical texture feature extraction methods are introduced, including moment characteristics, spatial autocorrelation function, GLCM, and so on. A wavelet domain statistical texture feature based texture classification method is proposed and applied to the texture recognition of visible components in urinary sediment images.更多还原