基于GVF Snake的宫颈细胞图像分割算法及分类识别的研究

发布时间：2018-03-26 07:32

本文选题：宫颈细胞图像　切入点：图像分割　出处：《广西师范大学》2016年硕士论文

【摘要】：本文以宫颈细胞图像为例,对细胞图像的分割、形态特征和极坐标特征、识别应用的技术进入了深入的研究,主要是包括宫颈细胞图像的细胞质、细胞核与背景的轮廓精确提取,宫颈细胞图像多特征融合及细胞分类识别方法。主要研究从以下几个部分：(1).提出了一种基于自适应阈值和射线梯度的GVF Snake主动轮廓模型,用来定位宫颈单细胞图像的细胞核与细胞质的边缘。GVF Snake主动轮廓模型应用比较广泛的主动轮廓模型的目标边缘跟踪算法,但是宫颈细胞图像,特别是细胞质边缘相对模糊、细胞核与细胞质的边缘相互吸附难以分开,还有干扰性的血细胞及炎症细胞、染色度分布不均匀,都是导致GVF Snake模型细胞边缘吸附到错误的位置。为了解决以上难题,本文研究了以下方法：首先是利用自适应阈值去除细胞背景,然后使用射线梯度方向的信息计算细胞灰度值,最后根据射线上的灰度值使用GVF Snake模型演化,在演化过程中使用栈的灰度差补偿算法,结合正向灰度差抑制能够很好的克服噪声、血细胞及炎症细胞等虚假边缘的信息影响。本文使用的Herlev数据库验证了该方法的有效性和可行性。(2).在对宫颈细胞图像进行精确分割的基础上,研究了宫颈细胞图像的形态特征参数,主要包括9种几何特征和4种纹理特征。9种几何特征分别为：细胞质的周长、细胞核的周长、竖直方向的最长轴、水平方向的最宽轴、细胞核与细胞质的比率、轴中心到周长的最长长度、轴中心到周长的平均长度、重心到周长的最长长度、重心到周长的平均长度；4种纹理特征：共生矩阵的熵、共生矩阵的对比度、对比度和粗糙度。宫颈单细胞图像是由细胞核、细胞质和背景三个区域都可以转化到极坐标系下,提取极坐标下的极经灰度值,360条极经的灰度值组成一个特征矩阵。本文将极坐标下的特征向量与前面的形态特征进行融合,来研究宫颈细胞的识别。(3).使用基于AdaBoost与SVM算法结合的向量机,改善了分类器的稳定性和差异性。采用的AdaBoost-SVM分类器将提取的宫颈细胞多特征进行融合,再识别分类应用。双重分类器结合可以弥补单个分类器的缺点,提升分类识别效率。通过特征提取方法与AdaBoost-SVM多特征融合分类器结合,实验结果证明：提高了宫颈细胞涂片筛查的效率和准确率,降低了宫颈癌的误诊率。本文对宫颈细胞图像的分割、特征提取和宫颈细胞的分类识别等进行了系统性的研究和改进。实验结果表明：本文的方法能较好的完成宫颈细胞的定量分析,对于宫颈细胞图像的自动筛查分析系统具有较好的应用价值。
[Abstract]:In this paper, we take cervical cell image as an example, the segmentation of cell image, morphological features and polar coordinate features, recognition of the application of technology into in-depth research, mainly including the cervical cell image of the cytoplasm, The precise contour extraction of nucleus and background, the multi-feature fusion of cervical cell image and the method of cell classification recognition are studied. A GVF Snake active contour model based on adaptive threshold and ray gradient is proposed from the following parts: 1. GVF Snake active contour model used to locate the edge of nucleus and cytoplasm of single cell image of cervix is widely used in target edge tracking algorithm of active contour model, but the image of cervical cell, especially the edge of cytoplasm, is relatively fuzzy. It is difficult to separate the nuclear and cytoplasmic edges from each other, as well as interfering blood cells and inflammatory cells. The uneven distribution of staining results in the GVF Snake model cells being adsorbed to the wrong position on the edges. In this paper, the following methods are studied: firstly, the adaptive threshold is used to remove the background of the cell, then the information of the direction of the ray gradient is used to calculate the gray value of the cell. Finally, according to the gray value on the ray, the GVF Snake model is used to evolve. In the evolution process, using the stack gray difference compensation algorithm, combining with the forward gray difference suppression, can overcome the noise very well. The effect of false edge information such as blood cells and inflammatory cells. The Herlev database used in this paper verifies the effectiveness and feasibility of this method. On the basis of accurate segmentation of cervical cell images, The morphological parameters of cervical cell images were studied, including 9 geometric features and 4 texture features, including the circumference of the cytoplasm, the circumference of the nucleus, the longest axis in the vertical direction, and the widest axis in the horizontal direction, respectively. The ratio of nucleus to cytoplasm, the longest length from the axis center to the circumference, the average length from the axis center to the perimeter, the longest length from the center of gravity to the perimeter, and the average length from the center of gravity to the circumference are four texture features: the entropy of the symbiotic matrix, The contrast, contrast and roughness of the symbiotic matrix. The single cell image of the cervix can be transformed from the nucleus, the cytoplasm, and the background into polar coordinates. The gray values of 360 poles in polar coordinates are extracted to form a feature matrix. In this paper, the feature vectors in polar coordinates are fused with the former morphological features. Using vector machine based on the combination of AdaBoost and SVM algorithm to improve the stability and difference of the classifier. The AdaBoost-SVM classifier is used to fuse the extracted multiple features of cervical cells. The combination of double classifiers can make up for the shortcomings of single classifier and improve the efficiency of classification recognition. The method of feature extraction is combined with AdaBoost-SVM multi-feature fusion classifier. The results show that the efficiency and accuracy of cervical smear screening are improved, and the misdiagnosis rate of cervical cancer is reduced. Systematic research and improvement of feature extraction and classification and recognition of cervical cells have been carried out. The experimental results show that the method presented in this paper can accomplish the quantitative analysis of cervical cells. It has good application value for automatic screening and analyzing system of cervical cell image.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R737.33;TP391.41

【相似文献】