宫颈细胞图像特征分析与自动识别方法研究

发布时间：2018-03-17 09:18

本文选题：宫颈细胞图像　切入点：图像识别　出处：《哈尔滨理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：宫颈癌是导致女性死亡的第二大恶性肿瘤,大约占总的癌症死亡率的十分之一。据WHO统计,2015年全球有51.8万新增病例和27.3万死亡病例。2015年,我国有9.89万新增病例和3.05万死亡病例;而美国仅有1.29万新增病例和0.41万死亡病例。预计到2050年,我国将会有18.7万新增病例。因此,寻找适合于我国的宫颈癌的筛查方法已经迫在眉睫。宫颈细胞图像识别技术是近年来兴起的一种新的宫颈细胞识别的方法。该方法克服了传统的人工判读筛查方式存在成本高、工作量大、可靠性与准确性受到医师专业技术和主观情绪的影响等问题。该方法首先采用液基薄层制片技术将脱落细胞制片,然后在显微镜视野下通过工业相机抓取细胞图像进行分析识别。不仅能够用于检测异常上皮细胞,而且能够应用于宫颈癌的筛查诊断。本文采用宫颈细胞图像识别技术辅助病理学医生诊断,最终目的是识别出宫颈内异常上皮细胞,减轻医生的工作量、降低诊断结果中的假阴性和假阳性。宫颈细胞图像识别方法主要包括宫颈细胞图像获取、宫颈细胞图像预处理及分割、宫颈细胞图像特征提取和宫颈细胞图像分类四个阶段。本文对宫颈细胞图像识别方法以下几个方面进行改进:在宫颈细胞图像类别划分方面:本文结合宫颈取材特点,提出了两种划分方法。第一种类别划分方法,首先将宫颈细胞图像分为上皮细胞、淋巴细胞、中性粒细胞和垃圾细胞,然后将上皮细胞分为正常上皮细胞和异常上皮细胞,并将该宫颈细胞图像划分方法应用于SMMCFBCCI分类器。第二种类别划分方法,将宫颈细胞图像直接划分为正常上皮细胞、异常上皮细胞、中性粒细胞、淋巴细胞和垃圾细胞,并将该宫颈细胞图像类别划分方法应用于PMMCFBCCI分类器。在宫颈细胞图像特征选取方面:本文结合了前人研究和宫颈细胞学特点,提出了NF和PF特征集合。其中,NF是根据宫颈细胞病理特征定义的常规特征集合,共22维;PF是指传统方法中没有考虑到的潜在特征集合,共14维。然后,采用Relief F算法选择出类别相关性高的24维特征。在宫颈细胞图像分类器设计方面:结合多分类器融合方法,本文提出了SMMCFBCCI分类器和PMMCFBCCI分类器。SMMCFBCCI分类器是基于两级级联的多分类器融合方法。其中,第一级粗分类采用C4.5分类器;第二级细分类采用LR分类器。PMMCFBCCI分类器是基于多数投票法的串行分类器融合。首先,采用NB、C4.5及KNN分类器得到预测结果;然后,采用多数投票法得到最终预测结果。
[Abstract]:Cervical cancer is the second largest malignant tumor in women, accounting for 1/10% of the total cancer mortality. According to WHO statistics, in 2015, there were 518,000 new cases and 273,000 deaths in the world. On 2015, there were 98,900 new cases and 30,500 deaths in China. In the United States, there are only 12,900 new cases and 4,100 deaths. By 2050, there are expected to be 187,000 new cases in China. It is urgent to find a suitable screening method for cervical cancer in China. The technology of cervical cell image recognition is a new method of cervical cell recognition developed in recent years. This method overcomes the high cost of traditional manual screening. The workload, reliability and accuracy are influenced by the professional technique and subjective emotion of the physician. Firstly, the technique of thin-film preparation based on liquid is used to prepare the shedding cells. Then under the microscope field of vision, the cells are captured by industrial cameras for analysis and recognition. Not only can they be used to detect abnormal epithelial cells, And it can be used in the screening and diagnosis of cervical cancer. In this paper, cervical cell image recognition technology is used to assist pathologist in diagnosis, and the ultimate goal is to identify abnormal epithelial cells in the cervix and reduce the workload of doctors. The methods of cervical cell image recognition mainly include cervical cell image acquisition, cervical cell image preprocessing and segmentation. There are four stages of cervical cell image feature extraction and cervical cell image classification. In this paper, the methods of cervical cell image recognition are improved in the following aspects: classification of cervical cell image: this paper combines the characteristics of cervical sampling, Two methods of classification are proposed. The first is to divide cervical cell images into epithelial cells, lymphocytes, neutrophils and garbage cells, then to classify epithelial cells into normal epithelial cells and abnormal epithelial cells. The cervical cell image was divided into normal epithelial cells, abnormal epithelial cells, neutrophils, lymphocytes and garbage cells. The method of classification of cervical cell image is applied to PMMCFBCCI classifier. In the aspect of feature selection of cervical cell image, this paper combines the previous studies and the characteristics of cervical cytology. In this paper, NF and PF feature sets are proposed, in which NF is a conventional feature set defined according to the pathological features of cervical cells, and 22 dimensional PF refers to the potential feature set which is not considered in the traditional method, which has a total of 14 dimensions. The Relief F algorithm is used to select 24 dimensional features with high class correlation. In the design of cervical cell image classifier, the fusion method of multiple classifiers is combined. In this paper, SMMCFBCCI classifier and PMMCFBCCI classifier SMMCFBCCI are proposed, which are multi-classifier fusion method based on two-stage cascade, in which C4.5 classifier is used in the first stage coarse classification. In the second stage, LR classifier. PMMCFBCCI classifier is a serial classifier fusion based on majority voting method. First, NBC4.5 and KNN classifier are used to obtain the prediction results, and then the final prediction results are obtained by majority voting method.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R737.33;TP391.41

【相似文献】