HEp-2样本图片阴阳性分类算法研究

发布时间：2018-01-03 05:41

本文关键词：HEp-2样本图片阴阳性分类算法研究　出处：《深圳大学》2017年硕士论文　论文类型：学位论文

【摘要】：人体表皮细胞(HEp-2)在医学疾病检测上有着重要作用,其样本图片的诊断一般由专业人员通过肉眼观察完成,工作强度大且容易受主观因素影响。近年来人们尝试用计算机视觉算法来进行HEp-2样本自动化判读。该过程分为两个部分,第一部分为阴阳性判断,第二部分是样本的核型判断。目前业界研究主要集中在第二部分,第一部分的工作还较少。针对该问题,本文采用了两种不同的方法来解决HEp-2样本图片阴阳性判断问题。第一种方法是结合全局特征与局部特征来对HEp-2样本图片进行阴阳性分类,首先根据样本的全局特征判别出那些较为明显的样本,剩下的样本通过一系列的预处理,图像增强,物体分割等方法确定感兴趣区域,再在这些区域上提取局部特征并对样本进行分类。第二种方法是用深度学习技术来解决该问题,本文首先以旋转、尺度变换等方法对训练数据进行扩充,再选取VGG-16和Goog LeNet这两种网络进行训练,将训练好的网络用于HEp-2样本图片分类。最后,本文还尝试将深度学习方法和SVM(支持向量机)结合起来,用卷积神经网络提取特征,SVM进行分类,发挥二者的优势共同解决问题。本文在含877张阳性样本,413张阴性样本的SZU数据库上进行测试,测试结果表明深度学习方法在整体结果上比结合全局特征和局部特征的方法更好,其总体准确率最高能达到99.87%。论文最后介绍了HEp-2样本图片阴阳性分类系统的软件开发过程和运行效果。
[Abstract]:Human epidermal cells (HEp-2) play an important role in the detection of medical diseases, and the diagnosis of the sample images is generally completed by the professionals through the naked eye observation. In recent years, people try to use computer vision algorithm to interpret HEp-2 samples automatically. The process is divided into two parts, the first part is the yin-yang judgment. The second part is the karyotype judgment of the sample. At present, the industry research is mainly focused on the second part, the first part of the work is still less. In this paper, two different methods are used to solve the problem of judging the Yin and Yang of HEp-2 samples. The first method is to classify the images of HEp-2 samples by combining global and local features. First of all, according to the global characteristics of the samples to identify the more obvious samples, the rest of the samples through a series of preprocessing, image enhancement, object segmentation and other methods to determine the region of interest. Then the local features are extracted from these regions and the samples are classified. The second method is to solve the problem by depth learning. Firstly, the training data are expanded by rotation, scale transformation and so on. Then select VGG-16 and Goog LeNet for training, and use the trained network for HEp-2 sample image classification. Finally. This paper also attempts to combine the depth learning method with SVM (support vector machine), and use convolution neural network to extract features and SVM for classification. Taking advantage of the two methods to solve the problem together. This paper was tested on the SZU database with 877 positive samples and 413 negative samples. The test results show that the depth learning method is better than the method combining global and local features in the overall results. The overall accuracy of the system can reach 99.87. Finally, the software development process and running effect of the HEp-2 sample image Yin and Yang classification system are introduced.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R446;TP391.41

【相似文献】