基于深度学习的白细胞分类计数的研究

发布时间：2018-03-31 01:22

本文选题：白细胞　切入点：分类　出处：《深圳大学》2017年硕士论文

【摘要】：在临床上,白细胞分类识别是血液检验的一项重要内容,准确、快速的对白细胞进行分类是医疗领域一项重要的研究。目前,临床上对白细胞的检验的方法是血细胞分析仪和人工镜检,即先用血细胞分析仪对样本进行筛查,如果发现异常样本,则进一步用显微镜肉眼观察,确定最终结果。人工镜检是白细胞分类的金标准,准确度能够达到95%以上。但是人工镜检效率低,分类速度慢,准确度受检验人员经验和状态的影响。近几年来,深度学习在图像识别领域取得重大突破,利用这种方法对白细胞图像识别成为一种新的研究方向。本文利用深度学习的方法设计了一种白细胞自动分类系统,整个系统包括从血涂片制作到图像采集分割以及最后的识别分类整个过程。本文完成的主要工作可以概括为以下几点:1)利用显微镜从血涂片中拍摄大量的血细胞显微图像,经过图像分割算法得到大量的单个白细胞图像。利用分割得到的白细胞图像,建立新的白细胞数据库,该数据库总共包含四个数据集,Train、Train*、Test和Test*。每一个数据集都包含五种类型的白细胞图像,中性粒细胞、嗜酸性粒细胞、嗜碱性粒细胞、淋巴细胞和单核细胞。2)设计了一种新的深度学习网络结构,包括两层卷积层、两层下采样层、一层全连接层。我们利用Train数据集去训练该网络,并利用Test数据集去验证网络模型的性能,白细胞图像平均识别率为98.58%。3)通过调整网络参数和样本数量优化系统结构。优化后的网络包含五层卷积层、五层下采样层和一个全连接层,每个卷积层和下采样层都包含39个特征映射。优化之后,白细胞图像平均识别准确率为99.27%。4)利用交叉验证的方法评估网络模型的性能。我们将Train*和Test*数据集中的所有白细胞图像分成10等份,9份用作网络训练,1份用作网络验证。按照这种方法,我们得到10个不同的白细胞数据库。我们利用白细胞数据库得到10组白细胞识别准确度,我们取其平均值作为系统性能的标准。本文提出了一种基于深度学习的白细胞自动分类方法,并与传统的白细胞图像自动识别方法进行了对比。本文提出的方法对中性粒细胞、嗜酸性粒细胞、嗜碱性粒细胞、淋巴细胞和单核细胞等五种白细胞的识别率分别为99.70%、99.63%、99.78%、99.49%、99.62%,平均识别率为99.64%。
[Abstract]:In clinic, leukocyte classification and recognition is an important content of blood test. Accurate and fast classification of white blood cells is an important research in medical field. At present, The methods of clinical examination of white blood cells are hematology analyzer and artificial microscope examination. That is to say, the samples are screened by blood cell analyzer first, and if abnormal samples are found, they are further observed with the naked eye of a microscope. To determine the final results. Artificial microscopy is the gold standard for the classification of white blood cells, and the accuracy can reach more than 95%. But the efficiency of artificial microscopy is low, the speed of classification is slow, and the accuracy is affected by the experience and state of the examiners. In recent years, Depth learning has made a great breakthrough in the field of image recognition. This method has become a new research direction for white blood cell image recognition. In this paper, a white blood cell automatic classification system is designed by using the method of depth learning. The whole system includes the whole process from blood smear making to image acquisition and segmentation, and the final recognition and classification. The main work accomplished in this paper can be summarized as follows: 1) taking a large number of blood cell microscopic images from blood smears by microscope. A large number of single white blood cell images are obtained by image segmentation algorithm, and a new white blood cell database is established by using the white blood cell image obtained by segmentation. The database contains a total of four data sets, Trainberg Trainberg Test and Test.Each dataset contains five types of white blood cell images, neutrophils, eosinophils, basophil, neutrophils, eosinophils, eosinophils, eosinophils, and eosinophils. Lymphocyte and monocyte. 2) designed a new deep learning network structure, which consists of two layers of convolution layer, two layers of lower sampling layer, one layer of full connection layer, and we use the Train data set to train the network. Using the Test data set to verify the performance of the network model, the average recognition rate of leukocyte image is 98.58. 3) by adjusting the network parameters and the number of samples, the system structure is optimized. The optimized network consists of five convolution layers. Five lower sampling layers and a fully connected layer, each convolution layer and lower sampling layer contain 39 feature maps. The average accuracy of leukocyte image recognition is 99.27.4) the performance of the network model is evaluated by cross-validation. We divide all white blood cell images in the Train* and Test* datasets into 10 equal parts and 9 white blood cell images for network training and 1 for network training. Validation. According to this method, We got 10 different white blood cell databases. We used the white blood cell database to get 10 sets of leukocyte recognition accuracy. This paper presents an automatic classification method for white blood cells based on deep learning, and compares it with the traditional method of automatic recognition of white blood cell images. The method proposed in this paper is applied to neutrophilic granulocytes. The recognition rates of eosinophil, basophil, lymphocyte and monocyte were 99.70 and 99.78 respectively. The average recognition rate was 99.64.
【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】：R446.1;TP18;TP391.41

【相似文献】