The Research Of Automatic Five Classification Of Leukocytes In Automated Blood Smear

Leukocytes are essential cells in the human immune system.They can remove antibodies secreted by foreign bodies and play an important role in the recovery of body damage,resistance to virus invasion,and immunity to blood-related diseases.Till today,leukocyte testing is still an important basis for modern medical diagnosis.By measuring the total number,relative ratio,and morphology of various leukocytes in human blood,it is possible to determine whether there are diseases,the causes and severity of diseases,and to develop a corresponding medical plan and observe the medical efficacy.Therefore,it is of great significance to accurately,efficiently and cheaply classify and count the various leukocytes in the blood.Image-based leukocyte intelligence five-category system is one of the important research contents of medical image processing.The system can replace artificial microscopy by simulating human visual processes to achieve automatic classification and counting of leukocytes.It mainly involves techniques such as cell localization,cell segmentation,feature extraction and intelligent identification.For image understanding such as image feature extraction and recognition,image segmentation is a critical step.Before the deep learning technique appeared,it is necessary to extract the various features of the nucleus and cytoplasm manually from the leukocytes separated from blood smear image.Then,these features are used by various intelligent identification strategies(such as BP neural network,and SVM etc)to perform leukocyte classification.In recent years,the deep learning has received more and more attention in the field of image recognition.Using this method for leukocyte image recognition has become a new trend.In this thesis,the machine-stained leukocyte smear automated five-category was studied using the automatic blood smear provided by NanChang Tecom Technology Company.The completed work in this thesis is summarized as follows:1.This thesis used both traditional machine learning method and deep learning method to classify leukocyte images.The former involves the processes of leukocyte localization,feature extraction,recognition and classification,while the latter requires preparation of a large number of well-characterized sample cell data,these data are used to train the corresponding deep learning networks,the trained network can be used for leukocyte recognition.2.To separate the leukocyte cells from the blood smear images containing the residual red blood cell impurities and the contamination of dyestuffs,we proposed an algorithm to extract a single cell by using a localization method.The method used global threshold method and morphological operations to locate the leukocyte,then extracted the leukocytes by calculating the self-adaptive window containing each leukocyte image based on the centroid of the connected region of the leukocyte and the value of the minimum circumscribed rectangle length and width.3.In the traditional leukocyte identification method,after extracting the leukocyte image from the slice image,we first extracted the Histogram of Oriented Gradient(HOG)and gray level co-occurrence matrix(GLCM)related features of leukocytes,and then used Support Vector Machine(SVM)to perform identification and classification of leukocytes.HOG is a descriptor on a local unit of an image.It maintains good invariability to image shape and lighting changes.The support vector machine is a classifier with excellent classification ability on small samples,which can effectively avoid “dimensional disasters”.The overall recognition rate of the system in the experiment reached 95.6%,the recognition rates of the five types of leukocytes in the entire system,basophils,eosinophils,lymphocytes,monocytesand neutrophilswere 97.1%?84.2%?96.3%?89.2%?99%.4.In the classification study of leukocyte images based on deep learning,a five-category leukocyte method based on fine-tuned AlexNet network was proposed on the deep learning framework(caffe platform).Although the entire processing of leukocytes classification does not require human participation,the features of the leukocyte can be more accurately extracted and analyzed.The recognition rates of the five types of leukocytes in the entire system,basophils(BAS),eosinophils(EOS),lymphocytes(LYM),monocytes(MON)and neutrophils(NEU)were 100%,79%,96.3%,94.6% and 99% respectively and the overall recognition rate was 96.3%.Through comparison and analysis of the two kinds of classification methods,we conclude that the deep learning network can automatically extract the essential features that are more conducive to data classification,this will have important implications for improving the accuracy of diagnosis of blood diseases.Therefore,deep learning will have better prospects in the field of medical image processing research.