Automatic Detection And Classification Of Leukocytes In Blood Images

Leukocyte is an important part of human immune system.The number of leukocytes in the blood will change when the disease occurs.It is of significance to count different types of leukocytes in the blood images because the number and proportion of leukocytes can help doctor diagnose the disease.The classification of leukocytes in blood images includes three steps:leukocyte location and segmentation,feature extraction and selection,leukocyte classification.This thesis completes the classification of five types of leukocytes in blood images according to the above three steps.In the part of leukocyte location and segmentation,this thesis locates the leukocyte by using the method based on channels combination of RGB and achieves a high locating accuracy.Then,this thesis segments nucleus by using the methods of RGB channels combination and Otsu.Lastly,the active basis model is used to segment cytoplasm.We use local edge model to solve the offset problem of active base model in leukocyte detection.In order to restrain the over-segmentation of cytoplasm,outlier detection and convex hull segmentation are used and the result of experiments shows that the improved method can obtain a better effect for cytoplasm segmentation.In terms of feature extraction and selection,this thesis extracts a total of 100 features,including morphological features,color features and texture features of nucleus and cytoplasm.According to the characteristics of the convex hull region of leukocyte nucleus,we also extract morphological features,color features and texture features of the convex hull region of nucleus.We add new leukocyte features,for example,perimeter difference rate and edge gradient integral.Fisher Score is used to select good features,at last,the number of features is reduced to 70.This thesis also uses Z-Score to standardize the features.In the part of leukocyte classification,this thesis uses three classifiers(SVM,Random Forest and K-NN)to do the experiment of leukocyte classification.The experimental results show that all three classifiers get high classification accuracies of lymphocyte and neutrophils.SVM has the best classification effect on leukocyte five classification.On the basis of the above study,this thesis designs a complete solution for the automatic classification of leukocytes in blood cell image.This solution includes seven parts:leukocyte localization based on RGB spatial component,leukocyte nucleus segmentation based on RGB spatial component,leukocyte cytoplasm segmentation based on the active basis model,feature extraction,feature standardization based on Z-Score,feature selection based on Fisher Score,SVM.