The Research Of Leukocytes Classification Based On Deep Learning

In the clinical,leukocyte classification is very important in blood examination and the quick and accurate classification is an essential part of today’s medical research.At present,the recognition of leukocytes depends clinically on the hematology analyzer and microscope.In the clinical,the blood samples from patients are measured firstly in the hematology analyzer,and observed under a microscope if the result is abnormal.The latter is the gold standard of the leukocyte classification with more accuracy than 95%.But this method is inefficient and the speed of classification is slower.The accuracy affected by the inspection personnel experience and mental states.In recent years,Deep Learning make a significant breakthrough in the field of image recognition,and the leukocyte image recognition with the method would be a new research direction.In this paper,a new system based on Deep Learning is designed to recognize leukocytes.The whole system includes blood smears preparation,image acquisition and segmentation,network training and recognition.The main work of this paper can be summarized as the following aspects:1)A large number of microscopic images were taken from blood smears using a microscope,and then a large number of individual leukocyte images were obtained through the image segmentation algorithm.These leukocyte images were used to establish a new leukocyte database.The database contains four datasets,Train,Train*,Test and Test*.Each dataset includes Neutrophil,Eosinophil,Basophil,Monocyte and Lymphocyte.2)A new network based on Deep Learning was proposed in this paper.The network contains two convolution layers,two pooling layers and one fully connected layer.We use the Train dataset to train the network,and use the Test dataset to test the network.The average accuracy of leukocyte is 98.58%.3)The network was optimized by adjusting the network parameters and the sample size.The optimized network contains five convolution layers,five pooling layers and one fully connected layer.And there are 39 feature maps in each convolution layer and pooling layer.The average accuracy of leukocyte from the optimized network is99.27%.4)The performance of the optimized network was evaluated by cross validation.All of leukocyte images in Train and Test were divided randomly into ten parts.Nine parts were used for training network and the rest for test.In this way,we obtained ten different leukocyte databases for training and test network.We take the average of ten accuracies from ten different leukocyte databases as the standard of the system performance.This paper proposed a new method based on Deep Learning for leukocyte classification,and compared with the traditional methods.The accuracy of Neutrophil,Basophil,Eosinophil,Lymphocyte and Monocyte are respectively 99.70%,99.63%,99.78%,99.49% and 99.62%.The average accuracy of leukocyte is 99.64%.