Research On Classification Of Blood Cell Image Based On Deep Learning

In clinical testing,the classification and recognition of blood cells can effectively assist in the diagnosis of many diseases.At present,blood cells are generally classified and recognized by artificial microscopy,which is time-consuming and labor-intensive,and is also susceptible to interference by human factors.Therefore,the intelligent classification methods are widely concerned and expected.The current research focuses on the use of deep learning methods instead of artificial white blood cell classification.The classification and recognition of white blood cells in blood cell images can be divided into three parts: the segmentation of single white blood cell images,the amplification and establishment of dataset,and the classification of white blood cell images.In the segmentation part of single white blood cell images,a blood cell recognition algorithm based on CNN was proposed to solve the tedious and time-consuming procedure of blood cell manual counting by traditional blood cell counter and other devices as well as the high complexity of traditional white blood cell image segmentation algorithms.The Yolov4-tiny object detection algorithm is used to automatically identify and count three types of blood cells by combining Res2 Net and CBAM,while segmenting single white blood cell images based on the upper-left coordinate of the leukocyte prediction box.The performance of the blood cell recognition model is enhanced by fusing Res2 Net and CBAM into the Yolov4-tiny model to extract multi-scale features with granular level representation and increase the range of the receptive field in each network layer.Through the experiment of public blood smear image dataset,it can automatically identify and count white blood cells,red blood cells and platelets with recognition accuracy of 93.44%,96.09% and 96.36%,respectively.Compared with other recognition models based on CNN,this algorithm has high recognition accuracy and strong generalization,which can significantly improve the efficiency of blood cell recognition.In the classification part of leukocyte images,firstly,in order to solve the problem of insufficient and unbalanced datasets,the images obtained by segmentation are processed by image transformation operation and GAN to increase the number of leukocyte images and to establish the classification dataset.Secondly,to effectively solve the problem of classification and recognition of leukocyte images,a white blood cell image classification model based on Inception-V4 is proposed.The classification network parameters were initialized by using transfer learning,and Res2 Net and Inception-V4 were fused into a new network to extract multi-scale features to improve the classification accuracy of the model.Finally the performance of the classification model is further enhanced by using the improved RBF-Softmax as the loss function of the model to enhance the discriminative ability of the deep learning features.The experimental results show that the proposed model can achieve 94.31% classification accuracy in the classification task of five types of leukocyte images.