Research On Leukocyte Classification Technology Based On Deep Learning With Object Detection

White blood cell（WBC）is an essential part of blood and plays significant role in the human immune system,which can be divided into many subtypes according to their different functions and morphologies.The content of various subtypes of WBCs is usually maintained within a certain range in the human body,while deviant levels are important warning signs for diseases.Hence,the detection and classification of WBCs is an essential diagnosis technique.Traditional WBC classification is usually realized by professionals with the help of microscopes.This manual classification is difficult to achieve unbiased estimation and is seriously time-consuming.Therefore,the automatic WBC classification technology based on pattern recognition and image processing has been actively studied.At present,there are two popular automatic WBC classification technologies:feature engineering based on machine learning and automatic feature extraction based on deep learning.The former method is often combined with miniaturized microfluidic systems or cell analysis instruments to realize WBC classification by manually selecting features.While the latter method relies on high-resolution imaging systems and the learning ability of neural networks to distinguish different WBC.These two technologies can effectively avoid the defects of manual classification and achieve good performance.However,they all need to segment the collected original cell image,which increases the workload to a great extent,especially when the cells are densely distributed.More importantly,the accuracy of segmentation will seriously affect the performance of classification,so it is difficult to realize the optimization of the system end-to-end.To overcome these shortcomings,this paper applies object detection based on deep learning to realize the classification of WBC,which can realize the classification and location of multiple targets in the original image at the same time.This method not only guarantees the integrity of the image but overcomes the adverse effects of segmentation operation.In this paper,Yolov4^[1],Mobile Netv2-SSD^[2,3],Faster-RCNN^[6]based on Inceptionv2^[4]and Res Net50^[5]are used to classify lymphocytes,eosinophils,neutrophils,and monocytes on the Blood Cell Classification Datasets（BCCD）.The accuracy achieved 96.50%,96.50%,98.00%and 97.27%,respectively,while the detection speed reached 59.99FPS,44.72FPS,29.4FPS and 25.24FPS respectively.In terms of the superior performance of Yolov4,this detector is then applied to the label-free WBCs classification.Through the rebuild and optimization of the dataset,the classification accuracy of about 91.33%was achieved.This image classification technology based on object detection can realize the simultaneous classification of multiple WBCs without segmentation,and make rapid predictions under the condition of high classification accuracy,which provides an effective reference for disease prevention,prediction,and point-of-care test.