Image Detection And Classification Technology Of White Blood Cells Based On Convolutional Neural Network

White blood cells are the defenders of the immune system in the human body,protecting the human body from germs.When the number of cells in five types of white blood cells exceeds the normal range,it often indicates that a certain disease may occur in the human body.Therefore,the correct recognition of five types of white blood cells have great significance for medical aid diagnosis.The gold standard for white blood cell recognition is microscopic examination.But microscopic examination is less efficient and susceptible to the level of the inspector.Electrical impedance method,flow cytometry and other method have better results,but the instrument cost is relatively high.The recognition method based on image processing has the advantages of being intuitive and easy to re-examine by doctors.However,the inaccurate segmentation,feature redundancy and classification model selection in traditional image processing methods affect the recognition effect.The image processing method based on convolutional neural network can simplify the recognition step and is expected to complete the low-cost,high-precision white blood cell recognition task.In this paper,convolutional neural networks and transfer learning methods are used to detect and classify white blood cells.The main research work is as follows:(1)White blood cell image detection based on convolutional neural network.In order to improve the detection accuracy,this paper makes improvements based on the YOLO method.White blood cells are treated as a positive class,and the background is regarded as a negative class for white blood cell detection.Firstly,data expansion is performed using data augmentation methods such as rotation and image stitching.Secondly,The Kmeans++algorithm is used to initialize the Kmeans clustering.The Kmeans algorithm is used to cluster the white blood cell bounding box size to obtain the initial candidate bounding box.The feature is extracted by the convolutional neural network with residual module.A bounding box regression is performed on the initial candidate bounding box to obtain a white blood cell prediction bounding box.Thirdly,use non-maximal suppression to select the optimal bounding box in the prediction box to get the white blood cell position.(2)Classification of white blood cell images based on transfer learning.For small data volumes,white blood cell data is not enough to fit the network model.In this paper,the natural image is used as the source domain to learn the bottom features of the image on the InceptionV3 network,the network structure and parameters are fixed and transfer to the neural network with the white blood cell image as the target domain,and multiple fully connected layers are added for classification training to realize the five types of white blood cells automatic classification function.This paper experiments on white blood cell images and compares them with other deep learning methods and traditional image processing methods.The experimental results show that the target detection method achieves white blood cell detection accuracy of 97.84%,which is better than YOLO.The average accuracy of the five types of white blood cell classification in the target classification network is 98.58%,which is better than support vector machine,random forest,AlexNet,VGG,Inception V3 without transferring.In this paper,the convolutional neural network is used to accurately detect white blood cells,and the transfer learning network is used to accurately classify white blood cells.