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

The classification of human leukocytes plays an important role in the clinical diagnosis of diseases.In clinical practice,white blood cells are generally recognized by artificial microscopic examination,which is time-consuming and labor-intensive.At the same time,there exist some human errors.Therefore,intelligent detection methods are widely concerned and expected.The current research focuses on the use of computers instead of artificial white blood cell recognition.In this paper,the images of white blood cells are automatically classified by using convolutional neural network and capsule network.Related literature shows that,in recent years,when researchers use computer to explore the process of the classification and recognition of white blood cell image,because the size,shape and edge features in cells and nuclei of white blood cell image are the main features of the automatic classification,researchers need to segment the white blood cells first,and then extract the relevant features,and classify and recognize the extracted features.Prophase research work shows that the effect of segmentation when using the existing white blood cell segmentation algorithm for cell segmentation is not ideal.Through comparison,it is found that in the dataset image used in this paper,the white blood cells are very similar to the background color and the brightness of the white blood cells in the image is uneven.In view of these problems in the dataset used in this paper,a white blood cell image segmentation algorithm based on RGB and C-Y color space is put forward.First of all,the white blood cell image in the RGB color space is segmented,and then the original RGB image is converted into C-Y image,and the B-Y color component containing complete information is extracted.The white blood cell image is obtained through the connected region area screening,opening operation,and pixel operation,and the G image after contrast stretching is extracted.The complete cell nuclei image is obtained after repeating the above process.Experiments show that the proposed algorithm has high segmentation accuracy for eosinophils,lymphocytes,monocytes and neutrophils,and achieves accuracy rates of 94.33%,91.60%,97.72%,and 98.66%,respectively,which lays a foundation for the classification accuracy of white blood cell image.In recent years,with the wide application of neural networks,the automatic classification of white blood cell images has become one of the hot research fields.In order to get closer to the actual needs,this paper adjusts the research strategy,eliminating the process of white blood cell image segmentation,and redesigns the automatic classification neural network adapted to the white blood cell image dataset in this paper.In this paper,the convolutional neural network and the capsule network are merged into a new network to perform four-category classification on white blood cell images,and two classification models are proposed.Aiming at the shortcoming that the pooling layer will lose some detailed information during the image processing,and considering its advantages of being more suitable for large-scale datasets,it complements the advantages of the capsule network and can improve the accuracy of white blood cell classification further.The first network classification model is the Dilated Convolution Capsule Network(DCCNet),which is a seven-layer network model composed of dilated convolutions.It replaces part of the pooling in the network with dilated convolutions,which simplifies the network and reduces the time required for training.Another classification model is the Deep Multi-Lane Capsule Network(DMLCN)for white blood cell classification.The DMLCN model belongs to the category of convolutional neural networks,and its functions are very powerful.It combines the advantages of Inception Network(Inception-v4),Residual Neural Network(ResNet)and Capsule Network(CapsNet).The classification accuracy of training,verification and testing of this model are 99.99%,97.19% and 82.39%respectively.The training time of each epoch is 171 seconds,half of the original capsule network.Compared with existing methods,the DMLCN model provides better classification performance in terms of global accuracy.