Classification And Recognition Of White Blood Cells Based On Deep Learning

At present,the classification of white blood cells is mostly classified by three-classification blood test instrument or five-classification blood test instrument,and these instruments cannot identify the classification very accurately.These instruments can only count the conventional three or five classification of white blood cells,and no more than 40 types of white blood cells cannot be measured in detail.If we can identify the specific type of white blood cells more intelligently,precisely and comprehensively,instead of macroscopically divided three classifications or five classifications,it will play a very good auxiliary role for doctors in the diagnosis and treatment of leukemia patients.In order to solve these problems,based on the knowledge of deep learning,this paper focuses on how to construct 40 types of white blood cell classification model and improve the classification accuracy.The main work of this paper is as follows:In this study,a classification method of white blood cells based on deep convolution neural network and residual algorithm was designed.The essence of the identification of different white blood cells is the classification of a visual image.Convolution neural network has unique advantages in the extraction of local features on the image,and it has three major features: local perception,parameter sharing and multicore convolution.These characteristics make the convolutional neural network greatly reduce the training parameters compared with the ordinary neural network,and make the feature extraction more efficient and convenient.Based on this cognition,we design a convolution neural network classifier with 33 layers of structure through a lot of experiments and practice.After training a large number of image samples,we can extract a network model with 82% classification accuracy.Through this model,40 kinds of white blood cells can be accurately identified and classified.The number of all types of white blood cell data we have obtained is extremely uneven and scarce.Images of three-classifications of white blood cells(large cells,medium cells,small cells)obtained from the usual blood test are easier to obtain and have more resources.But the more detailed 40-classifications of white blood cell image resources are relatively scarce.In this paper,a new method of white blood cell classification based on transfer learning is proposed to solve the problem of relatively scarce samples.In this paper,we first design a convolutional neural network structure with three-classifications,which can extract the salient features of white blood cell images automatically,and then fit the network parameters with sufficient three-classification training samples,then extract a better three-classification model.Then we design a convolutional neural network architecture with 40-classifications(CNN),and the feature extraction layer of the architecture is the same as the previous three-classification models.This design allows the feature extraction parameters of the three-classification model to be migrated to the 40-classification CNN architecture.Then we use the relatively inadequate 40-classification data sets to train the 40-classification architecture,and finally we can extract a network model with 89% classification accuracy.Finally,a lot of experiments show that the transfer learning method proposed in this paper is better than just using CNN or conventional methods,and proves the feasibility and accuracy of the proposed method.