The Research And Implementation Of The White Blood Cell Classification Method Based On Unbalanced Sample Sets

The analysis of peripheral blood smears is of great significance for evaluating human health.Doctors can find the cause of disease for patients according to the relevant image,and then make the corresponding treatment plan.At the same time,with the rapid development of modern science and technology and in order to assist doctors to make pathological decisions quickly,efficiently and accurately,the means of intelligent medical assistance has gradually entered the medical field to assist doctors to make relevant pathological decisions.Among them,the most popular method is to introduce artificial intelligence technology into the field of medical diagnosis.However,it is not easy for the researchers to collect a sufficient and balanced white blood cell data set.In addition,many methods for classifying white blood cells are relatively cumbersome and time consuming.At the same time,these methods usually require extremely experienced experts to choose some specific features for classification,which results in that the classification results are sometimes lack of objectivity due to the subjectivity of the experts.Therefore,this paper proposes a white blood cell image classification method for unbalanced sample sets.This method is mainly based on deep learning and uses convolutional neural network to automatically learn to obtain the high-level white blood cell features in the image for classification.The method in this paper can be divided into two major stages: the pre-processing stage for white blood cell images and the classification stage for white blood cells based on the pre-processed images.Among them,in the pre-processing stage of white blood cell images,first of all,this paper implements the target detection in the original white blood cell image based on the Faster R-CNN network model,and marks the white blood cell areas from the raw images.Then,based on the labeled white blood cell area,the color space RGB to HSV is mapped to label the nuclei;the improved SLIC algorithm is based on the labeled nuclei position to further segment the white blood cell area.After marking the white blood cell area,this paper then proposes a new method called ‘RRA method' to remove the redundant information based on the idea of sliding window algorithm.This method removes 75% of irrelevant areas and has improved the feature ratio of white blood cell area.Finally,based on the pre-processed images,this paper uses the DCGAN network to increase the data samples of the five classes at some extent.In the stage of classifying the pre-processed WBC images,considering that the current commonly used classification networks often sacrifice time efficiency at a certain extent while ensuring the classification accuracy,so aiming at such a situation,this paper builds a new classifier model based on multiple batch normalization layer(in this paper,this network is called ‘MBNnet')to complete the final classification task.Finally,the effectiveness of the preprocessing method proposed in this paper is proved through comparative experiments.Then,this paper conducts sufficient experiments on the preprocessed data set to compare MBNnet with commonly used classification networks: VGG16,Alex Net,and Res Net50.And the experimental results show that the proposed method achieves an overall accuracy of about 99.69% at a lower training time cost.