Research On Label-free Cell Classification Technology Based On Deep Learning

The types and contents of different cells in human blood are closely related to human health.Therefore,analyzing the content of different subtypes of white blood cells(WBCs)and the types of circulating tumor cells(CTCs)has become an vital clinical technique for evaluating immune system status and screening diseases.Based on fluorescence labeling,the manual cell classification method and the automatic classification method represented by flow cytometry adverse effects on classified cells,while they also have the problems of low classification efficiency and high cost,which makes it difficult for cell classification technique to meet the medical demand of immune status assessment and disease screening in less developed areas.Based on deep learning technology,this paper proposed a batch normalization free(BN-free)convolution neural network and a convolution neural network with feature fusion for the classification of label-free WBCs and label-free CTCs,respectively.The BN-free structure overcomes the inconsistency behavior between the training and testing caused by BN operation and the degradation of classification performance caused by non-independent identically distributed samples.The feature fusion structure compensates for the disadvantage that existing convolution neural networks could not fully exploit the detailed features of holographic cell images.These specially designed structures effectively improve the accuracy of classification.After full comparative analysis,the accuracy of proposed method in the classification of label-free WBCs and label-free CTCs is significantly better than the typical convolutional neural network,reaching 81.4% and 94% respectively.In addition,compared with the current mainstream classification methods,the proposed method in this paper also has significant advantages such as label-free,high efficiency,and low cost,making it have important application value.