Research On Automatic Analysis Method Of Flow Cytometry Data Based On Neural Network Model

Objective: In view of the limitations of the traditional flow cytometry(FCM)data analysis method,the neural network model is proposed to realize the automation of data analysis from the two aspects of cell subgroup identification and disease diagnosis,and the feasibility of applying neural network in FCM data analysis is explored by taking bone marrow cell subgroup classification and acute myeloid leukemia(AML)diagnosis as examples.Methods: In the classification of bone marrow cell subsets,the back-propagation algorithm(BP)neural network model is constructed,and the bone marrow FCM data of 10 healthy volunteers are extracted to form a data set.The model self-learning extracts the features on the training set to obtain the optimal parameters.Taking the results of manual gating analysis as the gold standard,the effect of BP neural network model on cell subpopulation classification was evaluated by cross validation,and compared with the decision tree and K-means model.In the diagnosis of AML,the Le Net-5 convolutional neural network(CNN)model is constructed,and the FCM data obtained from the Flow Repository database and the clinical testing center of the People’s Hospital of Xinjiang are extracted.The public data is divided into6:2:2,which are used for model training,verification and testing respectively.The local data are used for model testing Externally,and the data have been clinically confirmed whether AML is diagnosed.In order to adapt FCM data to the CNN model,a data structure based on the principle of image matrix is proposed.After preprocessing the original data,the variables related to the preliminary diagnosis of AML are extracted,including the level of antigenic expression of lateral scattered light(SSC)and CD45,CD13,CD33,HLA-DR,CD117 and CD34,and each variable is written into the matrix to form a data set.Cell sampling and data enhancement methods are used to increase the sample size of the training set.The model is iteratively learned on the training set,and the training is stopped and saved when the loss function of the verification set falls to the lowest point,and the performance of the model on the test set is evaluated.Results: In the problem of bone marrow cell subpopulation classification,the BP neural network model fits well on the data set,and the clustering profile of the model on the scatter diagram is basically consistent with that of the artificial gating,which can well reproduce the artificial analysis results.The clustering effect of BP neural network and decision tree model is similar,and the accuracy is better than that of K-means model(p<0.05),and the clustering accuracy is 0.970,0.972 and 0.899 respectively.In the problem of AML diagnosis,the accuracy of CNN in identifying AML on the two test sets is 0.931 and0.851,the sensitivity is 0.667 and 0.636,the specificity is 0.968 and 0.940,and AUC is 0.940 and 0.917.Conclusion:1.The automatic recognition of bone marrow cell subsets can be realized by constructing BP neural network model to analyze FCM data.2.The automatic identification of AML can be realized by constructing CNN model to analyze FCM data.3.The application of neural network model in FCM data analysis has certain feasibility,which can provide reference for subsequent research.