Diagnosis Of Cytomorphology For Leukemia Based On Convolutional Neural Network Combined With Transfer Learning

Leukemia is a proliferative disease of abnormal bone marrow and hematopoietic tissue.Clinically,the cellular morphological diagnosis of leukemia relies mainly on manual microscopy of bone marrow smears.However,this method is laborious,time-consuming,inefficient,and the test results are subjective.In view of the above problems,this paper proposes a method based on convolutional neural network(CNN)combined with transfer learning to establish a classification model for microscopic images of bone marrow smear and single image of white blood cell.And a rapid,efficient and accurate diagnosis of leukemia cell morphology has been achieved.The main contents are as follows:1.The cellular morphological classification diagnosis of leukemia based on microscopic images of bone marrow smear.Firstly,bone marrow smears microscopic images of acute myelocytic leukemia(AML),acute lymphoblastic leukemia(ALL),chronic myelocytic leukemia(CML)and healthy volunteers are collected to construct data set Data-1.Then,using GoogleNet,ResNet and DenseNet to establish the classification model of Data-1,DenseNet classification results are the best and the prediction accuracy is 78%.Finally,the transfer learning is combined with DenseNet to further improve the prediction accuracy of the model.The results show that the total accuracy of the DenseNet model combined with transfer learning is 95.2%,and the prediction accuracy of the normal group,AML,ALL and CML are 90%,99%,97% and 95%,respectively.2.The cellular morphological classification diagnosis of AML based on whole microscopic images of bone marrow smear.Firstly,bone marrow smears microscopic images of the four subtypes(M2,M3,M4,M5)are collected to construct data set Data-2.Then,using GoogleNet,ResNet and DenseNet to establish the classification model of Data-2,DenseNet classification results are the best and the prediction accuracy is 70%.Finally,the transfer learning is combined with DenseNet to further improve the prediction accuracy of the model.The results show that the total accuracy of the DenseNet model combined with transfer learning is 93.5%,and the prediction accuracy of the M2,M3,M4 and M5 are 91%,90%,93% and 100%,respectively.3.The cellular morphological classification diagnosis of AML based on single image of white blood cell.Firstly,Otsu based on genetic algorithm(GA)is used to perform dual-threshold segmentation on bone marrow smears microscopic images of AML,white blood cell images were extracted and data set Data-3 of eight white blood cell images are constructed.In order to solve the problem of imbalance in the distribution of various types of white blood cell images,a new data set Data-3* is generated by data augmentation.Then,using GoogleNet,ResNet and DenseNet to establish the classification model of Data-3*,DenseNet classification results are the best.Finally,the transfer learning is combined with DenseNet to further improve the prediction accuracy of the model.The results show that the total accuracy of the DenseNet model combined with transfer learning is 91.2%,and the prediction accuracy of eight white blood cells(myeloblast,promyelocyte,myelocyte,metamyelocyte,band,neutrophil,monocyte and other types)are 91%,95%,82%,99%,87%,92%,93% and 90%,respectively.The results of this study indicate that the cellular morphological diagnosis of leukemia based on CNN combined with transfer learning is feasible,and the method has the characteristics of fast,accurate and objective.