卷积神经网络在急性髓系白血病流式细胞术自动诊断中的应用

Application of convolutional neural network in flow cytometry diagnosis of acute myeloid leukemia

目的建立卷积神经网络(CNN)模型对流式细胞术(FCM)数据进行自动分析,实现急性髓系白血病(AML)的初步诊断,探究将CNN模型应用于FCM数据分析中的可行性。方法以FlowRepository数据库和新疆维吾尔自治区人民医院临床检测中心获得的骨髓FCM数据进行CNN应用的探索性研究,数据均已被临床确诊是否患有AML。其中,公开数据按照6∶2∶2划分训练集、验证集和测试集,本地数据作为外部测试集;为了使FCM数据能够适应CNN模型,提出一种基于图像矩阵原理的FCM数据结构,对原始数据进行预处理后,提取与AML初步诊断相关的变量,包括侧向散射光和CD45、CD13、CD33、HLA-DR、CD117、CD34的各抗原表达水平,将各变量写入矩阵;对训练集使用细胞抽样和数据增强方法增大样本量,在Python中使用keras软件包构建LeNet-5 CNN模型,将训练集和验证集分别用于模型的训练和调参,评价模型在测试集上的性能。结果CNN在两测试集上识别AML的准确率分别为0.931、0.851,灵敏度为0.667、0.636,特异度为0.968、0.940,受试者工作特征曲线下面积(AUC)为0.940和0.917。结论基于提出的FCM数据结构,CNN模型能够实现对AML的初步诊断,表明CNN在FCM数据分析中具有一定的应用价值。

Objective A convolutional neural network(CNN)model was established to automatically analyze flow cytometry(FCM)data to achieve the preliminary diagnosis of acute myeloid leukemia(AML),and explore the feasibility of applying CNN model to FCM data analysis.Methods The exploratory study of CNN application was carried out using the bone marrow FCM data obtained by the FlowRepository database and the Clinical Testing Center of Xinjiang Uygur Autonomous Region People's Hospital,and the data had been clinically confirmed whether AML was present.Among them,the public data was divided into training sets,validation sets and test sets according to 6:2:2,and local data was used for external test;In order to adapt the FCM data to the CNN model,an FCM data structure based on the image matrix principle was proposed,and after preprocessing the original data,the variables related to the preliminary diagnosis of AML were extracted,including sidescattered light and the expression levels of CD45,CD13,CD33,HLA-DR,CD117,CD34,and each variable was written into the matrix.Cell sampling and data augmentation methods were used to increase the sample size of the training set,the keras software package was used to build the LeNet-5 CNN model in Python,and the training set and the validation set were used for model training and parameter tuning respectively to evaluate the performance of the model on the test set.Results The accuracy of CNN to identify AML on the two test sets was 0.931,0.851,the sensitivity was 0.667,0.636,the specificity was 0.968,0.940,and the area under the receiver operating characteristic curve was 0.940 and 0.917.Conclusion Based on the proposed FCM data structure,the CNN model can realize the preliminary diagnosis of AML,indicating that CNN has certain application value in FCM data analysis.