白细胞散点图识别模型的建立与验证  

Establishment and validation of identification model for WBC scattergram

作　　者：赵天赐 李建英[1] 连荷清 刘丹[1] 王庚[1] 王欣[1] 李柏蕤 吴卫[1] ZHAO Tianci;LI Jianying;LIAN Heqing;LIU Dan;WANG Geng;WANG Xin;LI Bairui;WU Wei(Department of Clinical Laboratory,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences,Beijing 100730;Beijing Xiaoying Technology Co.,Ltd.,Beijing 100084,China)

机构地区：[1]中国医学科学院北京协和医院检验科,北京100730 [2]北京小蝇科技有限责任公司,北京100084

出　　处：《临床检验杂志》2022年第4期246-250,共5页Chinese Journal of Clinical Laboratory Science

基　　金：北京市临床重点专科医学检验科卓越项目(ZK201000);AI+健康协同创新培育项目(Z221100003522004)。

摘　　要：目的建立全血细胞分析白细胞分类(WDF)通道散点图的识别模型,验证其有效性。方法运用卷积自编码技术,算法训练采用pytorch框架,模型训练采用AlexNet为基准网络,加入先验知识算法形成双重验证机制。采用32729份WDF通道散点图作为初始数据集,由3位检验技师对散点图进行标定,根据2次标定结果是否完全一致将初始数据集分为特征显著散点图和特征不显著散点图,训练集、验证集和测试集A所含2类散点图数量按照8∶1∶1进行划分,即26185份散点图作为训练集,3272份散点图作为验证集,3272份散点图作为测试集A。再通过日常工作中11043份散点图(测试集B)验证模型精确率,并与检验技师的判断结果进行对比分析。结果模型和检验技师在测试集A、B的精确率分别为0.956、0.967和0.924、0.932,模型和检验技师在测试集B所含正常、异常散点图精确率分别为0.979、0.935和0.992、0.795,所含特征显著、不显著散点图精确率分别为0.985、0.921和1.000、0.662。结论当散点图特征不显著的情况时,该模型能够给出更加准确的结论,为检验技师提供有价值的参考,辅助检验技师作出正确的判断。Objective To establish an identification model for WBC scattergram of WDF channel in blood routine examination and verify its effectiveness.Methods In application of convolutional autoencoder,pytorch framework was used for algorithm training,AlexNet was used as the benchmark network for model training,and a dual validation mechanism was formed by adding a priori knowledge algorithm.A total of 32729 images of WBC scattergram from WDF channel were preprocessed as the initial data set.The scattergrams were calibrated by three skilled laboratory technicians.The initial data set was divided into featuresignificant or featureinsignificant scattergrams according to consistency of the two calibration results.The amount of two types of scattergram contained in the sets of training set,validation set and test set A was divided according to proportion of 8∶1∶1,i.e.,26185 scattergrams for training set,3272 scattergrams for validation set and 3272 scattergrams for test set A.The accuracy of the model was analyzed by using scattergrams of 11043 images from daily work(test set B)and compared with the judgment results of the laboratory technicians.Results The accuracy rates of model and laboratory technicians in test set A and test set B were 0.956,0.967 and 0.924,0.932.The accuracy rates of model and laboratory technicians in normal scattergram and abnormal scattergram contained in test set B were 0.979,0.935 and 0.992,0.795,and the accuracy rates of featuresignificant and featureinsignificant scattergram were 0.985,0.921 and 1.000,0.662 respectively.Conclusion When the features of scattergram were not significant,the model could give more accurate conclusions and may provide valuable reference for laboratory technicians to make correct judgments.

关 键 词：白细胞散点图 医学图像分析 卷积自编码技术 

分 类 号：R446.11[医药卫生—诊断学]