\r曹玉珍1，张乾昆1，孙敬来1，张力新1，余 辉1，庞天翔2\r
\r
AuthorsHTML:\r曹玉珍1，张乾昆1，孙敬来1，张力新1，余 辉1，庞天翔2\r
\r
AuthorsListE:\rCao Yuzhen¹，Zhang Qiankun¹，Sun Jinglai¹，Zhang Lixin¹，Yu Hui¹，Pang Tianxiang\r²\r
\r
AuthorsHTMLE:\rCao Yuzhen¹，Zhang Qiankun¹，Sun Jinglai¹，Zhang Lixin¹，Yu Hui¹，Pang Tianxiang\r²\r
\r
Unit:\r1. 天津大学生物医学检测技术与仪器天津市重点实验室，天津 300072；
2. 中国医学科学院北京协和医学院血液病与血液病研究所国家重点实验室，天津 300020\r
\r
Unit_EngLish:\r1. Tianjin Key Laboratory of Biomedical Testing Technology and Instruments，Tianjin University，Tianjin 300072，China；
2. State Key Laboratory of Experimental Hematology，Institute of Hematology and Blood Diseases Hospital，Chinese Academy of Medical Sciences and Peking Union Medical College，Tianjin 300020，China\r
\r
Abstract_Chinese:\r\r针对分选富集具有治疗疾病效果的干细胞\r(\rMSCs\r)\r亚群很难实现质量控制的问题，设计了以深度神经网络作为特征映射的主成分分析\r-\r岭回归模型，实现对骨髓间充质干细胞\r(\rBM\r-\rMSCs\r)\r的定量评价．通过三维重建细胞计算基于长轴的最大截面作为模型输入；训练\r4\r层神经网络将细胞分为正常细胞与病人细胞，提取全连接层输出作为特征映射；利用主成分分析降维后的前\r3\r项主成分向量作为自变量\r\rX\r\r，样本评分作为因变量\r\rY\r\r，使用岭回归模型进行拟合，将特征与细胞活性评分相联系，实现\rBM\r-\rMSCs\r活性定量评价，为后续分选高质量的活性细胞提供依据．第\r1\r阶段通过对\r176\r例细胞样本进行数据扩增，采用\r8\r折交叉验证输入二分类神经网络进行训练，第\r2\r阶段将其中标有专家评分的\r68\r例细胞样本输入到已训练的神经网络中提取全连接层输出作为特征，利用主成分分析\r-\r岭回归模型实现定量评价，结果表明：神经网络二分类准确率\r98.75\r%\r，敏感度为\r97.84\r%\r，特异度为\r99.43\r%\r，对于定量评价，模型总体样本的\r\rR\r\r^\r2\r\r为\r0.87\r3\r6\r，拟合效果良好，可以实现对\rBM\r-\rMSCs\r定量评价．\r\r
\r
Abstract_English:\r\rDuring sorting and enrichment of stem cell subgroups\r(\re.g.\r，\rmesenchymal stem cells\r)\rwith therapeutic effects\r，\rit is difficult to achieve quality control\r．\rA principal component analysis-ridge regression model with deep neural network as feature map is designed to achieve the quantitative evaluation of bone marrow mesenchymal stem cells\r(\rBM-MSCs\r)\r．\rThe maximum cross sections based on the long axis are calculated as model inputs through the three-dimensional reconstruction of cells\r．\rA four-layer neural network is trained to enable the binary classification of normal and AML BM-MSCs\r，\rand the fully connected layer output is extracted as a feature map\r．\rIn principal component analysis\r，\rthe first three principal component vectors after dimensionality reduction are used as the independent variable X\r，\rwhereas the sample score is used as the dependent variable Y\r．\rThe ridge regression model is used to fit and correlate the characteristics with the cell activity score to achieve the quantitative evaluation of BM-MSC activity\r，\rwhich provides a basis for the subsequent sorting of high-quality active cells\r．\rFirst\r，\rdata are augmented from 176 cell samples as input of the neural network using an eight fold cross-validation\r．\rThen\r，\r68 cell samples labeled with expert scores are imported into the trained neural network and the fully connected layer output is extracted as a feature map\r．\rThe principal component analysis-ridge regression model is used to achieve quantitative evaluation\r．\rThe model results show that the accuracy of neural network classification is 98.75\r%\r，\rits sensitivity is 97.84\r%\r，\rand its specificity is 99.43\r%\r．\rFor quantitative evaluation\r，\rthe R² value of the model sample is 0.87\r3\r6\r，\rand the fitting effect is good\r．\rThus\r，\rthe quantitative evaluation of BM-MSCs\r，\rwhich can provide the basis for cell sorting and enrichment\r，\rcan be achieved\r．\r\r
\r
Keyword_Chinese:干细胞；深度学习；特征映射；主成分分析；定量评价\r

Keywords_English:mesenchymal stem cells(MSCs)；deep learning；feature map；principal component analysis(PCA)；quantitative evaluation\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6471