基于实时荧光PCR时序数据的深度神经网络知识蒸馏的阴阳性识别  

Positive and Negative Recognition of Deep Neural Network Knowledge Distillation Based on Real-Time PCR Time Series Data

作　　者：侯剑平 张蕊 赵万里 刘玉凤 张烁 王超 HOU Jianping;ZHANG Rui;ZHAO Wanli;LIU Yufeng;ZHANG Shuo;WANG Chao(Autobio Labtec Instruments Co.,Ltd.,Zhengzhou Henan 450016,China)

机构地区：[1]安图实验仪器(郑州)有限公司,河南郑州450016

出　　处：《中国医疗设备》2024年第6期8-16,29,共10页China Medical Devices

基　　金：国家重点研发计划(2022YFC2406400)。

摘　　要：目的提出一种基于实时荧光聚合酶链式反应(Polymerase Chain Reaction,PCR)时序数据的深度神经网络知识蒸馏的阴阳性识别技术。方法通过使用神经网络进行时序分类,在增加鲁棒性的同时减少异常值对模型结果的影响;结合知识蒸馏技术压缩深度神经网络,降低神经网络对算力资源的需求,以适应较低的计算资源运行环境。结果收集全自动核酸提纯及实时荧光PCR分析系统Auto Molec 3000数据302260条,将数据集按80%和20%的比例分成训练集和测试集,在测试集上验证知识蒸馏后的全卷积神经网络(Fully Convolutional Networks,FCN)模型与长短时记忆结构(Long Short-Term Memory,LSTM)模型,在准确度、敏感度、特异性和F1均高于0.999的前提下,FCN模型可缩小21.3倍,LSTM模型参数可降低12.8倍;60452条样本的预测时长分别为5.6900、2.2516 s。结论模型可保证整体的准确度性能,满足对部署环境的算法要求低的需求。Objective To propose a new technique for positive and negative recognition based on deep neural network knowledge distillation based on real-time polymerase chain reaction(PCR)time series data.Methods Neural networks to accurately classify time series were used,which could increase robustness and reduce the influence of outliers on model results.Deep neural networks were compressed by knowledge distillation,the demand for computational power of neural networks was reduced to adapt to the operating environment of low computational resources.Results 302260 pieces of data from AutoMolec 3000 automatic nucleic acid purification and real-timefluorescent PCR analysis system were collected,the data set was divided into training set and test set according to 80%and 20%ratio,and the fully convolutional networks(FCN)and long short-term memory(LSTM)after knowledge distillation were verified on the test set.When the accuracy,sensitivity,specificity and F1 were all higher than 0.999,the FCN model could shrink 21.3 times,and the LSTM model parameters could be reduced by 12.8 times.The prediction time of 60452 samples took 5.6900 and 2.2516 s,respectively.Conclusion The model can guarantee the overall accuracy and performance of the model,and meet the requirement of low algorithm requirements for the deployment environment.

关 键 词：实时荧光PCR 时序分类 神经网络 知识蒸馏 

分 类 号：R197.39[医药卫生—卫生事业管理] TP391[医药卫生—公共卫生与预防医学]