首页|基于深度学习构建金黄色葡萄球菌和粪肠球菌的快速图像识别系统

基于深度学习构建金黄色葡萄球菌和粪肠球菌的快速图像识别系统

扫码查看
原文链接
  • 万方数据
  • 维普
目的 基于GoogleNet、ResNet101和Vgg19 3种深度学习模型,对血流感染病原菌(金黄色葡萄球菌、粪肠球菌)进行高置信度的识别,比较模型间的性能与分类能力,探讨深度学习模型对血流感染病原菌快速识别的应用可行性.方法 将革兰染色及摄像预处理后的细菌图像和空白对照图像输入模型,进行训练与验证,共采集1 682张金黄色葡萄球菌、1 723张粪肠球菌和688张空白对照显微图像,对其中1 344张金黄色葡萄球菌、1 376张粪肠球菌和544张空白对照图像进行训练,余下的图像用于验证.根据模型间的分类参数评估出性能最佳的模型.结果 ResNet101模型识别三类验证集图像的交叉熵损失值(0.008 710 3)最低,Epoch值(93)最大且准确率(99％)最高;GoogleNet模型识别三类验证集图像的交叉熵损失值为0.063 89,Epoch值为86,准确率为98.6％;Vgg19模型识别三类验证集图像的交叉熵损失值为0.035 682,Epoch值为86,准确率为97.7％.结论 ResNet101模型在对三类验证集图像的分类上性能最佳;深度学习模型可对金黄色葡萄球菌和粪肠球菌的革兰染色图像进行准确、可信的快速识别.
Construction of a rapid image recognition system for Staphylococcus aureus and Enterococcus faecalis based on deep learning
Objective To identify the pathogenic bacteria such as Staphylococcus aureus and Enterococcus faecalis in bloodstream infec-tions with high confidence based on three deep learning models such as GoogleNet,ResNet101,and Vgg19,compare the performance and classification ability of these models,and explore the feasibility of applying the deep learning models for the rapid identification of pathogenic bacteria in bloodstream infections.Methods The preprocessed Gram-stained bacterial images,including 1 682 images for Staphylococcus aureus and 1 723 for Enterococcus faecalis,and 688 blank control microscopic images were input into three models for training and validation,respectively.Among them,1 344 images for Staphylococcus aureus,1 376 for Enterococcus faecalis,and 544 blank control images were used for training,and the remaining images were used for validation.The model with the best performance was identified according to the classification parameters between the models.Results The ResNet101 model had the lowest cross-en-tropy loss value(0.008 710 3),the largest Epoch value(93),and the highest accuracy rate(99％)for identifying the three types of validation set images.The cross-entropy loss value,Epoch value,and accuracy rate of the GoogleNet model were 0.063 89,86 and 98.6％,respectively,for identifying the three types of validation set images.Those of the Vgg19 model were 0.035 682,86 and 97.7％,respectively.Conclusion The ResNet101 model has the best performance in the classification of three kinds of images.The deep learning model may accurately,reliably and rapidly identify the Gram-stained images of pathogenic bacteria such as Staphylococcus aureus and Enterococcus faecalis in bloodstream infections.

deep learningGram stainingStaphylococcus aureusEnterococcus faecalisbloodstream infectionrapid identification

罗远美、陈轲玮、李振彰、岳雨彪、陈凌娟、刘加伟、李齐光、李杨、徐令清

展开 >

广州医科大学附属清远医院/清远市人民医院检验医学部,广东清远 511518

广州医科大学生物医学工程学院,广州 511436

深度学习 革兰染色 金黄色葡萄球菌 粪肠球菌 血流感染 快速识别

广东省医学科学技术研究基金清远市科技计划基金清远市科技计划基金清远市科技计划基金清远市人民医院医学科研基金清远市人民医院医学科研基金清远市人民医院医学科研基金

A2021490221104197683799200808114560452202280811456047120190209202301-201202301-318

2024

10.13602/j.cnki.jcls.2024.07.01

临床检验杂志
江苏省医学会

临床检验杂志

CSTPCD
影响因子：0.746
ISSN：1001-764X
年,卷(期)：2024.42(7)
  • 7