胸部X线片人工智能联邦学习系统用于病原学诊断儿童社区获得性肺炎  

Artificial intelligence federated learning system based on chest X-ray films for pathogen diagnosis of community-acquired pneumonia in children

作　　者：魏子伊 汤奕 滕泽 李宏锋 彭芸[3] 操江峰 高天姿 张恒 韩鸿宾 WEI Ziyi;TANG Yi;TENG Z e;LI Hongfeng;PENG Yun;CAO Jiangfeng;GAO Tianzi;ZHANG Heng;HAN Hongbin(Institute of Medical Technology,Peking University Health Science Center,Beijing 100191,China;Department of Radiology,National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital,Chinese Academy of Medical Sciences and Peking Union Medical College,Beijing 100021,China;Department of Radiology,National Center for Children’s Health,Beijing Children’s Hospital,Capital Medical University,Beijing 100045,China;Institute of Information Engineering,Chinese Academy of Sciences,Beijing 100093,China;Departement of Biomedical Engineering,Chengde Medical University,Chengde 067000,China;Department of Radiology,Peking University Third Hospital,Beijing 100191,China;Beijing Key Laboratory of Magnetic Resonance Imaging Devices and Technology,Beijing 100191,China)

机构地区：[1]北京大学医学部医学技术研究院,北京100191 [2]国家癌症中心/国家肿瘤临床医学研究中心/中国医学科学院北京协和医学院肿瘤医院影像诊断科,北京100021 [3]国家儿童医学中心首都医科大学附属北京儿童医院影像中心,北京100045 [4]中国科学院信息工程研究所,北京100093 [5]承德医学院生物医学工程系,河北承德067000 [6]北京大学第三医院放射科,北京100191 [7]磁共振成像设备与技术北京市重点实验室,北京100191

出　　处：《中国介入影像与治疗学》2024年第6期368-373,共6页Chinese Journal of Interventional Imaging and Therapy

基　　金：国家自然科学基金(62301615)。

摘　　要：目的观察基于胸部X线片建立的人工智能联邦学习系统用于病原学诊断儿童社区获得性肺炎(CAP)的价值。方法回顾性选取2所医院共900例CAP患儿,包括细菌性、病毒性及支原体CAP各300例,对每例选取1幅胸部正位片。收集公开数据集GWCMCx中的5856幅儿童胸部正位片,分别来自4273例CAP患儿和1583例胸部无明显异常患儿。按8∶2比例将全部6756幅胸片分为训练集(n=5359)与验证集(n=1397)。建立基于注意力机制的病原学诊断儿童CAP模型,设计二分类及三分类诊断算法并进行联邦部署训练;与DenseNet模型对比,观察所获学习系统用于病原学诊断儿童CAP的效能。结果人工智能联邦学习系统模型针对全部数据诊断CAP的准确率为97.00%,曲线下面积(AUC)为0.990。基于来自医院的数据,本系统根据单一影像学数据及临床-影像学数据实现病原学诊断儿童CAP的AUC分别为0.858及0.836,均高于DenseNet模型的0.740(P均<0.05)。结论基于胸部X线片的人工智能联邦学习系统可用于病原学诊断儿童CAP。Objective To explore the value of artificial intelligence federated learning system based on chest X-ray films for pathogen diagnosis of community-acquired pneumonia(CAP)in children.Methods Totally 900 cases of CAP children from 2 hospitals were retrospectively enrolled,including bacterial,viral and mycoplasma CAP(each n=300),and chest posteroanterior X-ray films were collected.Meanwhile,chest posteroanterior X-ray films of 5856 children from the publicly available dataset GWCMCx were collected,including 4273 CAP images and 1583 healthy chest images.All above 6756 images were divided into training set(n=5359)and validation set(n=1397)at the ratio of 8∶2.Then a pathogen diagnosis model of children CAP was established based on attention mechanism.Binary and ternary diagnostic algorithms were designed,and federated deployment training was performed.The efficacy of this system for pathogen diagnosis of children CAP was analyzed and compared with DenseNet model.Results Based on all data,the accuracy of the obtained artificial intelligence federated learning system model for diagnosing children CAP was 97.00%,with the area under the curve(AUC)of 0.990.Based on hospital data,the AUC of this system using single imaging data and clinicalimaging data for pathogen diagnosis of children CAP was 0.858 and 0.836,respectively,both better than that of DenseNet model(0.740,both P<0.05).Conclusion The artificial intelligence federated learning system based on chest X-ray films could be used for pathogen diagnosis of children CAP.

关 键 词：肺炎 儿童 X线 人工智能 

分 类 号：R725.6[医药卫生—儿科] R814[医药卫生—临床医学]