联邦学习技术提升医疗AI模型能力的可行性研究  

A Feasibility Study on Enhancing Medical AI Model Capabilities by Using Federated Learning Technology

作　　者：王栋 淡智利 马梁 WANG Dong;DAN Zhili;MA Liang(Shanghai Pudong Institute for Health Development,Shanghai 200129,China)

机构地区：[1]上海市浦东卫生发展研究院,上海市200129 [2]上海豪影云图科技有限公司,上海市201908

出　　处：《中国卫生信息管理杂志》2025年第2期275-281,共7页Chinese Journal of Health Informatics and Management

基　　金：上海市浦东新区卫生健康委2022年度卫生科技项目“医疗健康大数据分类分级使用标准的研究”(PW2022A-40)。

摘　　要：目的针对目前医疗行业凸显的数据孤岛难题,研究利用基于联邦学习技术的影像组学平台高效地训练泛化能力强的诊断模型的可行性。方法模拟两所医院在数据不出院的条件下,分别进行本地训练和联邦学习,对比本地训练获得的2个私有模型和联邦学习所得聚合模型的性能指标来检验模型泛化能力。结果在独立测试数据集上,聚合模型Mf(ACC=0.7364,AUC=0.8041),相比2个私有模型,Ma(ACC=0.7182,AUC=0.7984)和Mb(ACC=0.6818,AUC=0.7668)显示更好的性能。基于DeLong测试比较Mf与Ma、Mb没有显示显著性差异(P>0.05),进一步使用NRI方法作对比验证,Mf相对Ma和Mb有正向性能提升。结论结合影像组学和联邦学习技术,可以有效地解决当前医疗AI模型训练和部署过程中遇到的一系列问题,获得泛化能力更强的模型。Objective To address the challenge of data silos in the current medical industry,this research explores the feasibility of training a diagnostic model with strong generalization capabilities using a federated learning-based radiomics platform.Methods A simulation was conducted in which two hospitals,under conditions that preserve data privacy,engaged in local training and federated learning,respectively.The performance metrics of two private models obtained from local training and the aggregated model obtained from federated learning were compared to test the models’generalization capabilities.Results On an independent test dataset,the aggregated model Mf demonstrated better performance(ACC=0.7364,AUC=0.8041)compared to the two private models Ma(ACC=0.7182,AUC=0.7984)and Mb(ACC=0.6818,AUC=0.7668).DeLong’s test showed no significant difference between Mf and Ma or Mb(P>0.05).Further validation using the NRI method confirmed that Mf had a positive performance improvement over Ma and Mb.Conclusion The combination of radiomics and federated learning technologies can effectively solve a series of issues encountered in the training and deployment of current medical AI models,resulting in models with stronger generalization capabilities.

关 键 词：影像组学 联邦学习 区块链 可信模型 医疗AI 

分 类 号：R197.323[医药卫生—卫生事业管理] R319[医药卫生—公共卫生与预防医学]