基于深度学习的胸部X线肺结核检测研究及多中心临床验证  

Development of a deep learning-based algorithm for pulmonary tuberculosis detection on chest radiograph:a multi-center study

作　　者：安超 张晨 郑广平[1] 曹义[1] 杨根东[1] 印宏坤 顾俊 邹彤 吴双 王立非[1] An Chao;ZHANG Chen;ZHENG Guang-pin(Department of Radiology,the Third People's Hospital of Shenzhen,Guangdong 518100,China)

机构地区：[1]深圳市第三人民医院放射科,广东518100 [2]广东省感染性疾病(结核病)临床医学研究中心,广东518100 [3]推想医疗科技股份有限公司,北京100025

出　　处：《放射学实践》2022年第6期704-709,共6页Radiologic Practice

基　　金：广东省科技计划项目(2020B1111170014)。

摘　　要：目的:构建基于深度学习的胸部X线肺结核检测模型并通过多中心研究验证其效能及临床价值。方法:回顾性搜集2600例来自3个中心的胸部X线图像并随机分为训练集、验证集和测试集,构建基于RetinaNet架构的肺结核深度学习检测模型,并在ChinaSet和MontgomerySet胸部X线公开数据集以及来自深圳三院的外部临床测试集上对深度学习模型的鲁棒性进行外部测试。采用受试者工作特征曲线(receiver operator characteristic curve,ROC)评估模型效能。同时通过临床检测评估深度学习模型的重复性和再现性。结果:深度学习模型在内部测试集的ROC曲线下面积(AUC)为0.967,在ChinaSet、MontgomerySet和深圳三院外部测试集的AUC分别为0.95、0.93和0.976,具有较高的准确性和良好的鲁棒性。临床一致性评估证实了模型的重复性和再现性。结论:深度学习模型具备良好的效能,可以作为胸部X线影像结核病检测工具用于临床决策支持。Objective:To develop a deep learning model for detecting tuberculosis on chest radiographs and evaluate its efficacy and clinical value on multi-center datasets.Methods:A total of 2600 chest radiographs from three hospitals were collected retrospectively,and a RetinaNet architecture-based convolutional neural network was proposed for the detection of tuberculosis.The chest radiographs were randomly divided into training,validation and internal testing datasets.Two public datasets(ChinaSetand and MontgomerySet)and an independent dataset from The Third People's Hospital of Shenzhen were used for external testing.The repeatability and reproducibility of the deep learning model was confirmed through clinical validation.Results:The deep learning model had an AUC of 0.967 in the internal testing dataset,and showed good robustness with AUCs achieving 0.95,0.93 and 0.976 in the ChinaSet,MontgomerySet and the independent testing dataset from The Third People's Hospital of Shenzhen.The clinical study demonstrated excellent repeatability and reproducibility of the deep learning model.Conclusion:The deep learning model showed good performance in detecting tuberculosis on chest radiographs,which could be used as a potential tool for assisting clinical decision.

关 键 词：深度学习 放射摄影术 胸部 结核 肺 多中心研究 

分 类 号：R-056[医药卫生] R816.4