摘要
目的 开发基于深度卷积神经网络的肺结核病灶检测模型,并评估其在肺结核大规模人群筛查及临床检测中的应用价值。方法 回顾性收集2019年3月至2020年7月于喀什地区第一人民医院影像中心就诊的1217例患者的影像数据,随机分为3个数据集,以7∶2∶1的比例在改进的RetinaNet肺结核病灶检测模型上进行训练、验证和测试,并收集两个肺结核公开数据集数据共800例,用于模型的外部验证。检测模型通过构造针对小病灶敏感的损失函数,引入注意力机制和多尺度特征提取等技巧,优化对微小病灶和隐匿性病灶的检出率。结果 改进的RetinaNet模型仅在测试集的曲线下面积(Area Under the Cure,AUC)略低于原始RetinaNet模型,其他数据集的AUC和准确度均高于原始RetinaNet模型。同时改进的RetinaNet模型在外部中心的公开数据集进行模型评价时,诊断性能较测试集和验证集表现更好(AUC为0.879,准确度为0.847)。放射科医生在人工智能(Artificial Intelligence,AI)系统辅助下对肺结核病进行诊断时的灵敏度、特异性、准确度较无AI系统辅助的诊断水平均有明显提升。在有AI系统辅助下放射科医生对于病例的影像数据进行阅片时间显著短于无AI系统辅助时(P<0.001)。结论 深度学习能用于快速检测和定位胸片中的肺结核病灶,并给出相应的置信指数和病灶位置信息,可大批量筛查肺结核高风险人群,大幅度地提高医疗资源匮乏地区放射科医生的工作效率和肺结核诊断的准确度。
Objective To develop a pulmonary tuberculosis lesion detection model based on deep convolutional neural network,and evaluate its application value in mass population screening and clinical detection of pulmonary tuberculosis.Methods In this study,the image data of 1217 patients in the Image Center of the First People’s Hospital of Kashi from March 2019 to July 2020 were retrospectively collected and randomly divided into 3 datasets for training,validation and testing on the improved tuberculosis lesion detection model of RetinaNet at a ratio of 7∶2∶1.Two public data sets of tuberculosis with a total of 800 cases were collected for external validation of the model.By constructing a loss function sensitive to small nidus,and introducing techniques such as attention mechanism and multi-scale feature extraction in this detection model,the detection rate of small or hidden nidi was optimized.Results The area under curve(AUC)of the improved RetinaNet model was slightly lower than the original RetinaNet model only in the test set,and the AUC and accuracy of other data sets were higher than the original RetinaNet model.At the same time,the improved RetinaNet model performed better diagnostic performance than the test and validation sets(AUC was 0.879,accuracy was 0.847)in the model evaluation of the public data set in the external center.The sensitivity,specificity and accuracy of the diagnosis of tuberculosis by radiologists with the aid of artificial intelligence(AI)system are significantly improved compared with that without the aid of AI system.With the aid of AI system,the time for radiologists to read the image data of the cases was significantly shorter than that without the aid of AI system(P<0.001).Conclusion Deep learning can be used to rapidly detect and locate tuberculosis lesions in chest radiographs,and provide corresponding confidence index and lesion location information,which can screen high-risk groups of tuberculosis in large quantities,and greatly improve the work efficiency of radiologists in a
作者
马依迪丽·尼加提
田序伟
米日古丽·达毛拉
阿布都克尤木·阿布力孜
阿里木江·阿卜杜凯尤木
戴国朝
董家科
MAYIDILI-Nijiati;TIAN Xuwei;MIRIGULI-Damaola;ABUDUKEYOUMU-Abulizi;ALIMUJIANG-Abudukaiyoumu;DAI Guochao;DONG Jiake(Image Center,First People’s Hospital of Kashi,Kashi Xinjiang 844000,China)
出处
《中国医疗设备》
2023年第10期7-13,共7页
China Medical Devices
基金
省部共建中亚高发病成因与防治国家重点实验室开放课题项目(SKL-HIDCA-2020-KS3,SKL-HIDCA-2021-JH6)
天山创新团队项目(2022D14007)。
关键词
人工智能
胸部X线检查
深度卷积神经网络
结核病诊断
医学影像
artificial intelligence
chest X-ray examination
deep convolutional neural network
diagnosis of tuberculosis
medical imaging
