摘要
目的:基于临床验证性研究,探索后疫情时代人工智能肺炎辅助诊断系统(AI-ADS)潜在的临床应用场景.方法:回顾性收集了来自三家医院的1049例胸部CT扫描数据,包括400例胸部CT表现正常的病例、233例新冠肺炎病例和416例其他社区获得性肺炎病例.六名高年资放射科医师参与了数据标注工作.采用敏感度、特异度、Dice系数和受试者操作特征(ROC)曲线下面积(AUC)评估人工智能系统在相应场景中的性能表现.结果:AI-ADS基于胸部CT识别各类型肺炎、细菌性肺炎、新冠肺炎、其他病毒性肺炎和其他社区获得性肺炎的AUC分别为0.968、0.983、0.992、0.941、0.958,检测各种肺炎的敏感度均超过0.90;鉴别病毒性肺炎和非病毒性肺炎的AUC达到0.950,敏感度为0.885,特异度为0.910;在新冠肺炎和其他社区获得性肺炎测试集中分割肺炎区域的平均Dice系数分别达到0.851和0.753.结论:AI-ADS在肺炎的检测预警、病灶定量分析以及鉴别诊断方面具有良好的性能,具备了后疫情时代的多场景应用价值.
Abstract
Objective:To investigate the potential clinical practice of artificial intelligence(AI)-assisted diagnostic system(AI-ADS)for pneumonia in the post-pandemic era by exploring various ap-plication scenarios in different patient cohorts.Methods:The study retrospectively collected 1049 sets of chest CT scans from patients either diagnosed as normal(n=400),COVID-19(n=233),or other community-acquired pneumonia(CAP)(n=416)at three hospitals.We explored the potential clinical practice by validating its performance in the detection,classification,and lesion measurement(segmen-tation)of different types of pneumonia.Six senior radiologists participated in the establishment of the gold standard for lesion labeling and segmentation.Sensitivity,specificity,Dice coefficient,and area un-der the receiver operating characteristic curve(AUC)were utilized to evaluate the performance.Results:AI-ADS displayed decent detection performance on different types of pneumonia,as evidenced by the AUC of 0.968,983,0.992,0.941,and 0.958 for overall types,bacterial,COVID-19,non-COVID viral,and other pneumonia,respectively.The detection sensitivity all reached above 0.9.Additionally,the system differentiated viral and non-viral pneumonia with an AUC of 0.950,a sensitivity of 0.885,and a specificity of 0.910.Of note,AI-ADS achieved good segmentation results on both COVID-19 ca-ses(internal test set,averaged DICE=0.851)and non-COVID cases(external test set,averaged DICE=0.753).Conclusion:With performance improvement,Al-ADS can detect various types of pneumonia and differentiate viral pneumonia from others.It shows a decent lesion segmentation capacity among different types of pneumonia,indicating its potential clinical application in the post-pandemic era.
基金项目
科技创新2030——"新一代人工智能"重大项目(2021ZD0111104)
NETL
NSTL
万方数据