人工智能与人工阅片不同联合方法在肺结节CT筛查中的比较  

The comparison of different combination methods of artificial intelligence and manual reading in CT screening of pulmonary nodule

作　　者：马宁强 赵子光 樊玮 王慧 王凯 范晓娥 齐敏[3] MA Ningqiang;ZHAO Ziguang;FAN Wei;WANG Hui;WANG Kai;FAN Xiao'e;QI Min(Imaging Center,Yangling Demonstration Zone Hospital,Yangling,Shaanxi Province 712100,China;Sigma Medical Technologies(Beijing)Co.Ltd,Beijing 100102,China;Imaging Center,Shaanxri Provincial People's Hospital,Xi'an 710068,China)

机构地区：[1]杨凌示范区医院影像中心,陕西杨凌712100 [2]图玛深维医疗科技(北京)有限公司,北京100102 [3]陕西省人民医院影像中心,陕西西安710068

出　　处：《实用放射学杂志》2020年第8期1317-1321,共5页Journal of Practical Radiology

基　　金：陕西省重点研发计划项目(2017SF-023).

摘　　要：目的探讨人工智能(AI)与人工阅片不同联合方法对胸部CT阅片时间及肺结节检出效能的影响.方法纳入肺结节CT筛查患者200例,共包含1836个肺结节,分别使用共同阅片(CR)(阅片时AI结果同时显示)、第2阅片(SR)(不知道AI结果的情况下先阅片)及人工双阅(DR)(住院医师阅片,主治医师审核)3种方法阅片,记录每例阅片时间、检出肺结节及其特征.以2名副主任医师一致意见为参考标准,分别计算出3种方法阅片时间、肺结节检出的敏感性、阳性预测值及假阳性率,阅片时间两两比较采用q检验,肺结节检出敏感性、阳性预测值及特征两两比较采用χ^2检验,假阳性率采用配对t检验.结果CR法平均每例阅片时间(184.45 s±91.21 s)明显短于SR法(345.72 s±130.71 s)(P<0.05)和DR法(522.88 s±130.08 s)(P<0.05).CR法的肺结节检出敏感性(95.48%)明显高于DR法(82.41%)(P<0.05),与SR法(96.57%)之间差异无统计学意义(P>0.05).CR法的肺结节阳性预测值和假阳性率(97.61%,0.22/CT)与SR法(97.74%,0.21/CT)、DR法(97.68%,0.18/CT),之间差异无统计学意义(P>0.05).结论在肺结节CT筛查中,使用AI共同阅片法,节省了人力和时间成本,相对于DR法又提高了肺结节检出的敏感性,便于临床推广使用.Objective To explore the efect of different combination methods of artificial itelligence(AI)and manual reading on the reading time and eficiency of deecting pulmonary nodules on chest CT.Methods Concurent reader(CR),scend reader(SR)and double reader(DR)were employed to analyze the CT reports of 1836 pulmonary nodules from 200 patients,by recording the reading time,results,and characteristics of positive dection.The agreement from two different vice directors was taken as a reference.The reading time,sensitivity of detetion,positive predictive value and the false positive rate were evaluated.The significant difference of the reading time among different methods was analyzed by q test.The significant comparisons of the sensitivity,positive predictive value,and characteristics among different methods were analyzed by chi-square test.The significant difference of the false positive rate among different methods was analyzed by t test.Results The mean reading time of CR(184.45 s±91.21 s)was significantly shorter than that of SR(345.72 s±130.71 s)(P<0.05)and DR(522.88 s±130.08 s)(P<0.05).The detecting sensitivity of pulmonary nodules with CR(95.48%)was higher than that with DR(82.41%)(P<0.05),whereas there was no significant difference between CR and SR(96.57%)(P>0.05).There were also no significant dfferences(P>0.05)in the positive predictire value of pulmonary nodules and false positive ate among the CR(97.61%,0.22/CT),SR(97.74%.0.21/CT)and DR(97.68%,0.18/CT).Condusion In CT screning of pulmonary nodule,Al concurent reader method could not only save time and cost,but also improve the detecting sensitivity of pulmonary nodules relative to DR,providing a more effective approach to monitor the pulmonary nodules in clinical practice.

关 键 词：人工智能 联合方法 肺结节 肺癌 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R563[自动化与计算机技术—控制科学与工程]