深度学习模型检测胸部CT肺结节的临床效能评估  

Evaluation of the efficacy of deep learning model in detecting pulmonary nodules on chest CT images

作　　者：刘晶[1] 鲜军舫[1] 李书玲[1] 姜虹[1] 刘国顺 杜润爽 陈青华[1] LIU Jing;XIAN Junfang;LI Shuling;JIANG Hong;LIU Guoshun;DU Runshuang;CHEN Qinghua(Department of Radiology,Beijing Tongren Hospital,Capital Medical University,Beijing 100730,China;Department of Radiology,Guangzhou First People's Hospital,the Second Affiliated Hospital of South China University of Technology,Guangzhou510180,China)

机构地区：[1]首都医科大学附属北京同仁医院放射科,北京100730 [2]华南理工大学第二附属广州第一人民医院放射科,广东广州510180

出　　处：《实用放射学杂志》2021年第5期732-735,767,共5页Journal of Practical Radiology

摘　　要：目的评估基于深度学习(DL)的人工智能模型检测胸部CT肺结节的诊断效能.方法回顾性收集100例胸部薄层CT扫描图像,以3名具有15年以上胸部CT诊断经验的医师分别对结节标注的结果作为金标准,采用基于DL的人工智能模型(简称DL模型)进行肺结节检测,评估DL模型总体诊断效能及其临床应用的鲁棒性.由另外3名放射医师(初级医师2名和高年资医师1名)分别进行独立阅片及DL模型辅助诊断,对比研究DL模型对放射科医师的辅助诊断作用.结果DL模型对标注的323个结节(实性结节263个,亚实性结节40个,钙化结节20个)的检测灵敏度、阳性预测值和假阳性率分别为96.90%、0.495和3.19FPs/Scan.DL模型对不同品牌CT扫描仪及层厚(0.625~2mm)图像中肺结节的检测灵敏度、阳性预测值及假阳性率均无显著差异(P>0.05),对不同类型、不同大小肺结节(<5mm及≥5mm)检测的灵敏度和阳性预测值无显著差异(P>0.05),假阳性率有显著差异(P<0.05).与独立阅片相比,DL模型辅助3位放射医师阅片对肺结节检测的灵敏度升高(P<0.05),初级医师1、初级医师2、高年资医师耗时分别缩短11.4s、10.2s和21s.结论基于DL的人工智能辅助诊断系统对肺结节的检测不受CT设备及层厚(0.625~2mm)的影响,可帮助放射医师提高肺结节的检出率并缩短阅片时间.Objective To evaluate the diagnostic efficacy of the deep learning(DL)-based artificial intligence(AI)model in detecting pulmonary nodules on chest CT images.Methods The chest thin-slice CT images of 100 patients were collected retrospectively.The gold standard was made by 3 radiologists with more than 15 years of experience who analyzed and labeled all the pulmonary nodules on each sequence of image.The overall diagnostic efficacy and robustness of the DL-based Al assisted diagnostic model(DL-based model)in detecting pulmonary nodules were evaluated.To study the diagnostic efficacy of DL-based model,three other radiologists(two junior radiologists and one senior radiologist)respectively performed independent reading and DL.Abased model-assisted diagnosis.Results 323 nodules(2863 solid nodules,40 subsolid nodules and 20 calcification nodules)were labeled with the gold standard in all the patients.The sensitivity of nodules detection for DL-based model was 96.90%,the positive predictive value was 0.495,and the false positive rate was 3.19 FPs/Scan.,There were no significant differences in sensitivity,postive predictive value and false positive rateof the DIL-based model in detecting pulmonary nodules among the images of different CT manufacturers and slice thickness(0.625-2 mm)(P>0.05).No significant differences in the sensitivity and positive predictive value were found among different types and size of pulmonary nodules(<5 mm and≥5 mm)when used the Dl-based model(P>0.05),and the significant differences were found in false positive rate(P<0.05).The DL-based model assisted diagnosis had higher sensitivity in detecting pulmonary nodules compared to independentreading without DL-based model(P<O.05).The dliagnosis time for thejurnior radiologist 1,2 and the senior radiologist was reduced by 11.4 s,10.2 s and 2l s respectively.Conclusion The DL-based Al model independent of CT manufactures and slice thickness(0.625-2 mm),which can significantly improve the diagnostic efficacy of detecting pulmonary nodules and short

关 键 词：人工智能 深度学习 肺结节 检出 

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]