人工智能辅助阅片模式对血管粘连型肺结节在低剂量胸部CT检出效能的影响  

Impact of a deep learning based artificial intellegence aided screening system for juxta-vascular pulmonary nodules detection in chest low-dose CT

作　　者：汪芳[1] 杨利莉[1] 许菲菲 顾俊 石骁萌 赵艳红[1] 哈若水[1] 沈云 WANG Fang;YANG Lili;XU Feifei;GU Jun;SHI Xiaomeng;ZHAO Yanhong;HA Ruoshui;SHEN Yun(Medical Imaging Center,Ningxia Hui Autonomous Region People’s Hos pital(the First Affiliated Hos pital of Northwest University for Nationalities),Yinchuan 750002,China;IMAGI NGRACE Medical Imaging Diagnostic Center,Yinchuan 750000,China;Beijing Infervision,Beijing 100000,China;CT Research Centre,GE Healthcare China,Beijing 100000,China)

机构地区：[1]宁夏回族自治区人民医院(西北民族大学第一附属医院)医学影像中心,宁夏银川750002 [2]影和医学影像诊断中心,宁夏银川750000 [3]北京推想科技有限公司,北京100000 [4]GE医疗中国GECT影像研究中心,北京100000

出　　处：《实用放射学杂志》2022年第2期213-216,231,共5页Journal of Practical Radiology

基　　金：2019年中央高校基本科研业务费项目(31920190183).

摘　　要：目的探讨基于深度学习人工智能(DL-AI)系统对辅助低年资医师在低剂量CT(LDCT)肺癌筛查中对血管粘连型肺结节(JVPN)检出效能的影响.方法选取接受256排宽体胸部低剂量成像早期肺癌筛查受检者104例.2名低年资影像科医师在传统阅片模式(方法A)和DL-AI系统辅助诊断阅片模式(方法B)下进行JVPN检测.记录每种方法检出的每个患者结节总数、结节的位置及大小.最终以2名高年资胸部影像诊断医师共同确认的结节作为金标准.分别计算2种检测模式对JVPN的检出率、漏诊率.利用χ2检验比较2种方法对JVPN检出率的差异;比较不同部位、不同大小JVPN的检测能力.结果共检出肺结节387个,其中JVPN 216个.方法B检测出真性JVPN 208个,检出率96.29%(208/216),漏诊率3.71%(8/216),方法A检出真性JVPN 156个,检出率72.22%(156/216),漏诊率27.78%(60/216),方法B的检出率明显高于方法A的检出率(96.29%vs 72.22%;χ2=47.193,P<0.001),低年资医师应用方法B多检出52个真结节,其在不同部位、不同大小结节的检出率均高于方法A(P<0.001).结论在肺癌早期低剂量筛查患者中,DL-AI系统显著提高低年资影像科医师对JVPN诊断敏感性;减少传统模式阅片的漏诊率.Objective To investigate the impact of deep learning-artificial intelligence(DL-AI)system on the effectiveness of assisting junior radiologists in the detecting of juxta-vascular pulmonary nodules(JVPN)of low-dose computed tomography(LDCT)lung cancer screening.Methods A total of 104 patients were enrolled for early lung cancer screening using 256-row wide detector low-dose chest scan.The CT images were interpreted by two junior radiologists in the traditional reading mode(method A)and DL-AI system aided reading mode(method B)for JVPN examination.The total number,location and size of detected JVPN were recorded.The reference standard was confirmed by the consensus readings of two senior radiologists.The detection rate and false negative rate in each reading mode were calculated accordingly.χ2 test was performed to test the difference between the detection rate of two methods;detection capabilities of JVPN with different sizes and in different locations were also compared.Results A total of 387 pulmonary nodules were detected,of which 216 were JVPN.Method B detected 208 authentic JVPN,the detection rate was 96.29%(208/216),and the false negative rate was 3.71%(8/216).Method A detected 156 authentic JVPN,the detection rate was 72.22%(156/216),and the false negative rate was 27.78%(60/216).The detection rate of method B was significantly higher than that of method A(96.29%vs 72.22%;χ2=47.193,P<0.001),junior doctor used method B to detect 52 more true nodules than method A.Also,the detection rates of nodules in different locations and sizes were higher than those of method A(P<0.001).Conclusion In patients with early low-dose CT screening for lung cancer,the DL-AI system sig-nificantly improves the sensitivity of junior imaging physicians to the diagnosis of JVPN;reduces the false negative rate compare to traditional mode reading.

关 键 词：深度学习 人工智能 血管粘连型肺结节 低剂量 计算机体层成像 

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]