人工智能在新型冠状病毒肺炎早期影像诊断中的应用  

Application of artificial intelligence in early imaging diagnosis of COVID-19

作　　者：陈辉 关春爽 闫铄 陈七一 张羽君 陈步东 谢汝明 CHEN Hui;GUAN Chunshuang;YAN Shuo;CHEN Qiyi;ZHANG Yujun;CHEN Budong;XIE Ruming(Department of Radiology,Beijing Ditan Hos pital,Capital Medical University,Beijing 100015,China;Shukun(Beijing)Technology Co.,Ltd.,Beijing 100102,China)

机构地区：[1]首都医科大学附属北京地坛医院放射科,北京100015 [2]数坤(北京)网络科技有限公司,北京100102

出　　处：《实用放射学杂志》2021年第4期558-562,共5页Journal of Practical Radiology

基　　金：北京市卫生健康科技成果与适宜技术推广专项项目(2020-TG-001).

摘　　要：目的 探讨人工智能(AI)辅助胸部CT在新型冠状病毒肺炎(COVID-19)早期定量诊断中的应用价值.方法 选取经病毒核酸检测确诊为COVID-19的患者60例,所有患者均为早期感染,在出现症状3 d内接受胸部CT平扫检查.定量分析患者胸部CT中不同密度范围内病变体积、病变平均密度、病灶/全肺体积和病灶/肺叶体积.由2位高年资放射诊断医师分别对病灶占每个肺叶的比例进行评估,分析二者评估的一致性,同时取2位诊断医师病灶/肺叶体积评估结果 的平均值,将平均值与AI测量的病灶/肺叶体积值进行比较.结果 胸部CT显示双肺未见明显异常14例(23.3%),累及2个及以上肺叶占65%.AI定量诊断结果 显示:在病变体积方面,左肺与右肺病变体积未见明显统计学差异[(37.94±74.15)cm^(3) vs(41.93±87.76)cm^(3),P=0.614],右肺中叶与右肺下叶病变体积有明显统计学差异[(4.41±9.10)cm^(3) vs(28.49±81.33)cm^(3),P=0.021];左肺下叶与右肺中叶病变体积有显著统计学差异[(29.09±67.15)cm^(3) vs(4.41±9.10)cm^(3),P=0.018].在病变密度方面,纯磨玻璃密度病变(-700~-450 HU)的体积及占比与其他不同密度病变(-450~-300 HU,-300~-100 HU,-100~0 HU,0~50 HU)比较均有明显统计学差异(P<0.05),而其他不同密度范围内病变(-450~-300 HU,-300~-100 HU,-100~0 HU,0~50 HU)体积均无明显统计学差异(P>0.05).2名放射诊断医师所测量的病灶/肺叶体积的平均值与AI软件测量值比较,各肺叶中病灶/肺叶体积值在两者间均未见明显统计学差异(P>0.05).结论 AI辅助胸部CT对COVID-19早期诊断与放射诊断医师相比较无明显差异.通过AI定量评价,发现COVID-19早期以双肺多发纯磨玻璃密度影(CT值-700~-450 HU)为主要表现,病变以双肺下叶为主.Objective To explore the application value of artificial intelligence(AI)assisted chest CT in early quantitative diagnosis of COVID-19.Methods 60 patients diagnosed as COVID-19 by virus nucleic acid test were collected.All patients had early infection and received chest CT scan within 3 days after symptoms appeared.Then the lesion volume,average lesion density,lesion/whole lung volume and lesion/lung lobe volume in different density ranges of patients'chest CT were quantitatively analyzed.T wo senior radiologists respectively evaluated the proportion of lesions in each lung lobe and analyzed the consistency of the two evaluations.At the same time,the average value of the lesion/lung lobe volume evaluation results of the two radiologists was taken,and the average value was compared with the lesion/lung lobe volume values measured by AL.Results In this study,14 patients(23.3%)showed normal chest CT,two or more lung lobes were involved in 65%patients.In terms of lesion volume,AI quantitative diagnosis results showed that there was no significant difference in lesion volume between the left and right lung[(37.94±74.15)cm^(3) vs(41.93±87.76)cm^(2),P=0.614].There was significant difference in the lesion volume between the right middle lobe and the right lower lobe[(4.41±9.10)cm^(2) vs(28.49±81.33)cm^(2),P=0.021];and there was significant difference in lesion volume between the left lower lobe and the right middle lobe[(29.09±67.15)cm^(2) vs(4.41±9.10)cm^(2),P=0.018].In terms of lesion density,the volume and proportion of pure ground glass opacities(-700-450 HU)were significantly different from those of other lesions with different densities(-450-300 HU,-300-100 HU,-100-0 HU and 0-50 HU)(P<0.05).While the volumes of other lesions(-450-300 HU,-300-100 HU,-100-0 HU and 0-50 HU)had no significant statistical difference between each other(P>0.05).The lesion/lung lobe volume values measured by the two radiologists were consistent.There was no significant difference between the average values of lesion/lung lobe volume

关 键 词：人工智能 新型冠状病毒肺炎 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] R563.1[自动化与计算机技术—控制科学与工程]