摘要
目的评价人工智能(AI)软件在胸部CT肺结节检出中的应用价值。方法随机选取2018年8月至10月广州医科大学附属第一医院健康体检人群218例的薄层胸部CT检查资料,由低年资、高年资放射医师及AI软件分别采用低敏和高敏算法分别进行阅片,检测肺结节的位置分布、密度、大小等特征,以2名高级医师在低年资医师及AI的基础上进行诊断的结果为金标准,评估各组的肺结节检测效能。结果共纳入218例胸部CT资料,金标准判定阳性(至少检出1个结节)176例、阳性结节619个。结节在肺叶及肺段的分布为右上叶170个、右中叶60个、右下叶122个、左上叶145个、左下叶122个;结节的密度分布为实性结节471个、钙化结节71个、磨玻璃结节65个、部分实性结节12个;直径大于4 mm的结节231个、小于等于4 mm的结节388个。低敏、高敏算法的AI检出结节的灵敏度(56.70%、78.84%)接近或高于高年资医师(56.38%),显著高于低年资医师(34.25%),低年资医师结合AI辅诊后灵敏度(49.92%)显著提升。AI总体假阳性率较高,低敏AI组和高敏AI组每例CT平均假阳性结节数为0.70个、3.79个。AI对结节的肺段定位准确率较低,均不足80%;而医师诊断结节密度准确率(均>90%)均高于AI(均<90%)。对结节直径的测量,低年资医师与金标准比较差异显著(t=2.73,P=0.007),其余各组与金标准差异均无统计学意义(均P>0.05)。低年资医师结合AI辅诊可以明显加快阅片速度,总阅片时间从12.5 h降至9.9 h。结论AI软件可辅助放射医师检出肺结节,提高诊断的灵敏度和效率。
Objective To evaluate the value of artificial intelligence(AI)software in detecting pulmonary nodules on chest CT.Methods Thin-slice chest CT imaging data were randomly selected from 218 health check-up subjects registered to First Affiliated Hospital of Guangzhou Medical University between August and October,2018.The images were read by junior and senior radiologists,and also by AI softwares with low-and high-sensitivity algorithms,respectively,to determine the patterns of pulmonary nodules including location,density,and size.By using the final diagnosis by two senior radiologists who reviewed the impressions by junior radiologists and AI softwares as the gold standard,we evaluated the detection yield of pulmonary nodules in each group.Results Among the 218 cases of chest CT data included,176 cases were positive with at least one nodule detected each and with 619 nodules detected in total,according to gold standard.By lobar/segmental distribution of nodules in the lung,there were 170 nodules in the right upper lobe,60 in the right middle lobe,122 in the right lower lobe,145 in the left upper lobe,and 122 in the left lower lobe.By density,there were 471 solid,71 calcified,65 ground-glass,and 12 part-solid nodules.There were 231 nodules with diameter>4 mm,and 388 with diameter≤4 mm.The sensitivity of AI software with either low-sensitivity(56.70%)or high-sensitivity algorithm(78.84%)in detecting nodules was close to or higher than that by senior radiologists(56.38%),and was significantly higher than that by junior radiologists(34.25%).The diagnostic sensitivity by junior radiologists was significantly improved(49.92%)with AI assistance.The overall false-positivity of AI was relatively high.The mean number of false-positive nodules per CT case in the low-sensitivity AI group vs high-sensitivity AI group was 0.70 vs 3.79.AI resulted in low accuracy in determining the location of nodules in lung segments(<80%in all cases).The radiologists yielded higher accuracy in determining nodule density(both>90%)compared with
作者
刘艳雯
曾庆思
邓宇
沈伟
梁锐烘
Liu Yanwen;Zeng Qingsi;Deng Yu;Shen Wei;Liang Ruihong(Department of Radiology,First Affiliated Hospital of Guangzhou University of Chinese Medicine,Guangzhou 510405,China;Department of Radiology,First Affiliated Hospital of Guangzhou Medical University,Guangzhou 510120,China;Yitu Healthcare,Hangzhou 310012,China;Department of Radiology,NanFang Hospital of Southern Medical University,Guangzhou 510000,China)
出处
《中华生物医学工程杂志》
CAS
2022年第6期638-644,共7页
Chinese Journal of Biomedical Engineering
基金
呼吸疾病国家重点实验室开放课题(SKLRD-OP-201906)
2021年度广东省卫生健康适宜技术推广项目
广州医科大学附属第一医院2018年度成果培育和临床转化项目。
关键词
人工智能
体层摄影术
X线计算机
肺结节
诊断
计算机辅助
Artificial intelligence
Tomography,X-ray computed
Pulmonary nodule
Diagnosis,computer-assisted
