摘要
目的:探讨在人工智能(AI)肺结节检测软件的辅助下能否提升疲劳状态的放射科规培医师对肺结节的检测效能。方法:搜集182例患者的1 mm薄层胸部CT图像,有一位放射科规培医师分别在3种模式下进行阅片:正常状态下独立阅片(A组)、疲劳状态下(即一天日常工作满8小时以上)独立阅片(B组)、疲劳状态下使用AI软件辅助阅片(C组),三种阅片模式均间隔洗脱期(2周),分别记录每次阅片时检出结节的位置、大小和数目。将3次肺结节检出结果与金标准(由2位从事胸部影像诊断超过8年的中级医师结合AI筛查结果分别作出诊断,再由1位从事胸部影像诊断超过15年的高级医师最终审核确定)进行比较,计算敏感度和(患者)人均假阳性(误诊)结节数来评价3种模式的检测效能。结果:经金标准确认1281个肺结节,A组检出真阳性结节592个、假阳结节297个,敏感度46.21%,人均误诊结节数为1.63;B组检出真阳性结节517个、假阳结节225个,敏感度40.36%,人均误诊结节数为1.24;C组检出真阳性结节995个、假阳结节165个,敏感度77.67%,人均误诊结节数为0.91。B组的敏感度和人均误诊结节数均较A组降低,差异均有统计学意义(P<0.05);C组的敏感度较B组提高,且人均误诊结节数降低,差异均有统计学意义(P<0.05);C组的敏感度较A组提高,人均误诊结节数降低,差异均有统计学意义(P<0.05)。结论:疲劳显著降低了放射科规培医师对肺结节的检测效能,但在AI软件辅助下能明显提高疲劳状态下放射科规培医师对肺结节的检出效能,甚至超过其正常状态下的水平。
Objective:The purpose of this study was to investigate whether the assistance of artificial intelligence(AI)-assisted pulmonary nodule detection software could improve the detection efficiency of pulmonary nodules in fatigued radiological residents with standardized training.Methods:A total of 182 patients who underwent chest CT scans was collected.Three readings methods were conducted by a standardization training intern with 4 months chest CT experience:the resident in daily work condition(not fatigued)without AI software(group A),the same fatigued resident(after 8 hours working)without AI-software(group B),and the fatigued resident with AI software(group C).Each of the three readings had a washout period of more than 15 days.The location,size and number of nodules in the three groups were recorded and compared with the gold standard(the diagnosis was made by two intermediate physicians who had been engaged in chest imaging diagnosis for more than 8 years in combination with AI screening results,and was finally confirmed by a senior physician who had been engaged in chest imaging diagnosis for more than 15 years),the number of true positive and false-positive nodules was calculated.The performance of three reading methods regarding lung nodule detection was assessed by sensitivity and the number of false positive nodules(error diagnosis)per patient(ED/P).Results:A total of 1281 nodules was confirmed to be gold standard nodules.Group A detected 592 true positive nodules and 297 false-positive nodules with a sensitivity of 46.21%,and the number of false positive nodules was 1.63 per CT scan.Group B detected 517 true positive nodules and 225 false-positive nodules with a sensitivity of 40.36%,and ED/P was 1.24.Group C detected 995 true positive nodules and 167 false-positive nodules with a sensitivity of 77.67%,and ED/P was 0.91.The sensitivity and ED/P in group B were lower than those in group A,and the differences were statistically significant(P<0.05);the sensitivity of group C was higher than that of group B,and ED/P was reduced,and the differences were both statistically significant(P<0.05);the sensitivity of group C was higher than that of group A,ED/P was reduced,and the differences were statistically significant(P<0.05).Conclusion:Fatigue can reduce the detection efficiency of pulmonary nodules for radiologists.However,fatigued intern aided with an AI-assisted lung nodule detection system could achieve a higher detection rate for pulmonary nodules,even exceeding the results obtained in their routine wor-king state without fatigue.
作者
王亮
许迪
孙丹丹
顾俊
伍建林
于晶
WANG Liang;XU Di;SUN Dan-dan(Department of Radiology,Zhongshan Hospital,Dalian University,Dalian 116001,China)
出处
《放射学实践》
CSCD
北大核心
2021年第4期475-479,共5页
Radiologic Practice
关键词
肺结节
人工智能
计算机辅助检测
疲劳
体层摄影术
X线计算机
Artificial intelligence
Computed aided diagnosis
Pulmonary nodules
Fatigue
Tomography,X-ray computed
