人工智能辅助低、高年资规培医师对CT肺结节检测效能的对比研究

目的:探讨不同年资规培医师在CT独立阅片及应用人工智能辅助阅片模式下对肺内结节的检测效能是否有差异,并分析低年资规培医师在应用人工智能辅助阅片后,能否达到高年资规培医师对肺结节的检测水平.方法:收集180例患者的胸部CT图像,分别由2名低年资、2名高年资规培医师进行独立阅片并标注肺结节,经过2周洗脱期后,再次对相同病例...

A comparative study about the detection capability of lung nodules on CT images based on AI-assisted software among primary and senior interns

Objective:To evaluate whether there is a difference in the detection capability of lung nodules by intern with different seniority under the mode of independent reading and the mode of using artificial intelligence(AI)assisted reading,and to analyze whether the detection capability of lung nodules by primary interns under AI-assisted software can reach the level by senior interns.Methods:The chest CT images of 180 patients were collected.Two primary interns and two senior interns read the images without AI software and the same intern read CT images with AI software after two weeks’washout period.The location,size,density of the nodules and the reading time were recorded.The labeling results of each group were compared with the gold standard,the number of true-positive nodules,false-positive nodules and false-negative nodules were recored,and the detection sensitivity and the number of false positive nodules per capita were calculated.The Kappa test was used to analyze the consistency of the detection of lung nodules by two primary interns and two senior interns,and the Wilcoxon rank test(paired samples)was used to compare the differences in sensitivity,the number of false positive nodules per capita and reading time between the two groups.Results:The primary interns used AI reading mode to detect 358 lung nodules,and the accuracy improved(77.07%vs.47.43%),and the number of false positive nodules per capita decreased(0.90 vs.1.48),the labeling time was shortened(361.07±163.07)s vs.(429.11±132.61)s],and there were statistically significant(all P<0.01).The senior interns used AI reading mode to detect 372 lung nodules,and the accuracy improved(81.29%vs.50.50%),and the number of false positive nodules per capita decreased(0.27 vs.0.48),and there were significant differences(all P<0.01),but the labeling time extended(380.16±135.40)s vs.(367.22±120.52)s],and there was not considered statistically significant(P>0.05).The primary interns used AI reading mode to detect 321 nodules more than the senior interns independent reading mode,and the accuracy improved(77.07%vs.50.50%),but the number of false positive nodules per capita increased(0.90 vs.0.48),and the differences were significant(both P<0.01);the labeling time shorten(361.07±163.07)s vs.(367.22±120.52)s],and there was not considered statistically significant(P>0.05).Conclusions:The AI assistant software has improved the detection efficiency of the lung nodules of the primary interns and senior interns,and greatly shortened the reading time of the primary interns.With the help of AI software,the number of misdiagnosed nodules of primary interns is slightly more than that of the senior interns.However,the number of true nodules detected by the former far exceeds the latter.