人工智能在体检肺CT中检出的假阳性结节研究

Study of False Positive Nodules Detected by Artificial Intelligence in Lung CT Examination

目的研究人工智能(Artificial Intelligence,AI)在体检肺CT中假阳性结节的分布与成因。方法连续收集2020年8至9月在我院做肺CT的体检者500例,所有肺CT图像经过AI软件进行检测,检出的肺结节由两名医师进行真假结节的认定,并记录假结节的数量、大小和密度。结果AI检出肺结节1518个,其中真结节740个,假结节778个,平均每例CT检出1.6个假结节。不同大小假结节中,以<5 mm假结节最常见,占68.6%(534/778)。不同密度假结节中,部分实性假结节的假阳性预测值为69.7%(23/33),高于实性假结节的48.3%(353/731)和纯磨玻璃密度假结节的53.3%(402/754),差异有统计学意义(P<0.05)。年龄与假结节的检出率有明显的正相关关系,Spearman等级相关系数rs=0.986,P<0.05。AI检出假阳性肺结节的主要成因有胸膜结节(21.5%)、索条影(17.9%)、血管增粗(13.8%)、血管分叉(12.2%)和肺小叶结构(9.0%)。结论AI检出的假阳性肺结节有一定的分布与构成规律,熟悉这些规律,有利于提高诊断医生识别假阳性结节的能力,并希望能为AI降低假阳性率提供参考。

Objective To study the distribution and causes of false positive nodules in lung CT of physical examination by artificial intelligence(AI).Methods A total of 500 patients who underwent lung CT examination in our hospital from August to September in 2020 were continuously collected.All lung CT images were detected by AI software.The detected lung nodules were identified as true-positive nodules and false-positive nodules by two radiologists,and the number,size and density of the false-positive nodules were recorded and classified.Results A total of 1518 lung nodules were detected by AI,including 740 true-positive nodules and 778 false-positive nodules,with an average of 1.6 false-positive nodules detected by CT.Among the false-positive nodules of different sizes,with a diameter of less than 5 mm were the most common false-positive nodules,accounting for 68.6%(534/778).Among the false-positive nodules of different densities,the false-positive predictive value of part solid false nodules was 69.7%(23/33),which was higher than the 48.3%(353/731)of solid false nodules and the 53.3%(402/754)of pure ground-glass false nodules,there was a statistical difference among them(P<0.05).There was a significant positive correlation between age and the detection rate of falsepositive nodules.Spearman rank correlation coefficient(rs)=0.986,P<0.05.The main causes of false-positive pulmonary nodules detected by AI were pleural nodules(21.5%),cord shadows(17.9%),vascular thickening(13.8%),vascular bifurcation(12.2%)and pulmonary lobular structure(9.0%).Conclusion The distribution and composition of false-positive pulmonary nodules detected by AI are regular.Being familiar with these regularity can improve the doctor’s ability to identify false-positive nodules,and hope to provide a reference for AI to reduce the false-positive rate.