人工智能系统在CT肺小结节筛查中的准确率及检出时间分析

Analysis of accuracy rate and detection time of artificial intelligence system in CT-guided small lung nodule screening

目的本研究旨在研究人工智能(AI)系统在CT肺小结节筛查中的准确率及检出时间情况,用以分析该AI系统在CT中的应用价值。方法回顾性分析2019年1—12月于牡丹江医学院附属红旗医院进行肺高分辨率CT体检检查的1316例患者,以经金标准诊断得出的198例患者的230个肺结节作为研究对象,分别运用AI检测法、人工检测法及AI+人工的检查方法对肺结节进行分析。对比分析各组对肺结节诊断准确率、检出时间、漏诊率及假阳性率情况。结果AI组共检测出336个肺结节,假阳性率30.95%(104/336),漏诊率0.87%(2/230);人工组共检测出190个肺结节,假阳性率为1.05%(2/190),漏诊率为18.26%(42/230);AI联合人工组共检测出341个肺结节,假阳性率为0.29%(1/341),漏诊率为0.87%(2/230),AI组假阳性率明显高于AI联合人工组及人工组(χ^(2)=6.067,5.596;P=0.002,0.004),人工组漏诊率明显高于AI联合人工组及AI组(χ^(2)=5.511,4.996;P=0.004,0.009)。AI组平均阅片时间为(24.35±7.37)s;人工组平均阅片时间为(633.45±212.37)s;AI联合人工组平均阅片时间为(189.84±61.28)s,AI组阅片时间明显短于AI联合人工组及人工组,差异有统计学意义(F=9.962,P=0.001)。AI组分析低至中度恶性概率结节114个,经病理证实17个结节为低至中度恶性结节,准确率为14.91%;AI组分析高度恶性概率结节14个,经病理证实8个结节为高度恶性结节,准确率为57.14%;AI联合人工组分析低至中度恶性概率结节24个,经病理证实17个结节为低至中度恶性结节,准确率为70.83%;AI联合人工组分析高度恶性概率结节10个,经病理证实8个结节为高度恶性结节,准确率为80.00%,AI联合人工组检测准确率高于单纯AI组(P<0.05)。结论AI系统在CT肺小结节筛查中的假阳性率较高,但漏诊率较低,需人工辅助AI检查将可获得较好的检出效果,且AI联合人工相较于传统人工阅片更省时。

Objective To investigate the accuracy rate and detection time of the artificial intelligence(AI)system in CTguided small lung nodules screening,so as to analyze the application value of AI system in CT.Methods A total of 1316 patients who underwent high-resolution CT-guided physical examination of lung in Hongqi Hospital of Mudanjiang Medical College from January to December 2019 were retrospectively analyzed,and 230 lung nodules from 198 patients who were diagnosed by the gold standard were selected as the study objects.The AI detection method,manual detection method and AI combined with manual detection method were used to analyze the lung nodules.The diagnostic accuracy rate,detection time,missed diagnosis rate and false positive rate of lung nodules in each group were compared and analyzed.Results A total of 336 lung nodules were detected in the AI group,with a false positive rate of 30.95%(104/336)and a missed diagnosis rate of 0.87%(2/230).190 lung nodules were detected in the manual detection group,with a false positive rate of 1.05%(2/190)and a missed diagnosis rate of 18.26%(42/230).341 lung nodules were detected in the AI combined with manual detection group,with a false positive rate of 0.29%(1/341)and a missed diagnosis rate of 0.87%(2/230).The false positive rate was significantly higher in the AI group than that in the AI combined with manual detection group and the manual detection group(χ^(2)=6.067,5.596;P=0.002,0.004).The missed diagnosis rate was significantly higher in the manual group than that in the AI combined with manual detection group and the AI group(χ^(2)=5.511,4.996;P=0.004,0.009).The mean imaging reading time was(24.35±7.37)s in the AI group,(633.45±212.37)s in the manual detection group and(189.84±61.28)s in the AI combined with manual detection group,the imaging reading time in the AI group was shorter than that in the AI combined manual detection and manual detetion group,with statistically significant difference(F=9.962,P=0.001).It was analyzed that there were 114 nodules with low to moderate malignant probability in the AI group,and 17 nodules were pathologically confirmed as low to moderate malignant nodule,with an accuracy rate of 14.91%.It was analyzed that there were 14 nodules with high malignant probability in the AI group,and 8 nodules were pathologically confirmed as highly malignant nodules,with an accuracy of 57.14%.It was analyzed that there were 24 nodules with low to moderate malignant probability in the AI combined with manual detection group,with an accuracy of 70.83%.It was analyzed that there were 10 nodules with high malignant probability nodules in the AI combined with manual detection group,and 8 nodules were pathologically confirmed as highly malignant nodules,and 17 nodules were pathologically confirmed as low to moderate malignant nodule,with an accuracy of 80.00%.The detection accuracy rate in AI combined with manual detection group was higher than that in the AI group alone(P<0.05).Conclusion The AI System has a higher false positive rate and a lower missed diagnosis rate in CT-guided small lung nodule screening.The AI examination with manual assistance will have better detection effect,and AI combined with manual assistance is more time-saving in comparison with traditional manual imaging reading.