人工智能肺小结节检测系统在低剂量CT肺筛查中的价值

The value of artificial intelligence pulmonary nodules detection system in low-dose CT lung screening

目的探讨人工智能(AI)肺结节检测系统在低剂量CT结节筛查及诊断中的检出时间、漏检率、假阳性率及准确率,判断其辅助诊断效能。方法对我院行肺小结节低剂量CT扫描的患者共6537例进行回顾性分析,选择经手术病理证实的52例共57个结节作为观察对象,然后分别用AI肺结节检测系统(AI组)、人工(人工组)及人工+AI(人工辅助AI组)进行分析,对每组肺结节的检出时间、漏检率、假阳性率及诊断准确率进行统计分析。结果AI组肺结节检出假阳性率为24.8%,漏检率为0;人工组假阳性率为0,对小于5 mm的肺结节漏检率为39.1%。人工辅助AI组假阳性率及漏检率均为0。人工辅助AI组肺结节的平均检出时间为(138±96)s,明显小于单纯人工组的(478±306)s,且差异具有统计学意义(P<0.001)。AI检测为低至中度恶性概率结节的准确率为5.8%,高恶性概率结节的准确率为86.5%,人工辅助AI组恶性结节诊断准确率为98.2%。结论尽管存在一定的假阳性率,但AI辅助诊断系统对肺结节的总体检出能力良好。人工阅片辅助AI阅片为最佳解决方案,可以显著减少阅片时间、提高阅片效率、降低漏检率及假阳性率,提高了肺小结节的诊断准确率。

Objective To investigate the detection time,missed detection rate,false positive rate and accuracy rate of an artificial intelligence(AI)pulmonary nodule detection system in low-dose CT screening and diagnosis,and to judge its auxiliary diagnostic efficiency.Methods A total of 6537 patients who underwent low-dose CT scanning of small pulmonary nodules in our hospital from May 2018 to December 2019 were retrospectively analyzed,and 57 nodules of 52 patients confirmed by operation and pathology were selected in this study.Then,the nodules were analyzed by using the AI pulmonary nodule detection system(AI group),artificial(artificial group)and artificial+AI(artificial auxiliary AI Group)respectively.The detection time,missed detection rate,false positive rate and diagnostic accuracy rate in each group were statistically analyzed.Results The false positive rate of pulmonary nodules in AI group was 24.8%,and the missed detection rate was 0;the false positive rate of artificial group was 0,and the missed detection rate of pulmonary nodules less than 5 mm was 39.1%.The false positive rate and missed detection rate of artificial auxiliary AI group were 0.The average detection time of pulmonary nodules in artificial assistant AI Group was(138±96)s,which was significantly shorter than(478±306)s in artificial group,and the difference was statistically significant(P<0.001).The accuracy rate of AI for low to moderate malignant nodules was 5.8%,for high malignant nodules that was 86.5%,and the accuracy rate of artificial auxiliary AI was 98.2%.Conclusion Although there is a certain false positive rate,the AI diagnosis system has a good overall detection ability for pulmonary nodules.Artificial auxiliary AI is the best solution modality to reduce the film-reading time,the missed detection rate and false positive rate,and to improve the review efficiency and diagnostic accuracy of pulmonary nodules.