人群肺亚实性结节CT筛查及人工智能应用研究初探

Population-based research of pulmonary subsolid nodule CT screening and artificial intelligence application

目的研究并探讨胸部低剂量CT(LDCT)应用于人群肺亚实性结节的筛查情况及人工智能的应用价值。方法回顾性分析山西省潞安区2015年1月至2017年12月间常规体检行LDCT筛查人群的临床资料,分析统计该地区筛查人群的特征、肺部亚实性结节的检出情况以及检出亚实性结节的独立预测因素,并评价人工智能阅片方法的准确性。结果该地区人群三轮筛查显示肺亚实性结节检出率分别为0.42%、0.69%和0.92%。完成三轮筛查的人群纳入队列研究(726例),以男性为主(83.2%),中位年龄43岁,47.0%有吸烟史,肺癌家族史(OR=8.753,95%CI:1.877~40.816,P=0.006)是检出亚实性结节的独立预测因素。110 kVp组(656例)模型和人工阅片方法的曲线下面积(AUC)分别为0.740、0.721,差异无统计学意义(P=0.502);NRI=-0.15,P=0.003,提示模型的准确性差于人工阅片方法。130 kVp组(98例)模型和人工阅片方法的AUC分别为0.888、0.756,差异无统计学意义(P=0.128);NRI=0.19,P=0.123,提示模型的准确性不亚于人工阅片方法。结论该地区常规体检人群亚实性结节检出率为0.42%~0.92%,肺癌家族史是其独立预测因素。人工智能肺结节检出模型的训练集数据参数与实际的应用参数匹配时,其结果具有一定的参考价值。

Objective To investigate the application of low-dose chest CT(LDCT)in the screening of pulmonary subsolid nodules in population and the application value of artificial intelligence.Methods People who received chest LDCT screening between January 2015 and December 2017 were included.A retrospective study was developed to analyze the enrolled population features,detection of pulmonary subsolid nodules and independent predictors of subsolid nodules,and to evaluate the accuracy of the artificial intelligence reading method.Results Result of three cross-sectional studies reveals that the detection rates of pulmonary subsolid nodules were 0.42%,0.69%and 0.92%in three rounds.726 cases who completed the three rounds of screening were included in the cohort study.The cohort population was predominantly male(83.2%),with a median age of 43 years,and nearly half of the subjects(47.0%)had a history of smoking.GEE revealed that the patient's family history of lung cancer(OR=8.753,95%CI:1.877-40.816,P=0.006)was an independent predictor of the detection of subsolid nodules.In the 110 kVp tube voltage group,AUC of AI model was 0.740,and AUC of the manual reading method was 0.721,no significant differences were observed(P=0.502);when the preseted cutoff value of AI model was 0.75,the NRI was-0.15,indicating the accuracy of AI model was inferior to manual method(P=0.006).In the 130 kVp tube voltage group,AUC of the model was 0.888,and AUC of the manual reading method was 0.756,no significant differences were observed(P=0.128);and the NRI was 0.19,indicating the accuracy of AI model was not inferior to manual method(P=0.123).Conclusion This population's detection rates of pulmonary subsolid nodules were 0.42%-0.92%.Family history of lung cancer was an independent predictor of subsolid pulmonary nodules.The result of AI pulmonary nodule detection model could be a reference when the training set data parameters match the actual application parameters.