Abstract: Background　Lung cancer is the most malignant tumor with the highest morbidity and mortality in the world.Early CT screening can reduce the mortality of patients with lung cancer，but at present，there are problems such as the large number of screenings，the heavy workload of doctors and a high rate of missed diagnosis.Artificial intelligence（AI） is an emerging method to screen early lung cancer.It is of great clinical significance to study its diagnostic value for pulmonary nodules.Objective　To explore the effect and clinical value of the application of AI in the diagnosis of pulmonary nodules.Methods　From July 2017 to July 2019，120 patients with pulmonary nodules diagnosed in the Third Affiliated Hospital of Jinzhou Medical University were selected as the study objects，and 256 pulmonary nodules were screened out.The positive rates of ground glass nodules（GGN） diagnosed by AI reading and physician reading of chest CT were compared，and the consistency of two methods was evaluated by Kappa test.Taking pathological examination results as the "gold standard"，the sensitivity，specificity，positive predictive value，negative predictive value and area under the receiver operating characteristic（ROC） curve of the diagnosis results by two methods in the malignant pulmonary nodules were calculated.Results　The positive rate of GGN diagnosed by AI reading〔87.2%（95/109）〕 was higher than that by physician reading〔68.8%（75/109）〕（χ2=10.686，P=0.001）.The coincidence of two diagnostic methods was weak，with a Kappa coefficient of 0.019.Of the 54 pulmonary nodules examined pathologically，47 were malignant and 7 were benign.The sensitivity of the diagnosis by AI reading and physician reading in malignant pulmonary nodules were 91.5% and 87.2%；the specificity were 57.1% and 85.7%；the positive predictive value were 93.5% and 97.6%；the negative predictive value were 50.0% and 50.0%，and the area under the ROC curve were 0.743 and 0.845.Conclusion　The AI reading in the diagnosis of GGN has a higher positive rate，which can be used to reduce the rate of missed diagnosis of GGN in clinical practice.Meanwhile，the sensitivity of AI reading in the diagnosis of malignant pulmonary nodules is higher than that of physician reading，but its specificity is lower.It is suggested that AI reading combined with physician reading should be used in clinical practice.

Key words: Artificial intelligence；Sarcoidosis, pulmonary；Lung neoplasms；Screening；Ground glass nodule

摘要： 背景　肺癌是世界上发病率和死亡率最高的恶性肿瘤，CT早期筛查可降低肺癌患者的死亡率，但目前存在筛查人数多、医师工作量大及漏诊率高等情况。人工智能（AI）为筛查早期肺癌的新兴手段，研究其对肺结节的诊断价值具有重要的临床意义。目的　探讨AI诊断肺结节的应用效果和临床价值。方法　选取2017年7月—2019年7月于锦州医科大学附属第三医院确诊的120例肺结节患者（共筛查出256个肺结节）为研究对象。分别采用AI阅片和医师阅片方法判读所有患者的胸部CT，比较两种阅片方式诊断磨玻璃结节（GGN）的阳性率，并采用Kappa检验评定两种方法的一致性。以病理检查结果为“金标准”，计算两种阅片方法在恶性肺结节诊断中的灵敏度、特异度、阳性预测值、阴性预测值及受试者工作特征（ROC）曲线下面积（AUC）。结果　AI阅片诊断GGN的阳性率〔87.2%（95/109）〕高于医师阅片〔68.8%（75/109）〕，差异有统计学意义（χ2=10.686，P=0.001），两种诊断方法的吻合度较弱，Kappa系数为0.019。54个经过病理检查的肺结节中，恶性肺结节47个，良性肺结节7个。AI阅片与医师阅片用于诊断恶性肺结节的灵敏度分别为91.5%、87.2%，特异度分别为57.1%、85.7%，阳性预测值分别为93.5%、97.6%，阴性预测值分别为50.0%、50.0%。AUC分别为0.743、0.845。结论　AI阅片诊断GGN的阳性率更高，用于临床可降低GGN的漏诊率；AI阅片诊断恶性肺结节的特异度低于医师阅片，灵敏度高于医师阅片，建议采取AI联合医师阅片方式应用于临床。

关键词: 人工智能；结节病, 肺；肺肿瘤；筛查；磨玻璃结节