基于肺CT的人工智能对良恶性肺结节的诊断价值

Diagnostic value of artificial intelligence based on lung CT for benign and malignant pulmonary nodules

目的通过比较人工智能(artificial intelligence,AI)与两位放射科医师(医师1为主治医师、医师2为副主任医师)对良恶性肺结节电子计算机断层扫描(computed tomography,CT)诊断的一致性和效能,探讨AI在肺结节良恶性诊断中的价值。方法回顾性分析2021年1月至2022年10月在浙江中医药大学附属杭州市中医院经手术病理证实的肺结节患者201例,共229个肺结节,其中良性结节74个,恶性结节155个。采用加权Kappa检验评估AI与两位放射科医师诊断的一致性,采用受试者操作特征(receiver operating characteristic,ROC)曲线评估AI与两位医师的诊断效能。结果在部分实性结节、磨玻璃结节、实性结节及部分实性+磨玻璃+实性结节良恶性诊断中,AI与医师2的一致性均高于AI与医师1,而医师2的曲线下面积(area under the curve,AUC)均高于AI与医师1,且在磨玻璃结节、实性结节及部分实性+磨玻璃+实性结节AUC间差异均有统计学意义(P<0.05)。在部分实性结节、磨玻璃结节的良恶性诊断中,医师1的AUC高于AI,但两者比较,差异无统计学意义(P>0.05)。在实性结节、部分实性+磨玻璃+实性结节良恶性诊断中,AI的AUC高于医师1,两者间差异有统计学意义(P<0.05)。在磨玻璃结节、实性结节及部分实性+磨玻璃+实性结节良恶性诊断中,AI的敏感度(97%、92%、94%)均高于医师1(58%、89%、72%)、医师2(83%、84%、85%)。结论AI在肺结节良恶性诊断中具有一定的诊断效能,本研究中所采用的AI系统的总体诊断效能介于医师1与医师2之间,但AI的敏感度高于后两者。

Objective To explore the value of artificial intelligence(AI)in the diagnosis of pulmonary nodules in terms of consistency and efficiency compared with two radiologists(physician 1 is a chief physician and physician 2 is a deputy chief physician)in the diagnosis of benign and malignant pulmonary nodules using computed tomography(CT).Methods Retrospective analysis of 201 patients with pulmonary nodules confirmed by surgery pathology at Hangzhou Municipal Hospital affiliated to Zhejiang Chinese Medical University from January 2021 to October 2022,including a total of 229 pulmonary nodules,of which 74 were benign and 155 were malignant.The consistency of AI diagnosis with two radiologists was evaluated by weighted Kappa test,and the diagnostic performance of AI with the two radiologists was evaluated by the receiver operating characteristic curve(ROC).Results In the diagnosis of the benign and malignant nature of partial solid nodules,ground-glass nodules,solid nodules,and partial ground-glass and solid plus ground-glass nodules,the consistency between AI and physician 2 was higher than that between AI and physician 1.Additionally,the area under the curve(AUC)of physician 2 was higher than that of AI and physician 1 with statistically significant differences between the AUCs of ground-glass nodules,solid nodules,and partial ground-glass and solid plus ground-glass nodules(P<0.05).In the diagnosis of the benign and malignant nature of partial solid nodules and ground-glass nodules,the AUC of physician 1 was higher than that of AI,but there was no statistically significant difference between the two(P>0.05).In the diagnosis of the benign and malignant nature of solid nodules and partial ground-glass and solid plus ground-glass nodules,the AUC of AI was higher than that of physician 1 with statistically significant differences between the two(P<0.05).In the diagnosis of the benign and malignant nature of ground-glass nodules,solid nodules,and partial ground-glass and solid plus ground-glass nodules,AI's sensitivity(97%,92%,and 94%)was higher than that of physician 1(58%,89%,and 72%)and physician 2(83%,84%,and 85%).Conclusion AI has a certain diagnostic efficacy in the diagnosis of pulmonary nodules malignancy.The overall diagnostic efficacy of the AI system used in this study is between that of physician 1 and physician 2,but its sensitivity is higher than that of the latter two.