人工智能在肺结节诊断中的应用研究

Application of artificial intelligence in distinguishing malignant and benign lung nodules

目的肺癌早期影像学表现为无明显症状的肺内结节,为提高肺癌筛查效率,本研究拟探讨人工智能(AI)对肺结节及肺结节良恶性诊断的效能。方法收集2016年3月-2021年4月于兰州大学第一医院胸外科就诊的397例(共416个肺结节)患者的相关临床资料,分析对比AI组、人工阅片+AI组、人工阅片组的诊断性能以及AI诊断系统对不同密度、不同大小结节的诊断准确率。结果AI、人工阅片+AI、人工阅片组对肺结节良、恶性鉴别诊断敏感度、特异度、准确率分别为94.40%、46.70%、73.80%,92.00%、93.90%、92.79%,81.30%、87.90%、83.89%(P<0.001)。AI对实性、亚实性、磨玻璃样、钙化结节诊断准确率分别为67.98%、78.13%、87.27%、100.00%(P<0.001)。AI对直径0~10、10~20、20~30 mm结节诊断准确率分别为70.97%、82.25%、67.03%(P<0.05)。AI对20~30 mm实性结节准确率较低(60.69%),其中肺结核诊断准确率为28.00%;对0~10 mm亚实性结节诊断准确率也较低(47.62%)。结论在肺结节良恶性鉴别诊断方面,AI联合人工阅片具有可接受的区分力,可在临床中广泛应用,但AI的误诊率仍较高,需提高AI在20~30 mm实性结节、0~10 mm亚实性结节及肺结核中的诊断准确率。

objective Early imaging manifestations of lung cancer are asymptomatic pulmonary nodules.In order to improve the efficiency of lung cancer screening,this study intended to explore the effectiveness of artificial intelligence(AI)in the diagnosis of lung nodules and distinguish between benign and malignant lung nodules.Methods The clinical data of 397 patients(416 nodules)were collected in the Department of Thoracic Surgery of the First Hospital of Lanzhou University from March 2016 to April 2021.The diagnostic performance of AI group,artificial film reading+AI group and artificial film reading group as well as the diagnostic accuracy of AI diagnosis system for nodules of different densities and sizes were analyzed and compared.Results The sensitivity,specificity and diagnostic accuracy of AI,artificial radiography+AI,artificial radiography for diagnosing benign and malignant pulmonary nodules were 94.40%,46.70%,73.80%;92.00%,93.90%,92.79%;81.30%,87.90%,83.89%(P<0.001).The diagnostic accuracy of AI for solid,subsolid,ground glass and calcified nodules was 67.98%,78.13%,87.27%,100.00%(P<0.001).The diagnostic accuracy of AI for 0~10,10~20,20~30 mm was 70.97%,82.25%,67.03%(P<0.05).The accuracy rate of AI for 20~30 mm solid nodules was low(60.69%),and the accuracy rate of pulmonary tuberculosis diagnosis was 28.00%.The diagnostic accuracy of 0~10 mm subsolid nodules was also low(47.62%).Conclusion In the differential diagnosis of benign and malignant pulmonary nodules,AI combined with manual film reading will have an acceptable discrimination and can be widely used in clinical practice.However,the misdiagnosis rate of AI is still high.It is necessary to improve the diagnostic accuracy of AI in 20~30 mm solid nodules and 0~10 mm subsolid nodules.