不同WHO肺癌病理分类下人工智能对肺结节良恶性的诊断效能  

Diagnostic Performance of Artificial Intelligence on Benign and Malignant Pulmonary Nodules Under Different WHO Pathological Classifications of Lung Cancer

作　　者：黄文君 李巧巧 敬文斌 葛艳明 董鹏 范丽 HUANG Wenjun;LI Qiaoqiao;JING Wenbin(Department of Radiology,The Second People's Hospital of Deyang,Deyang,Sichuan Province 618000,P.R.China)

机构地区：[1]德阳市第二人民医院放射科,618000 [2]海军军医大学第二附属医院放射诊断科,上海200003 [3]山东第二医科大学医学影像学院,261000 [4]山东第二医科大学附属医院影像中心

出　　处：《临床放射学杂志》2025年第1期76-82,共7页Journal of Clinical Radiology

基　　金：国家自然科学基金项目(编号:82171926,81930049,82202140);国家重点研发计划项目(编号:2022YFC2010002,2022YFC2010000);上海市科学技术委员会项目(编号:21DZ2202600);国家卫生健康委员会医学影像数据库建设项目(编号:YXFSC2022JJSJ002);上海长征医院临床创新项目(编号:2020YLCYJ-Y24);德阳市科学技术局社会发展领域重点研发指导类项目(2024SZY115);北京医学奖励基金会睿影科研基金(编号:YXJL-2024-0350-0183)。

摘　　要：目的验证商用的人工智能(AI)辅助诊断系统在世界卫生组织(WHO)新旧两种不同针对腺体前驱病变(PGL)的肺癌分类标准下预测肺结节良恶性的诊断效能。方法回顾性搜集2019年1月至2021年12月期间经病理证实的肺结节,共404例。根据WHO 2015版和2021版两种分类标准,将PGL分别纳入恶性病变和良性病变进行良恶性分组。分析AI诊断肺结节与术后病理的一致性,并构建不同分类标准下基于AI定量特征的恶性肺结节预测模型。结果当PGL被纳入恶性病变时(WHO 2015版),术后病理的良恶性与AI诊断的高低危间具有高度的一致性(Kappa=0.658),平均CT值(P<0.001)是肺结节恶性的独立保护因素,预测模型的受试者工作特征曲线曲线下面积(AUC)为0.806(95%CI:0.756~0.856);当PGL被纳入良性病变时(WHO 2021版),术后病理的良恶性与AI诊断的高低危间具有中等程度的一致性(Kappa=0.440),熵(P=0.008)是肺结节恶性的独立危险因素,预测模型的AUC为0.749(95%CI:0.700~0.797)。结论AI软件的肺结节诊断结果与术后病理具有中等到高度的一致性,显示出较好的诊断性能,但需要随着病理分类的更新进一步优化对PGL的评估。Objective This study aimed to assess the accuracy of a commercially available AI-assisted diagnostic tool for distinguishing between benign and malignant pulmonary nodules,in accordance with both the previous and updated WHO version of lung cancer classification criteria.Methods A total of 404 with pathologically confirmed pulmonary nodules after surgical resection or needle biopsy from January 2019 to December 2021 were retrospectively collected.According to the 2015 and 2021 World Health Organization(WHO)classification criteria,precursor glandular lesions(PGL)were divided into malignant and benign lesions,respectively.To analyze the consistency between AI-assisted diagnosis of pulmonary nodules and postoperative pathology and to construct a clinical prediction model of malignant pulmonary nodules based on AI quantitative features under different classification criteria.Results(1)When the PGL was malignant(2015 WHO classification),there was a high consistency between the benign and malignant results of postoperative pathology and the high and low risk diagnosed by AI(Kappa=0.658).The mean CT attenuation(P<0.001)was an independent protective factor for the malignancy of pulmonary nodules,and the AUC of the constructed clinical prediction model was 0.806(95%CI:0.756-0.856).(2)When the PGL was benign(2021 WHO classification),there was moderate consistency between the benign and malignant results of postoperative pathology and the high and low risk diagnosed by AI(Kappa=0.440).The entropy(P=0.008)was an independent risk factor for malignancy of pulmonary nodules.The AUC of the constructed clinical prediction model was 0.749(95%CI:0.700-0.797).Conclusion AI software's diagnostic results of pulmonary nodules have moderate to high consistency with postoperative pathology,showing good diagnostic performance.However,the evaluation of PGL by AI needs to be optimized with the update of pathological classification.

关 键 词：人工智能 肺结节 腺体前驱病变 体层摄影术 X线计算机 

分 类 号：R734.2[医药卫生—肿瘤] TP18[医药卫生—临床医学] TP391.41[自动化与计算机技术—控制理论与控制工程]