基于人工智能的CT影像加权评分系统对肺结节良恶性鉴别的价值  

Value of artificial intelligence-based weighted scoring system of CT findings for differentiating benign and malignant pulmo-nary nodules

作　　者：李志平 张永胜[1] 徐辉景 陈文 康书朝 李焕国 张育 崔凤[1] 杨勇[3] LI Zhiping;ZHANG Yongsheng;XU Huijing;CHEN Wen;KANG Shuchao;LI Huanguo;ZHANG Yu;CUI Feng;YANG Yong(Department of Radiology,Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University,Hangzhou 310007,China;不详)

机构地区：[1]浙江中医药大学附属杭州市中医院放射科,杭州310007 [2]浙江中医药大学附属杭州市中医院病理科,杭州310007 [3]浙江中医药大学附属杭州市中医院胸心外科,杭州310007

出　　处：《浙江医学》2024年第7期703-709,共7页Zhejiang Medical Journal

基　　金：杭州市生物医药和健康产业发展扶持科技专项(2021WJCY014);杭州市卫生科技计划项目(A20220438)。

摘　　要：目的探讨基于人工智能(AI)的CT影像加权评分系统对肺结节良恶性鉴别的价值。方法回顾性分析2021年1月至2022年9月在浙江中医药大学附属杭州市中医院经手术和病理证实的肺结节患者187例,共215个肺结节。术前CT图像导入AI肺结节CT影像辅助诊断系统,由该系统自动识别,得到结节的相关量化参数:结节数量、部位、类型、平均直径、体积,并提供每个肺结节的AI恶性概率数值。采用多因素logistic回归分析筛选良恶性肺结节鉴别诊断的独立影响因素,基于这些独立影响因素构建初级模型,并进行加权赋分得到综合加权评分。采用ROC曲线评估各个变量、初级模型和综合加权评分等相关指标对良恶性肺结节鉴别诊断的效能。结果215个肺结节中,良性肺结节69个,恶性肺结节146个。多因素logistic回归分析显示亚实性结节(OR=4.732)、AI恶性概率>0.6(OR=5.602)、医师阅片恶性(OR=28.654)是良恶性肺结节鉴别诊断的独立影响因素(均P<0.01),基于这些独立影响因素,构建初级模型。Hosmer-Lemshow检验显示初级模型具有良好的校准性(P=0.486)。比较综合加权评分、初级模型、结节特征、医师阅片、AI恶性概率的诊断效能,综合加权评分的AUC最高,为0.929。综合加权评分与初级模型的AUC比较差异无统计学意义(P>0.05),综合加权评分与医师阅片、结节特征、AI恶性概率的AUC比较差异均有统计学意义(均P<0.01)。为了简化放射科医师评估肺结节恶性风险的流程,将综合加权评分分成3个区间分数(<4分、4~6分、>6分)。随着综合加权评分的升高,肺结节被诊断为恶性的可能性也相应增加。结论基于AI技术构建的综合加权评分系统在良恶性肺结节鉴别诊断中具有较高的效能,有助于简化放射科医师的评估流程和临床决策的制定。Objective To explore the value of artificial intelligence(AI)-based weighted scoring system of CT findings for differentiating benign and malignant pulmonary nodules.Methods CT images of 187 patients with pulmonary nodules(215 nodules)confirmed by surgery and pathology in Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University from January 2021 to September 2022 were retrospectively analyzed.Pulmonary nodules on preoperative CT images were automatically identified by an AI pulmonary nodule imaging assistance diagnostic system,the related quantitative parameters were obtained,including nodule number,location,type,average diameter,volume;and the AI risk of malignancy probability for each nodule was assessed.Multivariate logistic regression analysis was used to screen for independent factors in the differential diagnosis of benign and malignant pulmonary nodules,and a primary model was constructed based on these factors;then a comprehensive weighted scoring system was developed.The diagnostic performance of various variables,primary models,and comprehensive weighted scoring system in distinguishing benign from malignant pulmonary nodules was assessed using ROC curves.Results Among 215 pulmonary nodules,69 were benign and 146 were malignant.Multivariate logistic regression analysis identified subsolid nodules(OR=4.732),AI malignancy probability>0.6(OR=5.602),and physician's reading of malignancy(OR=28.654)as independent influencing factors in the differential diagnosis(all P<0.01).A primary model was built based on these factors,and the Hosmer-Lemeshow test indicated good calibration of the primary model(P=0.486).In the comparison of diagnostic efficacy of comprehensive weighted scoring system,primary model,nodule characteristics,physician's reading and AI malignancy probability assessment,the area under ROC curve(AUC)of comprehensive weighted scoring system was the highest(0.929).There was no significant difference in AUC between comprehensive weighted scoring system and primary model(P>0.05),but the

关 键 词：人工智能 肺结节 评分系统 鉴别诊断 

分 类 号：R730.44[医药卫生—肿瘤] R734.2[医药卫生—临床医学]