基于AI技术CT直方图参数模型预测微小磨玻璃结节样肺腺癌浸润性  

The predictive value of CT histogram parameter model based on artificial intelligence technology for the invasiveness of microscopic ground-glass nodular lung adenocarcinoma

作　　者：邓琦[1] 潘爱珍[1] 徐志锋[1] 周涛[1] 李勤祥[1] 林景兴 王诚明 DENG Qi;PAN Ai-zhen;XU Zhi-feng(Foshan First People's Hospital,Guangdong 528000,China)

机构地区：[1]佛山市第一人民医院影像科,广东佛山528000

出　　处：《放射学实践》2022年第8期977-981,共5页Radiologic Practice

基　　金：佛山市卫生健康局医学科研课题(20220048);佛山市科技创新平台(FSOAA-KJ218-1301-0021);佛山市第一人民医院"登峰计划"移动医疗创新平台(2020B003)。

摘　　要：目的:探讨人工智能(AI)技术提取的CT直方图定量参数建立的诊断模型对微小磨玻璃结节样(长径≤10 mm)肺腺癌浸润性的预测价值。方法:回顾性分析经手术病理证实为早期肺癌的98例患者共102个长径≤10 mm的微小磨玻璃结节样(GGN)病灶的胸部HRCT图像,其中原位癌(AIS)32个、微浸润腺癌(MIA)21个、浸润性腺癌(IAC)49个。将AIS和MIA归为无浸润组,IAC归为浸润组。采用独立样本t检验(满足正态分布)或Mann-Whitney U检验(不满足正态分布)比较两组结节的长径和AI技术提取的直方图中各定量参数(包括实性成分占比、最大CT值、最小CT值、平均CT值、中位CT值、标准差、偏度、峰度和熵)值的差异。对组间差异有统计学意义的参数采用受试者工作特征曲线(ROC)评估其诊断价值,将AUC大于0.7的参数纳入logistic回归分析,筛选出GGN浸润性的独立预测因子并建立诊断模型,利用ROC曲线分析模型的诊断效能,绘制预测模型的nomogram图,并采用校准曲线评价其预测效果。结果:实性成分占比、最大CT值、平均CT值、偏度、峰度和熵在无浸润性组和浸润性组间的差异均有统计学意义(P<0.05);长径、最小CT值、中位CT值和标准差在两组间的差异无统计学意义(P>0.05)。ROC曲线分析结果显示,各参数的诊断效能由高到低依次为熵(AUC=0.860)、平均CT值(AUC=0.845)、实性占比(AUC=0.817)、最大CT值(AUC=0.690)、峰度(AUC=0.665)和偏度(AUC=0.652)。logistic回归分析结果显示熵(OR=16.647,P<0.05)和平均CT值(OR=1.021,P<0.05)是预测GGN浸润性的独立影响因子;其中,熵的诊断阈值为3.745,相应敏感度和特异度分别为87.8%和89.6%;平均CT值的诊断阈值为-479.500 HU,相应敏感度和特异度分别为81.6%和84.9%。nomogram图显示预测模型结果与实际情况之间的一致性良好,C指数为0.840(95%CI:0.757~0.923)。结论:基于AI技术建立的CT直方图定量参数模型对长径≤10 mm早期肺�Objective:To explore the predictive value of the diagnostic model established by the quantitative parameters of CT histogram extracted by artificial intelligence(AI)technology for the invasiveness of microscopic ground-glass nodular(long diameter≤10mm)lung adenocarcinoma.Methods:The chest HRCT images of 102 tiny ground-glass nodule(GGN)lesions with long diameter≤10mm were retrospectively analyzed in 98 patients with early-stage lung cancer confirmed by surgery and pathology.Among them,32 were carcinoma in situ(AIS),21 were minimally invasive adenocarcinoma(MIA),and 49 were invasive adenocarcinoma(IAC).AIS and MIA were classified as non-invasive group,and IAC was classified as invasive group.Independent sample t-test(normal distribution satisfied)or the Mann-Whitney U-test(normal distribution unsatisfied)were used to compare the difference of the long diameter and quantitative histogram parameters(including solid composition proportion,maximum CT value,minimum CT value,the average CT value,median CT value,standard deviation,skewness,kurtosis and entropy)extracted by AI technique between the two groups.Recei-ver operating characteristic curve(ROC)was used to analyze the diagnostic value of the parameters with statistically significant differences between the two groups,and the parameters with AUC greater than 0.7 were included in logistic regression analysis to select out independent predictors of GGN invasiveness and establish diagnostic model,the ROC curve was used to analyze their diagnostic perfor-mance,then the nomogram of the prediction model was drawn,and the calibration curve was used to evaluate its prediction effect.Results:The proportion of solid components,maximum CT value,average CT value,skewness,kurtosis and entropy were significantly different between the non-invasive group and the invasive group(P<0.05).There were no significant differences in long diameter,minimum CT value,median CT value and standard deviation between the two groups(P>0.05).The ROC curve analysis showed that the diagnostic ef

关 键 词：肺腺癌 磨玻璃结节 人工智能 直方图 浸润性 体层摄影术 X线计算机 

分 类 号：R814.42[医药卫生—影像医学与核医学] R734.2[医药卫生—放射医学]