基于XGBoost算法探究混合磨玻璃结节型肺腺癌浸润程度的关键影像学特征  

Key Imaging Features of the Degree of Invasion of Mixed Ground-Glass Module Lung Adenocarcinoma Based on the XGBoost Algorithm

作　　者：陈凯 吴天晨[1] 宋德胤[1] 沈广澍[1] 梁艳[1] CHEN Kai;WU Tianchen;SONG Deyin;SHEN Guangshu;LIANG Yan(Imaging Department,Nanjing Hospital of C.M.,Nanjing 210022,China)

机构地区：[1]江苏省南京市中医院影像科,210022

出　　处：《实用心脑肺血管病杂志》2025年第1期23-28,共6页Practical Journal of Cardiac Cerebral Pneumal and Vascular Disease

基　　金：国家中医药管理局第五批全国中医临床优秀人才研修项目(国中医药人教函[2022]1号);江苏省南京市中医药青年人才培养项目(ZYQ20047)。

摘　　要：目的基于XGBoost算法探究混合磨玻璃结节(mGGN)型肺腺癌浸润程度的关键影像学特征。方法回顾性选取2020年1月—2023年1月南京市中医院心胸外科收治的mGGN型肺腺癌患者104名,收集患者的性别、年龄及影像学特征。按照肿瘤细胞有无基底膜浸润将患者分为腺体前驱期组(n=21)和浸润性病变期组(n=83)。对两位资深病理学专家判定mGGN型肺腺癌浸润程度的结果进行一致性检验。按照3∶7的比例将104例mGGN型肺腺癌患者分为训练集(n=31)和测试集(n=73),其中训练集进行XGBoost算法,并构建暗箱模型,绘制ROC曲线以评价暗箱模型对测试集mGGN型肺腺癌浸润性病变期的预测价值,采用R包计算每个特征变量的SHAP值。结果两位资深病理学专家诊断mGGN型肺腺癌浸润程度的组内相关系数(ICC)为0.97。腺体前驱期组与浸润性病变期组患者结节边缘影像学征象、结节长径、结节短径、结节长短径均值、结节实性部分长径比较,差异有统计学意义(P<0.05)。采用XGBoost算法在训练集构建暗箱模型,选择对数损失函数(Logloss)最小的暗箱模型参数,按照相对权重系数大小排序,前五位特征变量分别是结节长径、结节实性部分长径、结节边缘毛刺征、结节边缘分叶征、结节实性部分/结节长径比值。SHAP图分析结果显示,结节长径、结节实性部分长径、结节实性部分/结节长径比值每增加一个标准单位,mGGN型肺腺癌浸润性病变期的风险分别升高65.3%、61.9%、45.0%;出现结节边缘毛刺征、结节边缘分叶征时,mGGN型肺腺癌浸润性病变期的风险分别升高38.1%、37.7%。ROC曲线分析结果显示,该暗箱模型预测测试集mGGN型肺腺癌浸润性病变期的AUC为0.88[95%CI(0.75~0.96)]。结论结节长径、结节实性部分长径、结节边缘毛刺征、结节边缘分叶征、结节实性部分/结节长径比值是mGGN型肺腺癌浸润程度的关键影像学特征。Objective To explore the key imaging features of the degree of invasion of mixed ground-glass module(mGGN)lung adenocarcinoma based on the XGBoost algorithm.Methods A total of 104 patients with mGGN lung adenocarcinoma admitted to the Department of Cardiothoracic Surgery,Nanjing Hospital of C.M.from January 2020 to January 2023 were retrospectively selected.The gender,age and imaging characteristics of the patients were collected.According to the presence or absence of basement membrane infiltration of tumor cells,the patients were divided into glandular precursor group(n=21)and invasive lesion group(n=83).Consistency test was performed on the results of two senior pathologists in determining the degree of invasion of mGGN lung adenocarcinoma.According to the ratio of 3∶7,104 patients with mGGN lung adenocarcinoma were divided into training set(n=31)and test set(n=73).The training set was performed by XGBoost algorithm,and the dark box model was constructed.The ROC curve was drawn to evaluate the predictive value of the dark box model for the invasive stage of mGGN lung adenocarcinoma in the test set.The SHAP value of each characteristic variable was calculated by R package.Results The intra-class correlation coefficient(ICC)of two senior pathologists in diagnosing the degree of invasion of mGGN lung adenocarcinoma was 0.97.There were statistically significant differences in the imaging signs of nodule edge,long diameter of nodule,short diameter of nodule,mean value of long and short diameter of nodule,and long diameter of nodule solid part between the glandular precursor group and the invasive lesion group(P<0.05).The XGBoost algorithm was used to construct the dark box model in the training set.The dark box model parameters with the smallest logloss function were selected.According to the relative weight coefficient,the top five characteristic variables were the long diameter of nodule,long diameter of nodule solid part,nodule edge burr sign,nodule edge lobulation sign and ratio of long diameter of nodule sol

关 键 词：肺腺癌 混合磨玻璃结节 影像学特征 XGBoost算法 

分 类 号：R734.2[医药卫生—肿瘤]