基于CT生境影像组学和深度学习模型预测进展期胃癌HER2表达状态的价值  

Value of CT habitat radiomics and deep learning models in predicting HER2 expression status in advanced gastric cancer

作　　者：任梦婷 陈基明[1] 昌杰 杨利 牛冉冉 翟建[1] REN Meng-ting;CHEN Ji-ming;CHANG Jie(Medical Imaging Centre,Yijishan Hospital of Wannan Medical College,Anhui 241001,China)

机构地区：[1]皖南医学院弋矶山医院影像中心,安徽芜湖241001 [2]皖南医学院医学信息学院,安徽芜湖241000

出　　处：《放射学实践》2025年第4期501-508,共8页Radiologic Practice

摘　　要：目的:探讨基于CT生境影像组学和深度学习(DL)模型在预测进展期胃癌(AGC)患者人表皮生长因子受体2(HER2)表达状态中的价值。方法:回顾性分析2013年12月-2023年9月在本院经术后病理证实的201例AGC患者的临床及影像学资料。根据HER2表达水平将患者分为HER2阳性组(60例)和阴性组(141例)。按照7∶3的比例采用随机分组法将患者分为训练集(n=141)和验证集(n=60)。在增强CT静脉期图像上手动逐层勾画肿瘤ROI,使用生境成像技术根据聚类情况将胃癌病灶划分为3个亚区(ITH1、ITH2、ITH3),提取肿瘤整体及各亚区的影像组学特征,并使用ResNet50网络提取肿瘤的深度学习特征;采用最小冗余最大相关(mRMR)和最小绝对收缩和选择算子(LASSO)两种方法进行特征降维,分别构建基于肿瘤整体、各亚区及DL算法的影像组学模型,并计算基于肿瘤整体及各亚区的影像组学得分。在训练集中,比较HER2阳性组与阴性组的临床病理和影像学特征的差异,并采用多因素Logistic回归分析筛选独立预测因子和构建临床模型。结合临床独立预测因子和ITH1的影像组学得分构建联合模型。根据ROC曲线下面积(AUC)评估各模型的预测效能,使用校准曲线和Hosmer-Lemeshow检验分析模型的拟合优度,应用决策曲线(DCA)分析模型的临床价值。结果:临床模型、肿瘤整体模型、ITH1模型、ITH2模型、ITH3模型、DL模型及联合模型在训练集中预测HER2阳性的AUC(95%CI)分别为0.833(0.757~0.909)、0.782(0.700~0.864)、0.868(0.802~0.934)、0.848(0.779~0.916)、0.806(0.726~0.886)、0.848(0.785~0.911)和0.918(0.869~0.968),在验证集中分别为0.718(0.574~0.863)、0.701(0.551~0.851)、0.821(0.693~0.950)、0.778(0.652~0.904)、0.738(0.600~0.876)、0.753(0.617~0.888)和0.873(0.780~0.966),以联合模型的AUC最大。Hosmer-Lemeshow检验结果显示联合模型校准曲线的拟合度良好(P>0.05)。DCA结果显示ITH1模型及联合模型的临床获�Objective:To explore the value of CT habitat radiomics and deep learning models in predicting HER2 expression status of advanced gastric cancer.Methods:Clinical and imaging data of 201 patients with AGC confirmed by postoperative pathology from December 2013 to September 2023 in our hospital were analyzed retrospectively.All patients were divided into a HER2 positive group(60 cases)or a HER2 negative group(141 cases)according to the level of HER2 expression.All patients were randomly allocated to a training cohort(n=141)and a validation cohort(n=60)at the ratio of 7∶3.Tumor regions of interest(ROIs)were manually delineated slice-by-slice on venous-phase contrast-enhanced CT images.Habitat imaging was employed to partition gastric cancer lesions into three intratumoral heterogeneity subregions(ITH1,ITH2,ITH3)through cluster analysis.Radiomic features were extracted from the whole tumor and three sub-regions respectively,and deep learning features were obtained via a ResNet50 network.Feature dimensionality reduction was performed using minimum redundancy maximum relevance(mRMR)and least absolute shrinkage and selection operator(LASSO)algorithms.Predictive models were subsequently constructed for the whole tumor,individual subregions,and deep learning radiomics.Radiomics scores for both the whole tumor and each subregion were subsequently calculated for every patient.In the training set,clinical features diffe-rences between HER2 positive and negative groups were compared.Independent clinical predictors were determined using multivariate logistic regression and were utilized to develop a clinical model.A combined model integrating independent clinical predictors with the ITH1 radiomics score was deve-loped.The predictive performance of each model was evaluated using the area under the receiver ope-rating characteristic curve(AUC).The goodness of fit of the models was analyzed using calibration curves and the Hosmer-Lemeshow test,and the clinical utility of each model was assessed by decision curve analysis(DCA).Re

关 键 词：胃肿瘤 生境分析 深度学习 影像组学 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] R735.2[医药卫生—放射医学]