乳腺X线影像组学标签在预测乳腺癌HER2表达中的价值  

The value of mammography based radiomics signature for preoperative prediction of HER2 expression in breast carcinoma

作　　者：帅鸽 郁义星[1] 董佳 杨玲[1] 胡春洪[1] SHUAI Ge;YU Yi-xing;DONG Jia(Department of Radiology,the First Affiliated Hospital of Soochow University,Jiangsu 215006,China)

机构地区：[1]苏州大学附属第一医院放射科,江苏215006 [2]苏州市立医院放射科,江苏215006

出　　处：《放射学实践》2022年第1期41-47,共7页Radiologic Practice

基　　金：国家重点研发计划数字诊疗装备研发基金资助项目(2017YFC0108900)。

摘　　要：目的:探讨乳腺X线摄影影像组学标签在预测乳腺癌HER2表达中的价值。方法:回顾性分析2018年1月-2020年10月在苏州大学附属第一医院及苏州市立医院经病理证实为乳腺癌患者的临床及X线资料。共入组222例女性患者,平均年龄(53.70±14.46)岁,其中HER2阳性患者59例,阴性患者163例,苏州大学附属第一医院患者设为训练集(n=154),苏州市立医院患者设为验证集(n=68),对比获得的双乳内外斜位(MLO)和头尾位(CC)X线图像,选取病灶面积较大的乳腺X线图像利用MaZda软件进行图像分割和影像组学特征提取。采用费希尔参数法(Fisher)、分类错误率联合平均相关系数法(POE+ACC)和相关信息测度法(MI)分别进行特征筛选,得到三组特征子集。将准确率最高的特征子集进行Z-Score标准化,再采用二元logistic回归进一步筛选,利用选择特征的线性融合构建影像组学标签,计算每例患者的组学标签得分,并进行受试者工作特征(ROC)曲线分析,计算其预测乳腺癌HER2表达的曲线下面积(AUC)、准确率、敏感度、特异度、阳性预测值和阴性预测值。结果:(POE+ACC)-NDA法筛选出的训练集特征子集准确率最高,为88.31%。在训练集利用逻辑回归筛选特征后获得影像组学标签Radscore=-2.149-0.548×WavEnLH;-4+0.475×Kurtosis-0.765×Perc.01%-0.703×WavEnHH;-5-0.513×Teta4+1.069×135dr;hrtREmp-3.831×WavEnHH;-1,在训练集HER2阳性组与阴性组乳腺癌的影像组学得分分别为0.159(-0.357,0.928)和-2.987(-3.997,-1.184),差异有统计学意义(Z=-8.088,P<0.001);在验证集HER2阳性组与阴性组乳腺癌的影像组学得分分别为0.475(-0.412,1.541)和-3.093(-4.126,-1.157),差异有统计学意义(Z=-4.865,P<0.001)。影像组学标签预测训练集乳腺癌HER2表达的AUC、准确率、敏感度、特异度、阳性预测值和阴性预测值分别为0.927(95%可信区间0.881~0.973)、85.4%、87.6%、71.4%、94.3%和87.0%;预测验证集乳腺癌HER2表达的AUC�Objective:To explore the value of mammography-based radiomics signature for preoperative prediction of HER2 expression in breast carcinoma.Methods:The clinical and X-ray data of patients with breast cancer confirmed by pathology in the first affiliated hospital of soochow university and Suzhou municipal hospital from January 2018 to October 2020 were retrospectively analyzed.A total of 222 female patients with an average age of 52.62±13.15 years old were enrolled,including 59 HER2 positive patients and 163 HER2 negative patients.Patients from the first affiliated hospital of soochow university were set as the training set(n=154),and patients from Suzhou municipal hospital were set as the validation set(n=68).Comparing the mediolateral oblique(MLO)and cranial cauda(CC)X-ray images,the mammography images with larger lesion areas were selected,and the image segmentation and icomic feature extraction were performed by Mazda software.Fisher coefficients(Fisher),classification error probability combined average correlation coefficients(POE+ACC)and mutual information(MI)were used to select 3 sets of feature subsets.Z-score standardization was carried out for the feature subset with the highest accuracy,and then binary logistics regression was used for further screening.The linear fusion of selected features was used to construct the Radiomics Signature.The score of each patient’s Radiomics Signature was calculated,and the receiver operating characteristic curve(ROC)was analyzed to calculate the AUC,accuracy,sensitivity,specificity,positive predictive value,and negative predictive value of HER2 expression in breast cancer.Results:The(POE+ACC)-NDA method had the highest accuracy of 88.31%.In the training set,radiomics signature was obtained by logistic regression:radscore=-2.149-0.548×WavEnLH_s-4+0.475×Kurtosis-0.765×Perc.01%-0.703×WavEnHH_s-5-0.513×Teta4+1.069×135 dr_ShrtREmp-3.831×WavEnHH_s-1.In the training set,the radiomics scores of breast cancer in HER2 positive group and HER2 negative group were 0.159(-0.

关 键 词：乳腺肿瘤 HER2表达 乳腺X线摄影 影像组学 

分 类 号：R737.9[医药卫生—肿瘤] R814.41[医药卫生—临床医学]