【摘要】 目的：探讨基于瘤内及瘤周水肿影像组学参数联合自编码器算法对乳腺癌表皮生长因子受体2（human epidermal growth factor receptor,HER-2）状态的预测价值，以为术前无创预测HER-2状态提供新的思路。方法：回顾性收集南京大学医学院附属盐城第一医院（中心1）的145例及南通大学附属建湖医院（中心2）的52例乳腺癌患者的临床资料及影像数据。根据免疫组化染色结果，分为HER-2阳性组（中心1为87例，中心2为30例）和HER-2阴性组（中心1为58例，中心2为22例）。将2018年12月至2024年10月收集的中心1有瘤周水肿的78例患者按照7∶3的比例随机分为训练集（55例）及验证集（23例），将2024年11月至2025年3月收集的中心1有瘤周水肿的26例患者作为时间验证集。将中心2的52例伴有瘤周水肿的乳腺癌患者作为外部测试集。首先，在T₂WI-FS序列图像上应用Mazda软件对瘤体最大层面及瘤周水肿区域进行感兴趣区勾画；其次，采用多变量方差分析（analysis of variance,ANOVA）、Kruskal-Wallis检验、递归特征消除（recursive feature elimination,RFE）及Relief算法筛选影像组学特征；最后，结合十折交叉验证，绘制ROC曲线，评价基于自编码器、支持向量机、线性判别分析、随机森林、Logistic回归、Lasso正则化逻辑回归、自适应提升、高斯过程、贝叶斯及决策树10种机器学习算法和影像组学参数分别构建的模型对乳腺癌HER-2状态的诊断效能。结果：基于瘤内MaxNorm、Variance及瘤周水肿SumAverg 3个特征参数，结合自编码器算法构建的模型诊断效能最优，训练集及验证集平均AUC值分别为0.808及0.735，时间验证集AUC值为0.746，外部测试集AUC值为0.732。结论：基于瘤内及瘤周水肿影像组学参数联合自编码器算法构建的模型能够术前无创预测乳腺癌HER-2状态，可为乳腺癌患者个性化治疗方案的制订提供参考。更多还原

【Abstract】 Objective To explore the predictive value of intratumor and peritumoral edema radiomics combined with the autoencoder algorithm for human epidermal growth factor receptor(HER-2) status in breast cancer to provide a new idea for preoperative noninvasive prediction of HER-2 status. Methods Totally 145 breast cancer patients from Yancheng Hospital Affiliated to Nanjing University Medical College(Center 1) and 52 ones from Jianhu Hospital Affiliated to Nantong University(Center 2) had their clinical and imaging data collected retrospectively, who were divided into a HER-2 positive group including 87 ones from Center 1 and 30 ones from Center 2 and a HER-2 negative group including 58 ones from Center 1 and 22 ones from Center 2. From December 2018 to October 2024 there were 78 patients with peritumoral edema from Center 1 randomly enrolled into a training set(55 patients) and a validation set(23 patients) in a ratio of 7∶3, and from November 2024 to March 2025 another 26 ones placed into a time validation set. The 52 patients with peritumoral edema from Center 2 were included into an external test set. Firstly, the Mazda software was used to delineate the regions of interest for the largest tumor layer and the peritumoral edema area. Secondly, multivariate analysis of variance(ANOVA), Kruskal-Wallis test, recursive feature elimination(RFE) and Relief algorithm were respectively employed to screen the radiomics features; finally, combined with ten-fold cross validation, the receiver operating characteristic curve was drawn, and the diagnostic efficacy of the models respectively constructed with radiomics parameters and ten types of machine learning algorithms, including auto encoder, support vector machine, linear discriminant analysis, random forest, Logistic regression, Logistic regression via Lasso, adaptive boosting, Gaussian process, native Bayes and decision tree, was evaluated for the HER-2 status in breast cancer. Results The model established by the auto encoder algorithm combined with three feature parameters including intratumor MaxNorm and Variance and peritumoral edema SumAverg behaved the best. The average AUC values of the training and validation sets were 0.808 and 0.735 resepctively, and the AUC values of the time validation and external test sets were 0.746 and 0.732 respectively. Conclusion The model developed with intratumor and peritumoral edema radiomics combined with the auto encoder algorithm can be used for preoperative noninvasive prediction of HER-2 status of breast cancer, which provides references for the preparation of individualized treatment scheme of breast cancer patients. [Chinese Medical Equipment Journal,2025,46(9):9-15]更多还原