目的

基于增强MRI 影像组学特征的生境分析，建立乳腺癌HER-2 表达状态预测模型。

方法

回顾性分析2018 年1 月至2023 年5 月在南京医科大学附属苏州医院接受增强MRI 检查的168 例乳腺癌患者增强T1 压脂序列第2 期图像数据，其中HER-2 阴性100 例，HER-2 阳性68 例。 对图像进行预处理后，手动分割得到全肿瘤感兴趣体积（VOI）。 提取24 项局部影像组学特征，并使用高斯混合模型（GMM）结合贝叶斯信息准则（BIC）进行聚类，获得生境亚区域。 分别提取亚区域及全肿瘤区域影像组学特征，并按照7 ∶3 比例随机划分训练集（117 例）和验证集（51 例）。 采用逻辑回归（LR）、支持向量机（SVM）和K-近邻（KNN）3 种算法，分别构建生境预测模型及全肿瘤预测模型。 通过验证集曲线下面积（AUC）结果选择最佳模型，并汇总其受试者操作特征（ROC）曲线，利用DeLong 检验比较2 种预测模型AUC 值的差异，同时使用决策曲线分析（DCA）评估模型的临床应用价值。

结果

每个全肿瘤VOI被划分为3 个生境亚区域。 SVM 为最佳建模方法，生境预测模型-SVM 在训练集的AUC 值为0.949（95%CI： 0.915～0.984），验证集的AUC 值为0.844（95%CI： 0.725 ～0.963）。 最佳的全肿瘤预测模型-SVM 在训练集的AUC 值为0.870（95%CI： 0.809 ～0.931），验证集的AUC 值为0.735（95%CI： 0.588 ～0.882）。 生境预测模型-SVM 在训练集和验证集的准确度、敏感度及特异度均优于全肿瘤预测模型-SVM，且DeLong 检验结果显示两者在训练集的AUC 值差异具有统计学意义（Z=2.134，P=0.033）。DCA 分析结果表明，生境预测模型-SVM 在预测HER-2 表达状态时具有更高的整体净获益。

结论

本研究基于增强MRI 影像组学的生境分析，建立了乳腺癌HER-2 表达状态预测模型，可为乳腺癌患者的精准治疗提供依据。

Objective

To establish a prediction model for HER-2 expression status in breast cancer using habitat analysis based on enhanced MRI radiomic features.

Methods

A retrospective analysis was conducted on the second-phase DCE-T₁ WI data of 168 breast cancer patients who underwent enhanced MRI examinations in the Affiliated Suzhou Hospital of Nanjing Medical University from January 2018 to May 2023.Among them， 100 cases were HER-2-negative， and 68 cases were HER-2-positive. After preprocessing the images， the whole tumor volume of interest （VOI） was manually segmented. Twenty-four regional radiomic features were extracted and clustered into habitat subregions using a Gaussian Mixture Model （GMM） combined with the Bayesian Information Criterion （BIC）. Radiomic features were separately extracted from the subregions and the entire tumor region. The data were randomly divided into a training set （117 cases） and a validation set（51 cases） at the ratio of 7 ∶3. Logistic regression （LR）， support vector machine （SVM）， and k-nearest neighbor （KNN） algorithms were used to construct habitat prediction models and whole-tumor prediction models. The optimal model was selected based on the area under the curve （AUC） value in the validation set，the receiver operating characteristic （ROC） curve was drawn. DeLong’s test was used to compare the AUC of the two prediction models， and decision curve analysis （DCA） was performed to evaluate the clinical utility of the models.

Results

Each tumor VOI was segmented into three habitat subregions. SVM was identified as the best modeling method. The habitat prediction model-SVM achieved an AUC of 0.949 （95%CI：0.915-0.984） in the training set and 0.844 （95%CI： 0.725-0.963） in the validation set. The best whole-tumor prediction model-SVM achieved an AUC of 0.870 （95%CI： 0.809-0.931） in the training set and 0.735 （95%CI： 0.588-0.882） in the validation set. The habitat prediction model-SVM demonstrated superior accuracy， sensitivity and specificity compared with the whole-tumor prediction model-SVM in both the training and validation sets.DeLong’s test indicated that the AUC differences between the models in the training set were statistically significant （Z=2.134， P=0.033）. DCA results showed that the habitat prediction model-SVM provided the highest net benefit.

Conclusions

This study established a prediction model for HER-2 expression status in breast cancer based on habitat analysis using enhanced MRI radiomic features， which may provide a reference for precision treatment of breast cancer patients.