基于特征融合的集成模型分类乳腺癌分子亚型的研究

An integrated model based on feature fusion for classifying molecular subtypes of breast cancer

目的融合传统影像组学特征和卷积神经网络特征,构建分类乳腺癌分子亚型的支持向量机(support vector machine,SVM)集成模型,探讨该模型分类乳腺癌分子分型的价值。材料与方法回顾性分析Duke-Breast-Cancer-MRI数据集中经病理证实的189例乳腺癌患者病例,其中Luminal型71例、人类表皮生长因子受体-2(human epidermal growth factor receptor 2,HER-2)过表达型57例、三阴性型61例。对所有患者的动态对比增强MRI(dynamic contrast-enhanced MRI,DCE-MRI)图像进行预处理后,按照8∶2的比例分为训练集(n=151)和测试集(n=38)。使用传统影像组学方法和DenseNet169网络模型对患者病变感兴趣区(region of interest,ROI)提取特征,然后使用Spearman相关系数和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法对传统影像组学特征进一步筛选,最后使用特征融合后的特征组构建SVM集成分类模型。使用宏观平均方法绘制受试者工作特征(receiver operating characteristic,ROC)曲线来判断集成模型的诊断效能。结果三种乳腺癌分子亚型的传统影像组学特征经过筛选后分别获得51、49、20个特征标签,将其分别与卷积神经网络提取的1664个特征进行融合并建模。其中Luminal型与HER-2过表达型构建的分类器曲线下面积(area under the curve,AUC)值为0.880[95%置信区间(confidence interval,CI):0.814~0.946],Luminal型与三阴性型构建的分类器AUC值为0.861(95%CI:0.791~0.931),HER-2过表达型与三阴性型构建的分类器AUC值为0.696(95%CI:0.571~0.822)。由3个二分类器组成的SVM集成模型的AUC值为0.820(95%CI:0.725~0.915)。结论基于特征融合的SVM集成模型在分类三种乳腺癌分子亚型时表现出良好的效果,对术前乳腺癌分子亚型的分类具有重要的指导价值。

Objective:To construct an integrated support vector machine(SVM)model for classifying molecular subtypes of breast cancer by fusing traditional radiomics features and convolutional neural network features,and the value of this model for classifying molecular subtypes of breast cancer was explored.Materials and Methods:One hundred and eighty-nine patients with pathologically confirmed breast cancer in the Duke-Breast-Cancer-MRI dataset were retrospectively analyzed,including 71 cases of Luminal type,57 cases of human epidermal growth factor receptor 2(HER-2)overexpression type,and 61 cases of triple-negative type.After preprocessing the dynamic contrast-enhanced MRI(DCE-MRI)images of all patients,the cases were divided into a training set(n=151)and testing set(n=38)in the ratio of 8∶2.The features were extracted from the region of interest(ROI)of the patient's lesion using traditional radiomics model and the DenseNet169 network model,then the traditional radiomics features were further filtered using Spearman correlation coefficient and the least absolute shrinkage and selection operator(LASSO)algorithm.Finally,the SVM-integrated classification model was constructed using the fused feature set.The macro-averaging method was used to plot the diagnostic effect of the integrated model using the receiver operating characteristic(ROC)curve.Results:The traditional imaging histology features of three molecular subtypes of breast cancer were filtered to obtain 51,49,and 20 feature labels,which were fused and modeled with 1664 features extracted by convolutional neural networks,respectively.The area under the curve(AUC)value of the classifier constructed by Luminal and HER-2 overexpression type was 0.880[95%confidence interval(CI):0.814-0.946],the AUC value of the classifier constructed by Luminal and triple-negative type was 0.861(95%CI:0.791-0.931),and the AUC value of the classifier constructed by HER-2 overexpression type and triple-negative type was 0.696(95%CI:0.571-0.822).The AUC value of the SVM integrated model consisting of three binary classifiers was 0.820(95%CI:0.725-0.915).Conclusions:The integrated SVM model based on feature fusion showed good results in classifying three molecular subtypes of breast cancer,which is an important guide for the preoperative classification of molecular subtypes of breast cancer.