基于融合双模态超声瘤内瘤周影像的乳腺肿瘤分类

Breast tumor classification by fusion of intratumoral and peritumoral ultrasound images based on dual-modality

本研究通过135例临床乳腺肿瘤的灰阶超声和应变弹性超声的双模态图像研究,并结合肿瘤感兴趣区域(region of interest,ROI)与瘤周组织超声信息进行乳腺肿瘤的良恶性分类。首先,分别提取肿瘤ROI区域的常规灰阶超声和应变弹性超声的影像组学特征:形态学特征(14个)、强度特征(18个)和纹理特征(75个),并提取瘤周区域的双模态超声强度特征和纹理特征;然后采用最小绝对收缩和选择算法(Lasso)进行特征筛选,得到最佳特征组合;最后,利用支持向量机进行良恶性分类。实验结果表明,将灰阶超声、应变弹性超声ROI区域和瘤周区域特征进行融合分析后,其受试者工作特性曲线下面积(area under curve,AUC)为(0.8895±0.0176)。其结果远高于单纯灰阶超声ROI区域得到的(0.8267±0.0150)。

We studied the B-mode ultrasound and strain elastic ultrasound images of 135 clinical breast tumors,and combined with the region of interest(ROI)and the ultrasound information of the peritumoral tissue to classify the benign and malignant breast tumors.First,radiomics features of the ROI area of B-mode ultrasound and strain elastic ultrasound were extracted,included of 14 pcs morphological features,18 pcs intensity features and 75 pcs texture features,and the intensity and texture features of dual-modality ultrasound images at peritumoral area were extracted.Then the features were selected by least absolute shrinkage and selection operator(Lasso).Finally,support vector machine was used to classify benign and malignant breast tumors.When used all features of ROI and peritumoral area based on B-mode ultrasound and strain elastic ultrasound,the final fusion modal achieved area under curve(AUC)of(0.8895±0.0176).It is much superior than AUC of only based on the ROI area of the B-mode ultrasound(0.8267±0.0150).