ABVS图像冠状面汇聚征特性的量化及其应用

Quantification of spiculation characteristics of ABVS coronal images and its application

目的:设计一组新颖、有效的特征来定量化描述超声自动乳腺全容积扫描(automated breast volume scanner,ABVS)图像冠状面的汇聚征,以辅助判别乳腺肿瘤良恶性。方法:首先,在ABVS图像冠状面自动分割出肿瘤区域。其次,经过多尺度、多角度滤波后得到最大能量图,并进行阈值筛选及形态学处理,然后提取描述汇聚征的6个特征。最后,建立分类器,验证所设计的特征在判别有无汇聚征、判别肿瘤良恶性等任务的有效性。结果:对195例乳腺肿瘤数据进行回顾性研究,采用交叉验证的方式评价所设计的特征的分类性能。在有无汇聚征的判别任务中,采用单个特征时的受试者工作特征(receiver operating characteristic,ROC)曲线的曲线下面积(area under curve,AUC)为0.83~0.89,综合6个特征的AUC为0.91。在肿瘤良恶性判别的任务中,单个特征判别结果的AUC为0.68~0.75,综合6个特征时的AUC为0.75。结论:所设计的特征能够很好地量化ABVS冠状面的汇聚征性质,在乳腺肿瘤良恶性判别中具有重要意义。

Objective:To design a set of novel and effective features to quantify the spiculation characteristic of the coronal image in automated breast volume scanner(ABVS),in order to assist in classifying benign and malignant breast tumors.Methods:Firstly,the tumor regions on the coronal planes of ABVS images were automatically segmented.Secondly,the maximum energy maps with multi-scale and multi-angle filtering were obtained,and the thresholding and morphological processing on the energy maps were performed.Then six features to describe the spiculation characteristic were extracted.Finally,we established a classifier and verified the effectiveness of the designed features in the tasks of recognizing the spiculation and classifying benign and malignant tumors.Results:The classification performances of the designed features by cross-validation on the data of 195 cases of breast tumors were evaluated.In the task of recognizing the spiculation,the area under curve(AUC)of the receiver operating characteristic(ROC)curve was between 0.83 and 0.89 when using only one of the six features,and 0.91 when using all six features.In the task of classifying benign and malignant tumors,the AUC was between 0.68 and 0.75 when using only one of the six features,and 0.75 when using all six features.Conclusion:The designed features can well quantify the speculation characteristic in the coronal plane of ABVS.It is of great significance in classifying benign and malignant tumors.