摘要
目的探讨基于支持向量机(SVM)分类器提取的数字乳腺X线影像纹理特征用于诊断乳腺肿块良恶性的可行性和准确性。方法选取384例经乳腺X线摄影诊断为BI-RADS 3、4、5类的乳腺肿块(良性188例,恶性196例)患者图像,所有乳腺肿块均手术经组织病理学证实。由两名副主任医师共同使用5M工作站将病灶区域在图像上用矩形框标记。对感兴趣的矩形区域(ROI)进行分割,从每个ROI中提取4类共455个影像组学特征,包括一阶直方图特征,二阶纹理特征及高阶Gabor特征。利用最大相关最小冗余(MRMR)特征选择算法对提取的特征集进行降维操作从而特征优化。通过构建SVM对降维后得到的特征进行分类,70%的数据作为训练集,另外30%作为测试集,分类器的稳定性通过十倍交叉验证进行评估。分别使用准确性、敏感性、特异性和AUC对分类器的性能进行评价。并且使用SAS对降维后的特征进行统计学分析。结果利用MRMR算法筛选出30个对良恶性肿块诊断有价值的影像组学特征。十倍交叉验证精度为92.54%,模型稳定。通过MRMR算法降维的SVM分类器的准确性为93.2%,敏感性为92.2%,特异性为94.1%,AUC为0.963。结论基于SVM分类器对数字乳腺X线影像进行纹理特征提取,可用于预测乳腺肿瘤的良恶性特征,该方法具有较高的准确性、敏感性和特异性,有望为影像医师提供更多的诊断信息。
Objective To explore the feasibility and accuracy of extracting texture features from digital mammograms for predicting benign and malignant breast masses using Radiomics.Methods 384 patients who were diagnosed with breast masses(188 Benign and 196 Malignant)by mammography were enrolled.All breast masses were classified as BI-RADS 3,4,and 5,and at last confirmed by histopathology.Lesion areas were marked with a rectangular frame on the images at the 5 MP workstation by two senior radiologists.The rectangular regions of interest(ROI)were segmented and 4 categories of 455 radiomics features were extracted from every ROI,including first-order histogram features,second-order texture features and high-order Gabor features.Extracted features were dimensioned by MRMR algorithm.Post-dimension features were classified using Support Vector Machine(SVM).70%of the data as a training set and the other 30%as a testing set.The reliability of the Classifier was evaluated by the 10-fold cross-validation,and the precision of the Classifier was evaluated by the accuracy,sensitivity,specificity and the AUC.In addition,SAS was used for statistical analysis of the post-dimension features.Results 30 radiomics features that were valuable for diagnosing benign and malignant tumors were screened by MRMR algorithm.10-fold cross-validation showed the accuracy was 92.54%,so the model was stable.In the testing sets,using the MRMR dimension reduction algorithm,the SVM Classifier could achieve an accuracy of 93.2%,a sensitivity of 92.2%,a specificity of 94.1%,and the AUC of 0.963,higher for predicting benign and malignant breast tumors.Conclusion Radiomics techniques can be used to extract texture features from digital mammograms for predicting the characterization of breast tumors.This method offers high accuracy and sensitivity,and it is expected to provide additional diagnostic information to the surgeons and radiologists.
作者
崔延华
朱健
李云
CUI Yanhua;ZHU Jian;LI Yun(Department of Radiology Oncology Physics and Technology,Shandong Cancer Hospital and Institute,Shandong First Medical University and Shandong Academy of Medical Sciences,Jinan 250117,P.R.China)
出处
《临床放射学杂志》
CSCD
北大核心
2020年第7期1296-1301,共6页
Journal of Clinical Radiology
基金
山东省医学科学院院级科学计划青年项目(编号:2017-44)
山东省医学科学院院级科学计划面上项目(2017-09)
山东省泰山学青年专家项目(tsqn201909140)
山东第一医科大学学术提升计划项目(2020RC003)。
