基于乳腺X线图像纹理特征建立机器学习模型在鉴别良恶性乳腺肿块中的价值  

Application of machine learning model based on the texture features on mammography in the differeniation diagnosis of benign and malignant breast masses

作　　者：黄栎有 高先聪 尤传文 HUANG Li-you;GAO Xian-cong;YOU Chuan-wen(Department of Oncology,the Affiliated Suqian Hospital of Xuzhou Medical University,the Suqian People's Hospital of Nanjing Drum Towr Hospital Group,Jiangsu 223800,China)

机构地区：[1]徐州医科大学附属宿迁医院,南京鼓楼医院集团宿迁市人民医院肿瘤科,江苏223800 [2]徐州医科大学附属宿迁医院,南京鼓楼医院集团宿迁市人民医院放射科,江苏223800

出　　处：《放射学实践》2021年第4期480-483,共4页Radiologic Practice

摘　　要：目的:探讨基于乳腺X线图像的纹理分析建立机器学习模型在鉴别乳腺肿块良恶性中的价值。方法:回顾性搜集经病理证实的124个乳腺良性肿块和139个乳腺恶性肿块的乳腺X线图像。并按照7﹕3的比例划将所有病灶随即分为训练集和验证集。使用MaZda软件,在X线图像上于乳腺病灶内手动勾画ROI,提取6类共133个纹理特征,经降维处理后,利用训练集数据得到线性判别分析(LDA)、Logistic回归(LR)、随机森林(RF)和支持向量机(SVM)共4种模型。在验证集中对这4种模型进行验证。通过符合率、Kappa系数和AUC值分别评价4种模型在训练集和验证集中的表现,并通过delong法比较4种模型间AUC值的差异。结果:RF模型在训练集和验证集中符合率、Kappa系数和AUC值均高于其它模型;其中,RF模型在验证集中的符合率为94.9%、Kappa系数为0.896、AUC值为0.946,与LDA模型、LR模型间AUC值的差异均具有统计学意义(P<0.05)。SVM模型的符合率和Kappa系数仅次于RF模型;在验证集中,SVM模型的AUC值高于LDA和LR模型,但差异无统计学意义(P>0.05)。结论:基于乳腺X线图像纹理特征建立的机器学习模型在鉴别乳腺肿块良恶性中具有一定优势。其中RF模型表现出较好的诊断效能,SVM模型的表现仅次于RF模型。Objective:The purpose of this study was to investigate the value of machine learning model based on the texture features of mammography image in the differentiation diagnosis of benign and malignant breast masses.Methods:Mammography images of 124 benign and 139 malignant breast masses were collected and analyzed retrospectively.All the masses were divided into two groups for training set and verification set according to the proportion of 7﹕3.The region of interest(ROI)was drawn,and 133 texture features of six types were extracted using MaZda software.After the extracted texture features were reduced in dimensionality,four models including linear discriminant analysis(LDA),logistic regression(LR),random forest(RF)and support vector machine(SVM)were obtained based on training set data;and verified in the verification set.The accuracy,Kappa coefficient,and AUC were used to evaluate the performance of the four models in the training and verification sets,and the AUC differences of the four models were analyzed by the delong test.Results:The accuracy,Kappa coefficient,and AUC of the RF model in the training and verification sets were higher than those of the other three models.In the verification set,the accuracy of the RF model was 94.9%,the Kappa coefficient was 0.896,and the AUC was 0.946,which was statistically different with the AUC value of LDA model and LR model(P<0.05).The accuracy and Kappa coefficient of SVM model were second only to RF model.The AUC of SVM model was higher than that of LDA and logistic regression model,but the difference was not statistically significant(P>0.05).Conclusion:The machine learning model based on the texture features of mammography image has certain advantages in differentiating benign and malignant breast masses.RF model showed better classification performance,and SVM model was second only to RF model.

关 键 词：乳腺肿瘤 乳房X线摄影术 纹理分析 机器学习 诊断效能 

分 类 号：R814.41[医药卫生—影像医学与核医学] R737.9[医药卫生—放射医学]