基于乳腺数字X射线图像纹理分析对乳腺BI-RADS 4类以上钙化灶定性诊断的价值  

Value of texture analysis in qualitative diagnosis of BI-RADS category 4 or above calcification lesions based on digital X-ray mammography image of breast

作　　者：罗成龙 宋一曼 于湛[1] 岳松伟[1] 张永高[1] 高剑波[1] 丁昌懋[1] LUO Chenglong;SONG Yiman;YU Zhan;YUE Songwei;ZHANG Yonggao;GAO Jianbo;DING Changmao(Department of Radiology,the First Affiliated Hospital,Zhengzhou University,Zhengzhou 450052)

机构地区：[1]郑州大学第一附属医院放射科,郑州450052

出　　处：《郑州大学学报（医学版）》2022年第2期233-237,共5页Journal of Zhengzhou University(Medical Sciences)

基　　金：国家重点研发计划项目(2017YFC0109404)。

摘　　要：目的:探讨基于乳腺数字X射线图像纹理分析对BI-RADS 4类以上钙化灶定性诊断的价值。方法:收集临床未触及明确肿块、仅由乳腺数字X射线摄影发现钙化灶(BI-RADS 4a及以上)的患者178例。病灶均在X射线引导下手术切除,并行病理学检查,其中良性117例,恶性61例。以7∶3的比例将所有病例分为测试集和验证集。应用MaZda软件,采用3种特征向量降维算法(费希尔参数法、最小分类误差与最小平均相关系数法、相关信息测度法)分别筛选出10个最佳纹理特征,合并后得到第4组最佳纹理特征(MPF)。方法一,依据上述4组最佳纹理特征采用线性判别分析(LDA)和非线性判别分析(NDA)对钙化灶进行良恶性判别。方法二,比较测试集中良恶性组间最佳纹理特征(MPF组,30个)的差异,以筛选出的最佳纹理特征构建Logistic回归模型,并用验证集评估模型诊断效能。结果:依据第4组最佳纹理特征,NDA的误判率最低,测试集误判率为13.60%,验证集为9.43%。Logistic回归分析结果表明S(0,4)SumOfSqs及S(3,3)Entropy为区分良恶性钙化灶的影响因素,该模型鉴别验证集良恶性钙化灶的AUC(95%CI)为0.771(0.642~0.901),准确度、敏感度和特异度分别为0.717、0.833、0.629。结论:基于乳腺数字X射线图像纹理分析的LDA/NDA或Logistic回归模型对BI-RADS 4类以上钙化灶均有较好的定性诊断价值,可为临床诊治提供较为客观的参考。Aim:To explore the value of texture analysis in the qualitative diagnosis of BI-RADS category 4 or above calcification lesions based on breast digital X-ray mammography image.Methods:A total of 178 patients with calcification lesions(above BI-RADS 4a)found by digital X-ray mammography without clinical palpation were collected.The lesions were located under the guidance of X-ray,surgically resected and had pathological examination.Among 178 patients,117 had benign lesion and 61 had malignant lesion.All the patients were allocated into test set and verification set at 7∶3.MaZda software was used to extract 300 texture features from the focus of the image.The best texture features were selected by Fisher method,POE+ACC(PA)method and MI method,and the 3 methods were combined to get the 4th group of best texture features(MI+PA+F,MPF).MethodⅠ,linear discriminate analysis(LDA)and nonlinear discriminate analysis(NDA)were used to classify the benign and malignant calcification lesions according to the 4 groups of best texture features,respectively.MethodⅡ,the difference of the best texture features(MPF group,n=30)between the benign and malignant calcification lesions in the test set was compared to further select the best texture features,a multi-factor Logistic regression model was constructed according to the chosen texture features,and the model was evaluated by the data in the verification set.Results:The error rate of NDA method was the lowest based on the texture features of MPF group,with 13.60%of test set and 9.43%of verification set.The results of Logistic regression analysis showed that S(0,4)SumOfSqs and S(3,3)Entropy were factors for distinguishing benign and malignant calcification lesions;in the verification set,the AUC(95%CI)of the model was 0.771(0.642-0.901),and the accuracy,sensitivity and specificity were 0.717,0.833 and 0.629,respectively.Conclusion:LDA/NDA or Logistic regression model based on texture analysis of digital X-ray images of breast is of great value in qualitative diagnosis of BI-RAD

关 键 词：乳腺数字X射线摄影 纹理分析 乳腺影像报告和数据系统 乳腺钙化灶 

分 类 号：R737.9[医药卫生—肿瘤]