多模态影像组学在鉴别乳腺良恶性病变中的应用研究  

Application Study of Multimodal Radiomics for the Differential Diagnosis of Benign from Malignant Breast Lesions

作　　者：张书海 王小雷 朱芸[2] 王祥芝 徐敏[2] 汤晓敏[2] 刘浩 马宜传[2] 谢宗玉[2] ZHANG Shuhai;WANG Xiaolei;ZHU Yun(School of Graduate,Bengbu Medical College,Bengbu Anhui 233030,P.R.China)

机构地区：[1]蚌埠医学院研究生院,233030 [2]蚌埠医学院第一附属医院放射科,233004 [3]北京医准智能科技有限公司,100088

出　　处：《临床放射学杂志》2021年第11期2098-2104,共7页Journal of Clinical Radiology

基　　金：安徽省教育厅自然科学基金重点项目(编号:KJ2019A0402、KJ2018A0225)。

摘　　要：目的探究MRI双序列联合X线乳腺钼靶摄影(MG)双体位的影像组学特征在鉴别乳腺良恶性病变中的价值,分析MRI联合MG影像组学模型的鉴别能力。方法回顾性搜集经手术病理或穿刺活检证实的194例乳腺病变患者,其中良性69例,恶性125例。选取患者的MRI-T_(2)WI抑脂序列、MRI-DCE图像上病灶最大面积层面及同一病灶的MG头尾位(CC)、内外斜位(MLO)图像勾画感兴趣区,提取病灶感兴趣区特征,按照7:3比例将样本随机分为训练集136例和测试集58例,通过统计及机器学习方法进行特征降维,筛选出一组最优特征,通过支持向量机(SVM)建立MRI联合MG的多模态影像组学模型,并根据MRI和MG的BI-RADS评分及病灶影像最大径建立传统影像诊断模型,分析各模型对乳腺良恶性病变鉴别能力。结果多模态影像组学模型训练集受试者工作曲线(ROC)的曲线下面积(AUC)、灵敏度、特异度、约登指数、准确率、阳性精确率、阳性召回率、阳性F1-score分别为0.99、0.94、0.96、0.90、0.96、0.95、0.99、0.97,测试集分别为0.97、0.97、0.90、0.87、0.93、0.95、0.95、0.95;传统影像诊断模型的训练集AUC、灵敏度、特异度、约登指数、准确率、阳性精确率、阳性召回率、阳性F1-score分别为0.88、0.98、0.77、0.75、0.88、0.89、0.93、0.91,测试集分别为0.83、0.95、0.71、0.66、0.84、0.85、0.92、0.88。结论多模态影像组学模型鉴别乳腺良恶性病变的效能优于传统影像诊断模型,基于MRI双序列及MG双体位的影像组学模型对乳腺良恶性病变具有较高的鉴别能力。Objective To explore the value of radiomics features of MRI double sequence and Mammography(MG)dual position in differentiating benign from malignant breast lesions,and to analyze the ability of MRI and MG radiomics model in differentiating benign from malignant breast lesions.Methods A retrospective analysis of 194 patients with breast lesions confirmed by biopsy or surgical pathology,including 69benign and 125malignant.The one-layer image with the largest area of the lesion in MRI and the same lesion in MG image were selected to sketch region of interest(ROI).Thefeatures of the ROI were eextracted,and the extracted ROI were randomly divided into the training group(136 cases)and the verification group(58 cases)according to the ratio of 7:3.Statisticaland machine learning methods were used to select the best image group features.We used support vector machine(SVM)based on MRI and mammography to establish multimodal radiomics model.The traditional imaging diagnosis model of breast lesions was established according to the BIRADS classification and maximum length diameter of the lesion.The value of each model in the differential of benign and malignant breast lesions was analyzed.Results The area under the curve(AUC)of the receiver operating characteristic(ROC),sensitivity,specificity,Youden index,accuracy,positive precision,positiverecall,positive F1-score of the training set in multimodal radiomics model were 0.99,0.94,0.96,0.90,0.96,0.95,0.99 and 0.97 respectively.The testing set were 0.97,0.97,0.90,0.87,0.93,0.95,0.95,0.95 respectively.The AUC,sensitivity,specificity,Youden index,accuracy,positiveprecision,positiverecall,positive F1-score of the traditional imaging diagnostic model of the training set in traditional imaging diagnostic model were 0.88,0.98,0.77,0.75,0.88,0.89,0.93,0.91 respectively;the testing set were 0.83,0.95,0.71,0.66,0.84,0.85,0.92,0.88 respectively.Conclusion The multimodal radiomics’model was superior to the traditional imaging diagnostic model,and the radiomics-based MRI combined with M

关 键 词：影像组学 乳腺病变 磁共振成像 X线乳腺钼靶摄影 多模态 

分 类 号：R737.9[医药卫生—肿瘤]