摘要
目的探讨多模态X线影像组学模型在鉴别乳腺BI-RADS 4类肿块型病变良恶性方面的价值。方法回顾性分析山东省千佛山医院2017年8月至2020年4月,经全屏数字化乳腺X线摄影(FFDM)和数字乳腺断层合成摄影(DBT)检查诊断为BI-RADS 4类乳腺病变并经病理证实的120例女性患者(4A 41例,良性34例、恶性7例;4B 30例,良性11例、恶性19例;4C 49例,良性4例、恶性45例),年龄20~86岁,平均(51.94±13.94)岁。患者均行乳腺FFDM及DBT头尾位(CC)和内外斜位(MLO)扫描。分别在FFDM及DBT图像的CC位和MLO位上对病变进行感兴趣区(ROI)勾画,其中DBT图像勾画病灶最清晰的五幅图像。利用组学软件(汇医慧影radcloud)进行分析,每个序列提取2818个特征值,并依次采用ANOVA和LASSO进行特征降维。以患者的病理结果为金标准,计算人工亚型分类的阳性预测值,并通过采用受试者工作特征(ROC)曲线评价不同亚型分类的诊断效能。分别构建FFDM、DBT、FFDM+DBT三种分类预测模型,利用K最近邻(KNN)模型进行机器学习,验证三种预测模型的诊断效能并采用ROC曲线成对对比(Z统计)进行三种预测模型之间的对比分析。使用SPSS 23.0进行统计分析。P<0.05为差异有统计学意义。结果 BI-RADS 4A、4B、4C亚型阳性预测值分别为17.1%(7/41),63.3%(19/30),91.8%(45/49),整体预测曲线下面积(AUC)值为0.87,诊断敏感度、特异度分别为83%和86%。在FFDM、DBT、FFDM+DBT三种预测模型中,DBT组AUC值(0.84)及特异度(70%)高于FFDM组(0.80,60%),FFDM+DBT组AUC值(0.91)及特异度(90%)均高于DBT组和FFDM组,三种预测模型敏感度相同,均为80%。三种预测模型两两比较后,仅FFDM+DBT组和FFDM组差异具有统计学意义(P<0.05),余两组之间不具有统计学差异。结论 FFDM结合DBT多模态X线影像组学模型有助于乳腺BI-RADS 4类肿块型疾病良恶性的鉴别,且其诊断效能高于临床亚型分类诊断方法。
Objective To investigate the value of the radiomics model of multimodal X-ray application in the differentiation of benign and malignant BI-RADS 4 breast masses. Methods From August 2017 to April 2020,120 female patients with BI-RADS 4 breast lesions [4 A,n=41(34 cases of benign, 7 cases of malignant);4 B,n=30(11 cases of benign, 19 cases of malignant);4 C,n=49(4 cases of benign, 45 cases of malignant)] diagnosed by full-field digital mammography(FFDM) and digital breast tomosynthesis(DBT) and confirmed by pathology in Qianfushan hospital were analyzed retrospectively. Among them, 49 patients had benign lesions whereas 71 had malignantlesions,with age ranging from 20 to 86 years(mean age, 51.94±13.94 years).Both craniocaudal(CC) and mediolateral oblique(MLO) mammographic views were taken for all the patients using FFDM and DBT.The regions of interest(ROI) of the lesions were delineated on FFDM and DBT images at CC and MLO views, respectively, among which five DBT images with the clearestlesions were delineated. Analysis was performed using radiomics software(Radcloud, Huiying Medical Technology Co.,Ltd.),with 2818 eigenvalues extracted from each sequence, and feature dimension reduced by ANOVA and LASSO,successively.With the pathological results of the patients as the golden standard, the positive predictive value(PPV) of artificial subtype classifications were calculated and the diagnostic efficacy of different subtypes was evaluated using ROC curve.FFDM,DBT and FFDM + DBT classification-based prediction models were established, respectively.Machine learning was carried out using the k-nearest neighbor(KNN) model. The diagnostic efficiency of the three prediction models were verified, and comparative analysis of the three prediction models was conducted by pairwise comparison of ROC curve(Z statistics). Statistical analysis which was performed using SPSS 23.0.P< 0.05 was considered as statistically significant. Results The PPV of subtypes BI-RADS 4 A,4 B and 4 C were 17%(7/41),63.3%(19/30) and 91.8%(45/49),respectively. The AUC of overall prediction was 0.87,and the diagnostic sensitivity and specificity were 83% and 86%,respectively. Among the FFDM,DBT and FFDM + DBT prediction models, the AUC value(0.84) and specificity(70%) of the DBT group were higher than those of the FFDM group(0.80,60%),and the AUC value(0.91) and specificity(90%) of the FFDM+DBT group were higher than those of the DBT group and the FFDM group. The sensitivity of the three prediction models were the same, both of which were 80%.Pairwise comparison of the three prediction models only showed that there was statistically significant difference between the FFDM+DBT group and the FFDM group(P< 0.05).No statistical difference was found between the other two groups. Conclusion The radiomics model of multimodal X-ray with FFDM combined with DBT is helpful for the differentiation of benign and malignant BI-RADS 4 breast masses, and its diagnostic efficiency is higher than that of clinical subtype classification.
作者
陈旭
宋歌声
李爱银
CHEN Xu;SONG GeshengLI Aiyin(School of Medicine,Shandong University,Jinan,Shandong Province 250012,P.R.China)
出处
《临床放射学杂志》
北大核心
2022年第1期54-58,共5页
Journal of Clinical Radiology
