基于多模态磁共振影像组学鉴别唾液腺多形性腺瘤和基底细胞腺瘤

Differentiating salivary gland pleomorphic adenoma from basal cell adenoma based on multimodal magnetic resonance imaging radiomics

目的探讨基于表观弥散系数(apparent diffusion coefficient,ADC)图、T1WI及T2WI序列构建的影像组学模型鉴别唾液腺多形性腺瘤(pleomorphic adenoma,PA)和基底细胞腺瘤(basal cell adenoma,BCA)的价值。材料与方法回顾性分析2015年1月至2021年10月来自济宁市第一人民医院的唾液腺129例PA和48例BCA患者的MR图像,并将其以8∶2的比例随机划分为训练集(n=141)与测试集(n=36)。在横断位ADC、T1WI及T2WI图像上手动勾画肿瘤的三维容积感兴趣区域,提取影像组学特征;采用方差阈值法、方差分析(analysis of variance,ANOVA)及基于5折交叉验证的最小绝对收缩与选择算法(least absolute shrinkage and selection operator,LASSO)筛选最有价值的特征,将筛选出的特征结合逻辑回归(logistic regression,LR)与支持向量机(support vector machine,SVM)两种分类器后进行模型训练,并在测试集中验证。绘制ROC曲线来评估LR模型与SVM模型鉴别PA和BCA的效能。此外,使用Delong Test对模型进行比较,使用决策曲线及校准曲线对模型进行评价。结果分别从ADC、T1WI、T2WI及联合序列(ADC+T1WI+T2WI)图像中得到15、3、15及23个最优特征。在训练集中,基于ADC图、T1WI图、T2WI图、联合模型构建的LR与SVM模型的曲线下面积(area under the curve,AUC)分别为0.955、0.961、0.812、0.813、0.939、0.949、0.994、0.995;基于ADC、T1WI、T2WI及联合序列图像构建的LR模型鉴别诊断PA和BCA的AUC值分别为0.906、0.780、0.868及0.972,SVM模型的AUC值分别为0.924、0.783、0.847及0.959;在训练集中,基于联合序列模型优于基于T1WI或T2WI影像组学模型(P<0.05),与基于ADC影像组学模型差异无统计学意义(P>0.05),联合序列模型的准确率、敏感度及特异度分别为98.6%~98.7%、96.4%~98.4%、98.8%~99.4%,ADC影像组学模型的准确率、敏感度及特异度分别为91.4%~91.8%、75.0%~79.7%、95.7%~98.1%;在测试集中,各模型间的AUC值均无显著性差异(P>0.05)。结论多序列联合模型及ADC影像组学模型鉴别多形性腺瘤和基底细胞腺瘤优于T1WI及T2WI序列,且与ADC影像组学模型比较,联合序列模型具有较高的准确率、敏感度及特异度。

Objective:To explore the value of radiomics models based on ADC,T1WI and T2WI in differentiating salivary gland pleomorphic adenoma(PA)from basal cell adenoma(BCA).Materials and Methods:The MR images of 129 cases with PA and 48 cases with BCA from Jining First People's Hospital from January 2015 to October 2021 were retrospectively analyzed,and then these data were randomly divided into training sets(n=141)and test sets(n=36)at a ratio of 8∶2.The three-dimensional volume region of interest of the tumor was manually delineated on the axial ADC,T1WI and T2WI images,and radiomics features were extracted;the variance threshold method,analysis of variance(ANOVA)and least absolute shrinkage and selection operator(LASSO)based on 5-fold cross validation were used to single out the most valuable radiomic features,and these selected features were combined with two classifiers,logistic regression(LR)and support vector machine(SVM),for training the models,and then the models were verified in the test sets.ROC curve was drawn to evaluate the efficacy of LR and SVM models in differentiating PA from BCA.In addition,the Delong Test was used to compare the models,and the decision curve and calibration curve were used to evaluate the models.Results:A total of 15,3,15 and 23 optimal features were obtained from ADC,T1WI,T2WI and combined sequence(ADC+T1WI+T2WI)image respectively.In the training set,the area under the curve(AUC)of the LR and SVM models constructed based on the ADC map,T1WI map,T2WI map,and joint model were 0.955,0.961,0.812,0.813,0.939,0.949,0.994,0.995,respectively.The AUC values of the LR model constructed based on ADC,T1WI,T2WI and combined sequence image for differential diagnosis of PA and BCA were 0.906,0.780,0.868 and 0.972,respectively,and the AUC values of the SVM model were 0.924,0.783,0.847 and 0.959,respectively.In the training sets,the combined sequence models were better than the T1WI or T2WI-based radiomics models(P<0.05),and there was no significant difference between the combined sequence models and the ADC-based radiomics models(P>0.05),the accuracy,sensitivity and specificity of the combined sequence models were 98.6%-98.7%,96.4%-98.4%,98.8%-99.4%respectively,the accuracy,sensitivity and specificity of the ADC radiomics models were 91.4%-91.8%,75.0%-79.7%,95.7%-98.1%respectively.In the test sets,there was no significant difference in AUC between the models(P>0.05).Conclusions:The combined sequence models and ADC-based radiomics models were better than the T1WI and T2WI-based radiomics models in differentiating pleomorphic adenoma and basal cell adenoma.Compared with ADC-based radiomics models,the combined sequence models had higher accuracy,sensitivity and specificity.