对比增强T1加权成像纹理分析鉴别小脑毛细胞型星形细胞瘤与血管母细胞瘤

Differentiation of pilocytic astrocytoma and hemangioblastoma in cerebellum by texture analysis based on contrast-enhanced T1-weighted imaging

目的:探讨基于对比增强T1加权成像(CE-T1W)的纹理分析在小脑毛细胞型星形细胞瘤(PA)与血管母细胞瘤(HB)鉴别中的价值。方法:收集本院2014年1月至2021年6月经组织病理学确诊的小脑32例PA和38例HB,提取CE-T1W序列肿瘤的直方图分析、灰度共生矩阵、灰度游程矩阵、绝对梯度、自回归模型和小波变换6种纹理特征,以筛选的纹理特征与影像特征建立支持向量机(support vector machine,SVM)分类模型,分别使用线性核、多项式核和径向基函数核。采用受试者工作特征(ROC)曲线评估模型效能。结果:279个特征参数中筛选出10个,2组影像特征中大囊小结节、流空血管、实性强化均匀比较差异有统计学意义(P<0.01),而是否呈实性、有无囊内分隔、囊壁是否强化、有无瘤周水肿比较差异无统计学意义(P>0.05)。纹理特征建立的SVM分类模型中,使用多项式核的效能最佳,敏感度、特异度、准确率及AUC分别为87.5%、94.7%、91.4%及0.897,优于以3个影像特征建立的3种SVM分类模型(P<0.05)。结论:基于CE-T1W的纹理分析可有效鉴别小脑PA与HB,且效能优于影像特征。

Objective:To investigate the value of texture analysis based on contrast-enhanced T1-weighted imaging(CE-T1W)in differentiating pilocytic astrocytoma(PA)and hemangioblastoma(HB)in cerebellum.Methods:Thirty-two patients with PA and 38 patients with HB diagnosed by histopathology from Jan 2014 to Jun 2021 in our hospital were selected.And six texture features,including histogram analysis(HA),gray-level co-occurrence matrix(GLCM),gray-level run length matrix(GLRLM),absolute gradient(AG),autoregressive mode(AR)and wavelet transform(WT),were extracted from CE-T1W.Then,support vector machine(SVM)classification model,including linear kernel(LK),polynomial kernel(PK)and radial basis function kernel(RBFK),was established by selected features and imaging features.Finally,receiver operating characteristic(ROC)curves were used to evaluate model performance.Results:Ten of the 279 feature parameters were selected,imaging features of large cysts with small nodules,flow viod vessels and solid uniform enhancement showed statistical difference between the two groups(P<0.01),while there was no significant difference between the two groups in terms of whether they were solid,whether there was intracapsular separation,whether the cystic wall was enhanced and whether there was peritumoral edema(P>0.05).SVM classification model with PK established by texture features had the best performance,which was better than three SVM classification models based on three imaging features(P<0.05),with sensitivity,specificity,accuracy and AUC of 87.5%,94.7%,91.4%and 0.897,respectively.Conclusion:Texture analysis based on CE-T1W can effectively distinguish PA and HB in the cerebellum,and the efficiency is better than that of imaging features.