基于^(18)F-FET PET影像组学分析预测脑胶质瘤IDH1基因表型模型的建立与验证

Establishment and validation of ^(18)F-FET PET radiomic features-based model in predicting IDH1 genotype in gliomas

目的构建基于O-(2-^(18)F-氟代乙基)-L-酪氨酸(^(18)F-FET)PET图像影像组学特征的回归分析模型,探讨其对未经治疗的脑胶质瘤患者异柠檬酸脱氢酶1(IDH1)基因表型的预测效能。方法回顾性分析2017年11月至2019年2月间复旦大学附属华山医院经病理学证实的58例脑胶质瘤患者[男36例、女22例,年龄(41.8±15.1)岁]的^(18)F-FET PET/CT脑显像数据,应用PyRadiomics软件包提取105个影像组学特征,使用最小绝对收缩和选择算子(LASSO)回归构建预测模型,计算每个病灶的影像组学评分(RS),使用受试者工作特征(ROC)曲线量化模型RS的预测效能,并与^(18)F-FET半定量参数{肿瘤靶本比[TBR,包括最大TBR(TBR_(max))、TBR峰值(TBR_(peak))和平均TBR(TBR_(mean))]、肿瘤代谢体积(MTV)及病灶总代谢摄取(TLU)}对IDH1基因表型的预测效能进行对比(Delong检验)。结果LASSO回归模型共纳入7个影像组学特征,分别为最大二维切片直径、一阶最大特征值、一阶灰度值范围、灰度共生矩阵_能量、灰度共生矩阵_反差方差、灰度相关矩阵_熵、灰度相关矩阵_大相关低灰度级增强。构建的LASSO回归模型对IDH1基因表型(突变型20例,野生型38例)的预测效能准确性为81.0%(47/58),灵敏度为65.0%(13/20),特异性为89.5%(34/38),曲线下面积(AUC)为0.842;^(18)F-FET半定量参数中TLU诊断效能较优,其诊断准确性为60.3%(35/58),灵敏度为85.0%(17/20),特异性为47.4%(18/38),AUC为0.661;LASSO回归模型对IDH1基因表型的诊断效能优于传统参数(z=3.426,P<0.01)。结论基于^(18)F-FET PET脑显像的影像组学分析可提高对未经治疗的脑胶质瘤患者IDH1基因表型的预测效能。

Objective To establish O-(2-[^(18)F]fluoroethyl)-L-tyrosine(^(18)F-FET)PET radiomics features-based model and investigate its predictive efficacy for isocitrate dehydrogenase type 1(IDH1)genotyping in untreated gliomas.Methods From November 2017 to February 2019,58 pathologically confirmed glioma patients(36 males,22 females;age(41.8±15.1)years)with preoperative ^(18)F-FET PET/CT imaging in Huashan Hospital,Fudan University were retrospectively enrolled.PyRadiomics software package was used to extract 105 radiomics features.Least absolute shrinkage and selection operator(LASSO)algorithm with 5-fold cross-validation was used to build the logistic regression model.And radiomic scores(RS)of each lesion were calculated according to their weighted coefficients.The area under the receiver operating characteristic(ROC)curve was used for evaluating the predictive efficacy for IDH1 prediction.The predictive efficacies of radiomics model and traditional semi-quantitative parameters including tumor-to-background ratio(TBR;maximum TBR(TBR_(max)),mean TBR(TBR_(mean)),peak TBR(TBR_(peak))),metabolic tumor volume(MTV)and total lesion tracer uptake(TLU),were compared by Delong test.Results Seven radiomics features including maximum 2-dimensional(2D)diameter slice,first order_maximum,first order_range,gray level co-occurrence matrix(GLCM)_joint energy,GLCM_inverse variance,gray level dependence matrix(GLDM)_dependence entropy and GLDM_large dependence low gray level emphasis were selected for the LASSO regression model building and RS calculation.ROC analysis results showed that the predictive accuracy of RS for IDH1 genotyping(mutation,n=20;wild-type,n=38)was 81.0%(47/58),with sensitivity of 65.0%(13/20),specificity of 89.5%(34/38),and area under curve(AUC)of 0.842,respectively.The traditional ^(18)F-FET semi-quantitative parameter TLU ranked the second regarding the diagnostic performance,with accuracy of 60.3%(35/58),sensitivity of 85.0%(17/20),specificity of 47.4%(18/38),and AUC of 0.661(z=3.426,P<0.01).Conclusion Radiomics analysis based on ^(18)F-FET PET images can improve the predictive efficacy for IDH1 genotyping in untreated adult glioma patients.