The non-invasive prediction analysis of radiomic features from 18F-FET PET/CT dual imaging modality for tumor grading in untreated adult gliomas
华涛 1周维燕 1周支瑞 2黄琪 1李明 1朱毓华 1管一晖1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 复旦大学附属华山医院核医学科/PET中心,上海 200235
2. 复旦大学附属华山医院放射治疗中心,上海 200040
折叠
摘要
目的:基于18F-FET正电子发射体层成像(positron emission tomography,PET)/计算机体层成像(computed tomography,CT)双模态影像组学特征构建模型,并对成人胶质瘤的病理学分级进行非侵入性预测分析.方法:回顾并分析58例经组织病理学检查证实的未治疗成年胶质瘤患者的18F-FET PET/CT影像学数据,根据病理学分级将患者分成低级别胶质瘤组[世界卫生组织(World Health Organization,WHO)Ⅱ级,共计32例]和高级别胶质瘤组(WHO Ⅲ级13例、WHO Ⅳ级13例,共计26例).在PET、CT影像模态中分别提取105个影像组学特征参数进行分析,采用基于R语言机器学习算法的5折交叉验证最小绝对收缩与选择算子(least absolute shrinkage and selection operator,LASSO)回归分析方法,构建成人胶质瘤病理学分级的3个独立的影像组学预测模型(PET-Rad模型、CT-Rad模型和PET/CT-Rad模型),然后再采用全子集回归对影像组学预测模型进行校正.采用受试者工作特征曲线的曲线下面积(area under curve,AUC)对预测模型进行评价.结果:基于4个18F-FETPET影像组学参数建立PET-Rad模型的AUC为0.845(95％CI 0.726～0.927);基于3个CT影像组学参数构建的CT-Rad模型的AUC为0.802(95％CI 0.676～0.895);而联合3个CT和2个PET影像组学特征的18F-FET PET/CT-Rad模型的AUC为0.901(95％CI 0.794～0.964),准确度为86.21％.DeLong检验结果显示PET/CT-Rad模型优于CT-Rad模型(P=0.032).尽管PET/CT-Rad模型效能优于PET-Rad模型,但差异无统计学意义(P=0.146).构建胶质瘤病理学级别预测的PET/CT-Rad模型中,3个CT影像组学参数分别为firstorder_10Percentile、glrlm_LowGrayLevelRunEmphasis、ngtdm_Busyness,其中glrlm_LowGrayLevelRunEmphasis是最重要的预测变量,其相对重要性为30.97％;2个PET组学特征为firstorder_Maximum和ngtdm_Contrast,其相对重要性分别为21.99％、21.01％.结论:基于18F-FET PET与CT双模态影像组学特征相结合的预测模型能够有效地预测未经治疗成人胶质瘤的病理学分级,可为临床诊治决策提供依据和重要参考.
Abstract
Objective:To build radiomics models based on the features from 18F-FET positron emission tomography(PET)/computed tomography(CT)dual imaging modality,and investigate the predictive efficacy for tumor grading in untreated adult gliomas.Methods:The 18F-FET PET/CT imaging data of 58 histopathologically confirmed untreated adult gliomas were retrospectively analyzed.Based on pathological grading,the patients were divided into low-grade glioma groups[World Health Organization(WHO)Grade Ⅱ,32 cases in total]and high-grade gliomas(13 cases for WHO grade Ⅲ,13 cases for WHO grade Ⅳ,26 cases in total).105 radiomics features were extracted from PET and CT modalities respectively.Five-fold cross-validation least absolute shrinkage and selection operator(LASSO)regression analysis based on R-language machine learning algorithm and all-subset regression were adopted to filter and optimize the identify the optimal feature combinations with higher distinguishing power for glioma grading.Three independent radiomics prediction models(PET-Rad model,CT-Rad model and PET/CT-Rad model)for adult glioma pathological grading were constructed.The area under the receiver operating characteristic curve(AUC)was used to evaluate the prediction model.Results:The AUC of PET-Rad model based on four 18F-FET PET radiomics features was 0.845(95％CI 0.726-0.927)and the AUC of CT-Rad model consisting of three CT radiomics features was 0.802(95％CI 0.676-0.895).The AUC of the 18F-FET PET/CT-Rad model combined with three CT and two PET radiomics features was 0.901(95％CI 0.794-0.964),and the accuracy was 86.21％.DeLong test results showed that PET/CT-Rad model was superior to CT-Rad model(P=0.032).The efficacy of PET/CT-Rad model was better than that of PET-Rad model,but there was no statistical difference(P=0.146).The three CT imaging parameters in PET/CT-Rad model were firstorder 10Percentile,glrlm_LowGrayLevelRunEmphasis,and ngtdm_Busyness,while Glrlm_LowGrayLevelRunEmphasis was the most important predictive variable with a relative importance of 30.97％.The relative importance of the other two selected PET radiomic features firstorder_Maximum and ngtdm_Contrast was 21.99％and 21.01％,respectively.Conclusion:The non-invasive prediction model based on the combination of integrated 18F-FET PET and CT modality radiomic features can effectively help grading the untreated adult glioma,facilitating clinical decision-making for patients with varied glioma grades.
万方数据