Department of Neurosurgery, Xiangya Hospital, Central South University, Changsha, 410008, China.
Department of Neurosurgery, Yale School of Medicine, New Haven, CT, 06520, USA.
Eur Radiol. 2020 Jun;30(6):3073-3082. doi: 10.1007/s00330-019-06632-8. Epub 2020 Feb 5.
To establish a quantitative MR model that uses clinically relevant features of tumor location and tumor volume to differentiate lower grade glioma (LRGG, grades II and III) and glioblastoma (GBM, grade IV).
We extracted tumor location and tumor volume (enhancing tumor, non-enhancing tumor, peritumor edema) features from 229 The Cancer Genome Atlas (TCGA)-LGG and TCGA-GBM cases. Through two sampling strategies, i.e., institution-based sampling and repeat random sampling (10 times, 70% training set vs 30% validation set), LASSO (least absolute shrinkage and selection operator) regression and nine-machine learning method-based models were established and evaluated.
Principal component analysis of 229 TCGA-LGG and TCGA-GBM cases suggested that the LRGG and GBM cases could be differentiated by extracted features. For nine machine learning methods, stack modeling and support vector machine achieved the highest performance (institution-based sampling validation set, AUC > 0.900, classifier accuracy > 0.790; repeat random sampling, average validation set AUC > 0.930, classifier accuracy > 0.850). For the LASSO method, regression model based on tumor frontal lobe percentage and enhancing and non-enhancing tumor volume achieved the highest performance (institution-based sampling validation set, AUC 0.909, classifier accuracy 0.830). The formula for the best performance of the LASSO model was established.
Computer-generated, clinically meaningful MRI features of tumor location and component volumes resulted in models with high performance (validation set AUC > 0.900, classifier accuracy > 0.790) to differentiate lower grade glioma and glioblastoma.
• Lower grade glioma and glioblastoma have significant different location and component volume distributions. • We built machine learning prediction models that could help accurately differentiate lower grade gliomas and GBM cases. We introduced a fast evaluation model for possible clinical differentiation and further analysis.
建立一种定量磁共振（MR）模型，使用肿瘤位置和肿瘤体积的临床相关特征来区分低级别胶质瘤（LRGG，II 级和 III 级）和胶质母细胞瘤（GBM，IV 级）。
我们从 229 例癌症基因组图谱（TCGA）-LGG 和 TCGA-GBM 病例中提取了肿瘤位置和肿瘤体积（增强肿瘤、非增强肿瘤、瘤周水肿）特征。通过两种采样策略，即机构采样和重复随机采样（10 次，70%训练集与 30%验证集），建立并评估了 LASSO（最小绝对收缩和选择算子）回归和九种机器学习方法的模型。
对 229 例 TCGA-LGG 和 TCGA-GBM 病例进行主成分分析表明，提取的特征可区分 LRGG 和 GBM 病例。对于九种机器学习方法，堆叠建模和支持向量机取得了最高的性能（机构采样验证集，AUC>0.900，分类器准确性>0.790；重复随机采样，平均验证集 AUC>0.930，分类器准确性>0.850）。对于 LASSO 方法，基于肿瘤额叶百分比和增强与非增强肿瘤体积的回归模型取得了最高的性能（机构采样验证集，AUC 0.909，分类器准确性 0.830）。建立了 LASSO 模型最佳性能的公式。
肿瘤位置和成分体积的计算机生成的、具有临床意义的 MRI 特征产生了具有高性能（验证集 AUC>0.900，分类器准确性>0.790）的模型，可区分低级别胶质瘤和胶质母细胞瘤。
• 低级别胶质瘤和胶质母细胞瘤的位置和成分体积分布有显著差异。• 我们构建了机器学习预测模型，可以帮助准确区分低级别胶质瘤和 GBM 病例。我们引入了一种快速评估模型，用于可能的临床区分和进一步分析。
