Xiao Dongdong, Zhao Zhen, Liu Jun, Wang Xuan, Fu Peng, Le Grange Jehane Michael, Wang Jihua, Guo Xuebing, Zhao Hongyang, Shi Jiawei, Yan Pengfei, Jiang Xiaobing
Department of Neurosurgery, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
Department of Neurosurgery, Taihe Hospital, Hubei University of Medicine, Shiyan, China.
Front Oncol. 2021 Aug 20;11:708040. doi: 10.3389/fonc.2021.708040. eCollection 2021.
Meningioma invasion can be preoperatively recognized by radiomics features, which significantly contributes to treatment decision-making. Here, we aimed to evaluate the comparative performance of radiomics signatures derived from varying regions of interests (ROIs) in predicting BI and ascertaining the optimal width of the peritumoral regions needed for accurate analysis.
Five hundred and five patients from Wuhan Union Hospital (internal cohort) and 214 cases from Taihe Hospital (external validation cohort) pathologically diagnosed as meningioma were included in our study. Feature selection was performed from 1,015 radiomics features respectively obtained from nine different ROIs (brain-tumor interface (BTI)2-5mm; whole tumor; the amalgamation of the two regions) on contrast-enhanced T1-weighted imaging using least-absolute shrinkage and selection operator and random forest. Principal component analysis with varimax rotation was employed for feature reduction. Receiver operator curve was utilized for assessing discrimination of the classifier. Furthermore, clinical index was used to detect the predictive power.
Model obtained from BTI4mm ROI has the maximum AUC in the training set (0.891 (0.85, 0.932)), internal validation set (0.851 (0.743, 0.96)), and external validation set (0.881 (0.833, 0.928)) and displayed statistically significant results between nine radiomics models. The most predictive radiomics features are almost entirely generated from GLCM and GLDM statistics. The addition of PEV to radiomics features (BTI4mm) enhanced model discrimination of invasive meningiomas.
The combined model (radiomics classifier with BTI4mm ROI + PEV) had greater diagnostic performance than other models and its clinical application may positively contribute to the management of meningioma patients.
脑膜瘤侵袭可通过影像组学特征在术前识别，这对治疗决策有重要意义。在此，我们旨在评估从不同感兴趣区域（ROI）得出的影像组学特征在预测脑膜瘤侵袭（BI）及确定准确分析所需瘤周区域的最佳宽度方面的比较性能。
我们纳入了武汉协和医院的505例患者（内部队列）和太和医院的214例病例（外部验证队列），这些患者均经病理诊断为脑膜瘤。使用最小绝对收缩和选择算子以及随机森林，从对比增强T1加权成像上分别从九个不同ROI（脑肿瘤界面（BTI）2 - 5mm；整个肿瘤；两个区域的合并）获得的1015个影像组学特征中进行特征选择。采用具有方差最大化旋转的主成分分析进行特征降维。使用受试者工作特征曲线评估分类器的辨别力。此外，使用临床指标检测预测能力。
从BTI4mm ROI获得的模型在训练集（0.891（0.85，0.932））、内部验证集（0.851（0.743，0.96））和外部验证集（0.881（0.833，0.928））中具有最大的曲线下面积（AUC），并且在九个影像组学模型之间显示出具有统计学意义的结果。最具预测性的影像组学特征几乎完全由灰度共生矩阵（GLCM）和灰度依赖矩阵（GLDM）统计生成。将表观扩散系数（PEV）添加到影像组学特征（BTI4mm）中增强了侵袭性脑膜瘤的模型辨别力。
联合模型（具有BTI4mm ROI + PEV的影像组学分类器）比其他模型具有更高的诊断性能，其临床应用可能对脑膜瘤患者的管理有积极贡献。
