Abstract:

Objective: To develop and validate a preoperative prediction model for meningioma brain invasion using radiomics features derived from multiparametric magnetic resonance imaging (MRI) - based brain - tumor interface (BTI). Methods: A total of 656 meningioma patients diagnosed and treated were included at The First Affiliated Hospital of Nanjing Medical University from September 2014 to April 2023. Using stratified random sampling, patients were randomly divided in a 4∶1 ratio into training set (524 cases) and testing set (132 cases). The training set was used for model construction and optimization, and the testing set for evaluating generalization ability. All patients underwent preoperative MRI examination including axial T₁WI, enhanced T₁WI and T₂WI. After image preprocessing and segmentation, the meningioma region of interest was identified, and BTI with thicknesses of 0.80, 1.00 and 1.20 cm were constructed. Radiomics features were extracted from the regions of interest (ROI) across the 3 sequences. Following single - value elimination and interclass correlation coefficient [ICC (2, k) > 0.90] stability screening, features were selected using five - fold cross - validated least absolute shrinkage and selection operator (LASSOCV). Six machine learning (ML) algorithms, including light gradient boosting machine (LightGBM), Logistic regression (LR), multilayer perceptron (MLP), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) were utilized to build predictive models. The performance of each model was assessed using receiver operating characteristic (ROC) curve and the area under the curve (AUC). The significance of differences between ROC curves were compared using the Delong test. Decision curve analysis (DCA) was performed to evaluate the clinical net benefit of the models across different threshold probabilities. Results: Among the 656 meningioma patients, 152 cases (23.17%) exhibited brain invasion, with 123 cases (23.47%) in the training set and 29 cases (21.97%) in the testing set. Through five - fold cross - validation in the training set and evaluation in the testing set, comparative analysis of the predictive performance of 18 model - thickness combinations (6 ML algorithms × 3 BTI thicknesses) showed that the XGBoost model constructed with a 1.00 cm BTI thickness demonstrated exceptional performance. This model achieved an AUC of 0.913 (95%CI: 0.886-0.937, P = 0.000), accuracy of 0.86, sensitivity of 0.77, and specificity of 0.88 in the training set; and an AUC of 0.897 (95%CI: 0.821- 0.961, P = 0.000), accuracy of 0.90, sensitivity of 0.72, and specificity of 0.95 in the testing set. Further Delong test showed that this model's AUC was significantly higher than all other models (P < 0.05, for all). DCA showed that this model demonstrated the best clinical utility with the highest net benefit area in both the training set (0.087) and the testing set (0.094). Conclusions: The XGBoost model based on 1.00 cm BTI exhibited outstanding predictive performance, providing an accurate and reliable non - invasive method for preoperative evaluation of meningioma brain invasion. This method offers substantial clinical utility in facilitating personalized surgical planning, risk assessment, and prognosis evaluation.

Key words: Meningioma, Neoplasm invasiveness, Magnetic resonance imaging, Radiomics (not in MeSH), Machine learning, ROC curve

摘要：

目的: 开发并验证一种基于术前多参数MRI的脑-肿瘤界面影像组学脑膜瘤脑侵犯无创性预测模型。方法: 纳入2014年9月至2023年4月南京医科大学第一附属医院收治的656例脑膜瘤患者，按照4∶1比例随机分为训练集(524例)和测试集(132例)，训练集用于预测模型的构建和优化，测试集用于模型泛化能力的评估。术前均行MRI检查(包括横断面T₁WI、增强T₁WI和T₂WI)，图像经预处理和分割后确定脑膜瘤感兴趣区，构建厚度为0.80、1.00和1.20 cm的脑-肿瘤界面。从3个序列感兴趣区中提取影像组学特征，经单一值筛除、稳定性筛选组内相关系数[ICC(2，k)> 0.90]后，采用五折交叉验证最小绝对收缩和选择算子算法筛选特征，采用轻量梯度提升机、Logistic回归、多层感知器、随机森林、支持向量机和极端梯度提升算法(XGBoost)共6种机器学习算法构建脑侵犯预测模型，绘制受试者工作特征(ROC)曲线并计算曲线下面积评估模型预测效能，Delong检验比较不同模型的曲线下面积，决策曲线分析评估不同模型在不同阈值概率下的临床净收益。结果: 共656例脑膜瘤患者中152例(23.17%)存在脑侵犯，训练集有123例(23.47%)、测试集有29例(21.97%)。通过训练集五折交叉验证和测试集评估，比较18个模型-厚度组合(6种机器学习算法× 3种脑-肿瘤界面厚度)的预测效能，1.00 cm脑-肿瘤界面的XGBoost模型表现优异，其训练集的曲线下面积为0.913 (95%CI：0.886 ~ 0.937，P = 0.000)，准确度为0.86、灵敏度为0.77、特异度为0.88；测试集的曲线下面积为0.897(95%CI：0.821 ~ 0.961，P = 0.000)，准确度为0.90、灵敏度为0.72、特异度为0.95；Delong检验显示，该模型曲线下面积大于其他所有模型(均P < 0.05)。决策曲线分析显示，该模型在训练集(决策曲线分析正净收益面积0.087)和测试集(决策曲线分析正净收益面积0.094)中均表现出最佳的临床净收益。结论: 基于1.00 cm脑-肿瘤界面的XGBoost模型展现出优异的预测效能，为脑膜瘤脑侵犯的术前评估提供一种准确、可靠的无创性预测方法，对术前制定个性化手术方案、评估手术风险和预后具有重要临床意义。

关键词: 脑膜瘤, 肿瘤浸润, 磁共振成像, 影像组学(非MeSH词), 机器学习, ROC曲线