Zhang Steven, Richter Jesse, Veale Jonathon, Hieu Phan Vu Minh, Candy Nick, Poonnoose Santosh, Agzarian Marc, To Minh-Son
Faculty of Health and Medical Sciences University of Adelaide Australia.
College of Medicine and Public Health Flinders University Australia.
J Clin Neurosci. 2025 May;135:111118. doi: 10.1016/j.jocn.2025.111118. Epub 2025 Mar 5.
To develop and compare machine learning models for distinguishing low and high grade meningiomas on multiparametric MRI.
Pre-operative T1-weighted(T1), contrast-enhanced T1-weighted(T1CE), T2-weighted, T2 FLAIR, and DWI/ADC MRI sequences of meningiomas performed between 2000 and 2020 were retrospectively collected from a single tertiary hospital dedicated neurosurgical department. Tumours were manually segmented and handcrafted radiomic features were extracted. Deep learning features were extracted using a fine-tuned foundation model. Various oversampling techniques, feature selection algorithms and classifiers were trialled to build Handcrafted radiomics only (HRO) and handcrafted with deep learning radiomics (HDLR) models. Bootstrap was used for internal validation of model performance and calculating confidence intervals of metrices. Discrimination, calibration, feature importance and clinical utility of models were assessed via ROC AUC, calibration curve, Shapley values and decision curve analysis, respectively.
The analysis included 97 low grade and 18 high grade meningiomas. HRO and HDLR models had comparable diagnostic performance, using Random Forest and XGBoost respectively. They achieved mean (95 %CI): ROC AUC 0.825[0.662,0.952] and 0.794[0.662,0.948], specificity 0.913[0.793,0.952] and 0.892[0.796,0.983], sensitivity 0.499[0.204,1] and 0.509[0.225,0.851], NPV 0.909[0.851,0.971] and 0.909[0.851,0.972], and PPV 0.529[0.238,0.924] and 0.465[0.263,0.846], respectively for HRO and HDLR models. HRO and HDLR models selected 11-12 features, with T1 and T1CE having consistent importance.
HRO and HDLR can effectively predict meningioma grades preoperatively. Challenges remain in achieving consistent sensitivity and PPV. Larger, multi-centre studies are warranted to confirm our findings, but it holds promise for improving personalized treatment strategies and patient outcomes in meningioma management. Code is available on Github https://github.com/stephano41/radiomics_ai.
开发并比较用于在多参数磁共振成像（MRI）上区分低级别和高级别脑膜瘤的机器学习模型。
回顾性收集了2000年至2020年间在一家单一的三级医院神经外科进行的脑膜瘤术前T1加权（T1）、对比增强T1加权（T1CE）、T2加权、T2液体衰减反转恢复序列（T2 FLAIR）以及扩散加权成像/表观扩散系数（DWI/ADC）MRI序列。对肿瘤进行手动分割并提取手工制作的放射组学特征。使用微调后的基础模型提取深度学习特征。尝试了各种过采样技术、特征选择算法和分类器，以构建仅手工制作放射组学（HRO）模型和结合深度学习放射组学的手工制作模型（HDLR）。采用自助法对模型性能进行内部验证并计算指标的置信区间。分别通过ROC曲线下面积（ROC AUC）、校准曲线、夏普利值和决策曲线分析评估模型的鉴别能力、校准、特征重要性和临床实用性。
分析纳入了97例低级别和18例高级别脑膜瘤。HRO模型和HDLR模型分别使用随机森林和极端梯度提升（XGBoost），具有可比的诊断性能。它们分别实现了以下均值（95%置信区间）：HRO模型和HDLR模型的ROC AUC分别为0.825[0.662,0.952]和0.794[0.662,0.948]，特异性分别为0.913[0.793,0.952]和0.892[0.796,0.983]，敏感性分别为0.499[0.204,1]和0.509[0.225,0.851]，阴性预测值分别为0.909[0.851,0.971]和0.909[0.851,0.972]，阳性预测值分别为0.529[0.238,0.924]和0.465[0.263,0.846]。HRO模型和HDLR模型选择了11 - 12个特征，T1和T1CE具有一致的重要性。
HRO模型和HDLR模型可以有效地术前预测脑膜瘤级别。在实现一致的敏感性和阳性预测值方面仍存在挑战。需要更大规模的多中心研究来证实我们的发现，但它有望改善脑膜瘤管理中的个性化治疗策略和患者预后。代码可在Github上获取，网址为https://github.com/stephano41/radiomics_ai 。
