基于DWI生境分析与Transformer深度学习术前预测脑膜瘤孕激素受体表达
本研究回顾性纳入两个中心经手术病理证实的脑膜瘤患者,利用DWI生境分析量化肿瘤内异质性并提取生态多样性特征构建ITH模型,同时采用Transformer深度学习模型预测,两者经逻辑回归融合为联合模型。联合模型在训练集和测试集的AUC分别为0.910(95%CI: 0.866, 0.947)和0.902(95%CI: 0.831, 0.957),显著优于临床模型,可用于脑膜瘤PR表达的术前无创预测。
要点
- 联合DWI生境分析与Transformer深度学习可术前无创预测脑膜瘤PR表达。
- 联合模型预测性能最佳,训练集AUC 0.910,测试集AUC 0.902。
- 联合模型显著优于仅基于临床和影像特征的临床模型。
核心结论
- 基于DWI的生境分析ITH模型与Transformer深度学习模型联合,在脑膜瘤PR表达术前预测中表现出良好性能,并显著优于临床模型。
研究设计与样本
- 回顾性收集两个中心2014年1月至2024年4月经手术病理证实的脑膜瘤患者数据,构建并验证预测模型。
- 采用DWI生境分析手动勾画肿瘤VOI并提取生态多样性特征,经ICC、Mann-Whitney U检验、SVM-RFE和5折交叉验证筛选特征,用随机森林构建ITH模型输出PR表达预测概率。
- 将VOI图像输入Transformer深度学习模型生成预测概率,并与ITH模型预测概率经逻辑回归融合构建联合模型。
- 临床和影像特征经单因素及逐步逻辑回归筛选,用多因素逻辑回归构建临床模型。
- 以ROC曲线下面积AUC评估模型性能,并在独立外部测试集验证。
诊断标准与关键数值
- 最终保留13个生境特征和4个临床及影像特征与脑膜瘤PR表达高度相关。
- 联合模型在训练集AUC为0.910(95%CI: 0.866, 0.947),在测试集AUC为0.902(95%CI: 0.831, 0.957)。
- 联合模型显著优于临床模型。
临床应用与阅片要点
- 可将DWI生境分析与Transformer深度学习联合模型用于脑膜瘤术前PR表达的无创预测,辅助临床决策。
- 模型基于手动勾画的肿瘤VOI,日常阅片需准确勾画肿瘤范围以保证特征提取可靠性。
局限性
- 本研究为回顾性设计,可能存在选择偏倚。
- 模型基于两个中心数据构建和验证,外部泛化能力仍需多中心更大样本验证。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: This study aims to develop and validate a model based on diffusion-weighted imaging (DWI) habitat analysis and Transformer-based deep learning (DL) for preoperative prediction of progesterone receptor (PR) expression in meningiomas.
MATERIALS AND METHODS: This study retrospectively collected data from patients with meningiomas confirmed by surgical pathology from two centers during the period from January 2014 to April 2024. Habitat analysis was used for the non-invasive quantitative measurement of intratumoral heterogeneity (ITH). Ecological diversity features were extracted from the manually segmented tumor volume of interest (VOI), the optimal features were retained using intraclass and interclass correlation coefficients (ICCs), Mann-Whitney U test, support vector machine recursive feature elimination and 5-fold cross-validation, and a random forest (RF) was used to construct the ITH model, which outputs the predicted probability of PR expression in meningiomas. In addition, VOI images were fed into a Transformer-based DL model to generate predicted probabilities. Finally, the predicted probabilities from the ITH and DL models were combined, and a combined model was constructed using logistic regression. Clinical and imaging features were screened via univariate logistic regression and stepwise logistic regression, and a clinical model was constructed using multivariable logistic regression. The predictive performance of the model was evaluated using the area under the receiver operating characteristic curve (AUC) and validated in an independent external test set.
RESULTS: Ultimately, 13 habitat features and 4 clinical and imaging features were retained as highly correlated with meningioma PR expression. In both the training and testing sets, the combined model demonstrated the best predictive performance, with AUC values of 0.910 (95 % CI: 0.866, 0.947) and 0.902 (95 % CI: 0.831, 0.957), respectively, and it significantly outperformed the clinical model.
CONCLUSION: The combined model constructed by combining DWI-based ITH and Transformer-based DL demonstrated good predictive performance in preoperative prediction of PR expression in meningiomas.
引用
- Gui Y, Diao J, Zhang F, et al. Preoperative prediction of progesterone receptor expression in meningiomas based on diffusion weighted imaging habitat analysis and Transformer-Based deep learning. Eur J Radiol. 2026. PMID: 42476061. doi:10.1016/j.ejrad.2026.113094
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。