基于影像组学和多序列MRI的胶质瘤相关癫痫预测

Prediction of glioma-associated epilepsy based on radiomics and multi-sequence MRI

基于多序列MRI建立无侵入性、准确、客观的胶质瘤相关癫痫诊断模型。回顾性地收集403例胶质瘤患者的T_(1)WI、T_(2)WI、T_(1)Gd和T_(2)-FLAIR图像。使用预训练的深度学习模型分割包含胶质瘤及瘤周水肿的感兴趣区域,从中提取一阶统计学特征、形态学特征和纹理特征。采用皮尔逊相关系数、递归特征消除等方法进行特征筛选,并将特征分类建立子模型,最终建立含有15个特征的影像模型用于癫痫诊断,在独立测试集上获得0.836的AUC值。在影像模型的基础上加入年龄及性别进行重新建模,获得含有14个特征的临床-影像模型,在独立测试集上获得0.872的AUC值。结合基本临床信息和多序列MRI影像特征的组学模型可以作为胶质瘤相关癫痫的有效诊断工具。

A non-invasive,accurate and objective model is developed based on multi-sequence MRI for diagnosing glioma associated epilepsy.The T_(1)WI,T_(2)WI,T_(1)Gd and T_(2)-FLAIR images of 403 glioma patients are collected retrospectively.A pre trained deep learning model is used to segment the region of interest containing the tumor and peritumoral edema,from which the first-order statistical characteristics,morphological features and texture features are extracted.After feature selection using Pearson correlation coefficient,recursive feature elimination and other methods,a scout model is built for each group of features,and finally a radiomics model containing 15 features is established for epilepsy diagnosis.The radiomics model achieves an AUC of 0.836 on the independent test set.A clinical-radiomics model containing 14 features is further built by incorporating basic clinical information(age and gender)to the radiomics model for remodeling,and it achieves an AUC of 0.872 on the independent test set.The model combining basic clinical information and multi-sequence MRI radiomics signatures can serve as an effective tool for the diagnosis of glioma-associated epilepsy.