影像组学在脑胶质瘤分级中的应用研究

目的利用影像组学与常规磁共振图像对脑胶质瘤进行分级。材料与方法利用BRATS2017公开数据集,自动从图像中的感兴趣区域提取多种特征,包括形态特征、灰度特征、纹理特征等,并运用特征选择方法进行特征选择,最后根据选出的特征对脑胶质瘤的高、低评级进行区分。对支持向量机、决策树与K近邻法等3种分类方法进行比较,也比较了基于遗传算法的包装方法和过滤方法这两种特征选择算法。结果采用过滤方法进行特征选择,支持向量机方法具有最高的准确率91.93%,受试者工作特征曲线下面积(area under the receiver operating characteristic curve,AUC)值为0.90。进一步采用遗传算法进行特征选择,准确率提升至93.33%,AUC值为0.94。结论基于常规磁共振图像,利用影像组学方法,选择合适的特征组合,可以对脑胶质瘤进行自动分级。

Application of radiomics in the grading of brain tumor

Objective:Using radiomics and conventional magnetic resonance images to grade brain tumor.Materials and Methods: This paper uses the open dataset:BRATS2017.Kinds of features,including shape features,intensity features and texture features,were extracted from region of interest and a subset of features were selected by hybrid feature selection methods to distinguish the high grade glioma(HGG)from the low grade glioma(LGG).Three classification algorithms including support vector machine(SVM),decision tree,and K nearest neighborhood were compared.The wrapper method using genetic algorithm(GA)was compared with filter method.Results: SVM was found to exhibit the best accuracy: 91.33％and AUC: 0.90 when using the feature selection method of filter.The accuracy and AUC raise to 93.33％and 0.94 when further using the method of genetic algorithm.Conclusions: We could use radiomics methods and conventional magnetic resonance images to automatically grade brain tumor by selecting suitable features.