Annu Int Conf IEEE Eng Med Biol Soc. 2017 Jul;2017:616-619. doi: 10.1109/EMBC.2017.8036900.
Radiomics can convert digital images to mineable data by extracting a huge number of image features. Because of the high dimensions of radiomics features, feature selection is a very important step which affects the performance of the final prediction or classification. In this paper, we propose a feature selection criterion for radiomics analysis of glioma based on Magnetic Resonance Image (MRI). The proposed method named as minimum Redundancy, Maximum Relevance and Maximum Sparse Representation Coefficient (mRMRMSRC) criterion, which take three factors into consideration at the same time: relevance between features and labels with or without the influence of all other features, and redundancy between each couple of features. Different from traditional feature selection method, the mRMRMSRC manifests the best performance compared with the methods based on sparse representation coefficient (SRC), minimum redundancy maximum relevance (mRMR), F_score and ReliefF. We conducted our methods on glioma Isocitrate Dehydrogenase 1 (IDH1) estimation. The experiment showed that mRMRMSRC produced area under the ROC curve (AUC) of 90% compared with 77%-89% of state-of-art methods.
放射组学可通过提取大量图像特征将数字图像转化为可挖掘的数据。由于放射组学特征的高维度性，特征选择是影响最终预测或分类性能的非常重要的一步。在本文中，我们提出了一种基于磁共振成像（MRI）的胶质瘤放射组学分析的特征选择标准。所提出的方法名为最小冗余、最大相关性和最大稀疏表示系数（mRMRMSRC）标准，该标准同时考虑了三个因素：特征与标签之间的相关性（有无所有其他特征的影响）以及每对特征之间的冗余性。与传统特征选择方法不同，与基于稀疏表示系数（SRC）、最小冗余最大相关性（mRMR）、F_score和ReliefF的方法相比，mRMRMSRC表现出最佳性能。我们将我们的方法应用于胶质瘤异柠檬酸脱氢酶1（IDH1）估计。实验表明，mRMRMSRC产生的ROC曲线下面积（AUC）为90%，而现有方法的AUC为77%-89%。
