基于预建模的影像组学特征选择方法  

Feature selection method of radiomics based on scout models

作　　者：淡一波 陶虹月[2] 王一达 王成龙[1] 陈爽[2] 杨光[1] DAN Yi-bo;TAO Hong-yue;WANG Yi-da;WANG Cheng-long;CHEN Shuang;YANG Guang(Shanghai Key Laboratory of Magnetic Resonance,Department of Physics,East China Normal University,Shanghai 200062,China;Huashan Hospital,Fudan University,Shanghai 200040,China)

机构地区：[1]华东师范大学,上海市磁共振重点实验室,上海200062 [2]复旦大学附属华山医院,上海200040

出　　处：《信息技术》2022年第4期1-6,12,共7页Information Technology

基　　金：国家自然科学基金资助项目(61731009)。

摘　　要：提出一种新的、基于预建模的影像组学特征选择方法,可用于影像组学研究中的特征选择。在预建模过程中,首先将提取的影像特征按照特征类别,如一阶、纹理特征等,划分为特征子集,然后采用五折交叉验证的方法,用训练集中每个特征子集中的特征分别建立分类模型。当基于某个特征子集建立的模型的交叉验证AUC高于特定阈值时,该模型中使用的所有子类特征即被选择用于最终的建模。在BraTS2019公开数据集和踝关节不稳定性数据集上对该方法进行了实验,在独立测试集上的AUC分别达到0.947和0.767,优于传统算法的结果。A new and pre-modeling based radiomics feature selection method,which can be used for feature selection in radiomics studies is raised.In the pre-modeling process,firstly,the image features extracted are divided into subsets by their categories,such as first-order and texture features.Then,5-fold validation is used to build and evaluate models with features from each feature subset.When the cross-validation AUC of a model is higher than a certain threshold,all the features retained in this model are kept for the final model building.Experiments are carried out on the BraTS2019 open data set and an ankle arthritis data set,and the AUC on independent test set reaches 0.947 and 0.767,respectively,which are better than those achieved with the previous algorithms.

关 键 词：特征选择 降维 影像组学 预建模 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]