基于正则化Softmax回归的全脑功能性磁共振成像数据特征选择框架*

针对功能性磁共振成像(fMRI)数据高维小样本特性给分类模型带来的过拟合问题,文中基于Softmax回归提出结合L2正则与L1正则的全脑fMRI数据特征选择框架.首先,基于大脑认知的特点,将全脑分成感兴趣区域和非感兴趣区域.然后,使用可以缩小权值系数的L2正则对感兴趣区域建模以选出感兴趣区域的全部体素,使用具有稀疏作用的L1正则对非感兴趣区域建模以选出非感兴趣区域中的激活体素.最后,结合感兴趣区域和非感兴趣区域的体素构成全脑fMRI数据的正则化Softmax回归模型.在Haxby数据集上的实验表明,L2与L1的正则化策略可有效提升全脑分类的准确率.

Feature Selection Framework of Whole-Brain Functional Magnetic Resonance Imaging Data Based on Regularized Softmax Regression

To solve the classification model overfitting problem caused by the high dimension and small sample properties of functional magnetic resonance imaging (fMRI) data, a feature selection framework of whole-brain fMRI data combining L1-norm regularization and L2-norm regularization in softmax regression is proposed. Firstly, the whole brain is divided into the region of interest (ROI) and the region of non-interest (RONI) in terms of the characteristics of brain cognition. Then, L2-norm regularization shrinking the weighting coefficients is used to model all voxels in ROI while L1-norm regularization with a sparse effect is employed for modeling the activated voxels in RONI. Finally, the regularized softmax regression model of whole-brain fMRI data is constructed by integrating all voxels in ROI and the activated voxels in RONI. The experimental results on Haxby datasets show that the regularization strategies of L2-norm and L1-norm effectively improve the whole-brain classification performance compared to some other methods.