The Research Of Feature Selection Algorithm Based Sparse Optimization On Brain MRI Image

Feature selection plays an important role in the study of functional magnetic resonance imaging(fMRI).It can help identify abnormalities of brain function and structure caused by the disease or detect brain activation corresponding to certain tasks.However,the feature dimension of fMRI data is high,the number of samples is small,and the features have strong correlation.These properties give a strong challenge to feature selection methods.Most traditional feature selection methods mainly aim to construct a concise classifier.They often select only a minimum subset of features,ignoring those correlated or redundant but informative features,or involve some noisy features which may not affect the prediction accuracy.In addition,the traditional feature selection methods may select significantly various features on account of small changes of the samples(even from the same system).In recent years,the stability selection method attracts much attention,because of good performance of feature selection: 1)guarantee the false positive for finite sample;2)providing a transparent principle to choose a proper amount of regularization for structure estimation.However,due to without fully addressing the structural information,stability selection may produce many false negatives.In this paper,we combine the stability selection framework with the sparse structure model and propose a method,structural stability selection.Particularly,we propose a "block correlated subsampling" method for the functional magnetic resonance data,which shows strong spatial local correlation.This method can effectively link the stability selection and the sparse structure model to estimate the structure of the discriminative features.The experiments of simulation data and real fMRI data all show that our new method can effectively control the false positive and the false negative simultaneously.Moreover,compared with other feature selection methods,our method is more consistent with the prior hypothesis,that is,the relevant discriminative voxels are grouped into several distributed clusters.