Robust multiple external stimuli classification in functional MRI images

Functional magnetic resonance imaging (fMRI) is an effective imaging modality in analyzing neural activation of human brain to external stimuli. The general analysis process includes feature selection and classification. Feature selection can be limited to specific anatomical regions or be implemented by using univariate or multi-variate statistical methods. However, both of these introduce some limitations for decoding different brain states. This paper introduces a new hybrid univariate and multivariate feature selection scheme, and explores different feature selection methods, i.e. orthogonal matching pursuit (OMP), smoothed Io (SL0) and principal feature analysis (PFA). The selected feature voxels are then classified by support vector machine (SVM) to produce predictions of perceived stimuli. Experimental results on Haxby single subject multiple stimuli dataset show the high effectiveness of this new multivoxel pattern analysis (MVPA) method.