Recently, sparse representation has received a great deal of attention in voxel selection and decoding analysis of functional magnetic resonance imaging (fMRI) data. Due to contamination of large noise in fMRI data, t...
详细信息
Recently, sparse representation has received a great deal of attention in voxel selection and decoding analysis of functional magnetic resonance imaging (fMRI) data. Due to contamination of large noise in fMRI data, the conventional sparse representation methods may not get stable results. Moreover, the selected activated brain regions may lose clustering effects and are less biologically interpretable. In order to overcome the above mentioned problems, we exploit the error-tolerant formulation of sparse representation and introduce multi-dimensional derivative constraints (smoothness constraints) in optimization. Two new methods are proposed in this paper. One is robust voxel selection with multi-dimensional constraint (RVSMDC). With the error-tolerant formulation and smoothness constraints on regression coefficients, RVSMDC is robust against noise/error and achieves clustering effects. To directly decode neural activities from fMRI data, we also proposed robust sparse decoding with multi-dimensional constraints (RSDMDC), which minimize the regression error of fMRI data to the task function with sparsity and smoothness constraints on regression coefficients. Numerical results validate the effectiveness of the two proposed methods.
暂无评论