New dimension reduction technique for brain decoding

Afrasiyabi, Arman
A new architecture for dimension reduction, analyzing and decoding the discriminative information, distributed in function Magnetic Resonance Imaging (fMRI) data, is proposed. This architecture called Sparse Temporal Mesh Model (STMM) which consists of three phases with a visualization tool. In phase A, a univariate voxel selection method, based on the assumption that voxels are independent, is used to select the informative voxels among the whole brain voxels. For this purpose, one of feature selection methods namely one way analysis of variance (ANOVA) or mutual information (MI) is employed. Then, in phase B, a multivariate voxel selection method, based on the multivariate form of the brain, known as recursive feature elimination (RFE) is employed. The last phase, phase C, contains two parts. In phase C.1, a local mesh with fix size around each voxel called seed voxel is formed. Next, the relationships, called arc weights, between the seed voxel and the neighbouring voxels are estimated. In phase C.2, ANOVA feature selection method is used to eliminate the unnecessary arc weights. Additionally, a visualization tool known as t Distributed Stochastic Neighbor Embedding (tSNE) is used to analyse the effect of each phase. The results indicate that STMM can successfully use for brain decoding purpose.
Citation Formats
A. Afrasiyabi, “New dimension reduction technique for brain decoding,” M.S. - Master of Science, Middle East Technical University, 2015.