Effect of voxel selection on temporal mesh model for brain decoding

In this study, we combine a voxel selection method with temporal mesh model to decode the discriminative information distributed in functional Magnetic Resonance Imaging (fMRI) data. We first employ one way Analysis of Variance (ANOVA) feature selection to select the most informative voxels. Then, we form meshes around selected voxels with their spatial and functional neighbors by employing the Mesh Model with Temporal Measurements (MM-TM). We estimate the arc weights of meshes, which represent the relationships among voxels within the selected neighborhood. In order to get rid of the redundant relationships, we prune the estimated mesh weights using ANOVA. By doing so, we obtain a sparse representation of discriminative information in the brain. Finally, we train k-Nearest Neighbor (kNN) and Support Vector Machine (SVM) classifiers using the sparse mesh arc weights. We used fMRI recordings from a visual object recognition experiment. Our results show that employing the selected voxels in classification performs better than employing all voxels in the brain. Moreover, mesh arc weights formed around selected voxels outperform the intensity values of selected voxels. Finally, pruning the mesh arc weights leads to a slight increase in the classification performance.