Automated Classification of Discrete Human Thoughts Using Functional Magnetic Resonance Imaging (fMRI): Comparison between Voxel-Based and Atlas-Based Feature Selection Methods

Abstract

It has been reported that human thoughts processes of sensory-motor functions as well as high level of cognitive processes may be highly reproducible between multiple trials as measured via functional MRI data. This trend of the reproducibility seems consistent between multiple subjects as well. We have also presented in our earlier study that six distinct thought processes were shown highly consistent spatial patterns of activations as evaluated from automated classification performance. In the present study, this automated classification performance was compared depending on the feature vector selection methods. A general linear model (GLM) was adopted to define a neuronal activity and voxel-based or atlas-based approaches were adopted as feature vector selection methods. The classification results showed superior performance from the voxel-based feature selection method than the atlas-based method. Nonetheless, when multiple atlases were used to defined feature vector elements, the resulting performance was comparable to that of the voxel-based method with greatly reduced computational time.