T-distribution stochastic neighbor embedding for fine brain functional parcellation on rs-fMRI

In human functional connectome and network analyses, brain subregions that are functionally parcellated have a better consistency than do anatomical subregions. Resting-state functional magnetic resonance imaging (rs-fMRI) signals can function as coherent connectivity patterns that can be used to assess human brain functional architecture. In this paper, an optimized framework that combines automatic spectral clustering with dimensionality reduction is presented for fine-grained functional parcellation of rs-fMRI of the human brain. First, the t-distribution stochastic neighborhood embedding (t-SNE) algorithm extracts features from the high-dimensional functional connectivity patterns between voxels of the brain regions to be segmented and the whole brain. Then, the number of clusters is determined, and each voxel in the regions is parcellated by the automatic spectral clustering algorithm based on the eigengap. A quantitative validation of the proposed methodology in synthetic seed regions demonstrated its accuracy and performance superiority compared to previous methods. Moreover, we were able to successfully divide the parahippocampal gyrus into three subregions in both the left and right hemispheres. The distinctive functional connectivity patterns of these subregions, educed from rs-fMRI data, further established the validity of the parcellation results. Notably, our findings reveal a novel insight into brain functional parcellation as well as the construction of functional atlases for future Cognitive Connectome analyses.