Structural MRI Based Parcellation of The Human Brain Using Spatial Hierarchical Clustering

Abstract

Parcellating the human brain into neurobiologically meaningful regions serves two (key) purposes , i.e. to gain neurobiological insight on the modules of the brain and to achieve dimensionality reduction for large -scale studies . Data- driven parcellation approaches based on magnetic resonance imaging ( MRI ) mainly differ from each other by the modalities that they are derived from and the parcellation algorithm. Here we extensively evaluate parcellation approaches which are based on one specific modality , i.e. grey matter volume deduced from T1- weighted structural MRI scans . The newly proposed spatial hierarchical variable clustering method SPARTACUS is employed as parcellation algorithm and its performance is compared with other spatial hierarchical agglomerative clustering ( SHAC ) algorithms as well as with spatial spectral clustering on both simulated data and a single- site structural MRI data set of older adults using two metrics , i.e. clustering quality and clustering stability across subsamples . Our analysis reveals that SPARTACUS and Ward's SHAC outperform the other competing methods . Using spatial ensemble clustering methods we could further improve the quality of the parcellations . Performing parameter tuning of the granularity ( considering 2-1000 brain regions ) we identify multiple interesting numbers of brain regions , where the corresponding parcellations may reflect different levels of brain organization . Finally , we observe that our parcellations converge well above chance with other structural MRI based parcellations but hardly better or even worse than chance with parcellations from another modality .