Brain Lobes Revealed by Spectral Clustering

The cortical surface is composed of a mosaic of distinct regions that are functionally or micro-anatomically homogeneous. Parcellating the cortical surface into few lobes is a common approach in neuroimaging resulting from a long tradition in physiology. However, defining such sub regions consistent across subjects is more difficult. If macro-anatomical landmarks such as central sulcus and parieto-occipital sulcus are clear boundaries between distinct lobes, appropriate separators still have to be defined e.g. in the posterior temporal lobe. Several approaches have been proposed, but they are all built from supervised information in general from manually defined segmentation onto an atlas brain. However, some of the boundaries imposed in such atlas actually rely on ad hoc approaches rather than anatomical or functional considerations regarding the underlying structure of the cortex. In this work, we propose an original technique that allows to define a parcellation of the cortical surface based on its intrinsic properties with no a priori information. Our approach is based on spectral clustering applied to the first Eigen functions of Laplace-Beltrami Operator of the cortical mesh. We demonstrate a good reproducibility of clusters across subjects as well as striking visual similarities between our segmentation and traditional lobar parcellation.