SPANOL (SPectral ANalysis of Lobes): A Spectral Clustering Framework for Individual and Group Parcellation of Cortical Surfaces in Lobes

Understanding the link between structure, function and development in the brain is a key topic in neuroimaging that benefits from the tremendous progress of multi-modal MRI and its computational analysis. It implies, , to be able to parcellate the brain volume or cortical surface into biologically relevant regions. These parcellations may be inferred from existing atlases (e.g., Desikan) or sets of rules, as would do a neuroanatomist for lobes, but also directly driven from the data (e.g., functional or structural connectivity) with minimum a priori. In the present work, we aimed at using the intrinsic geometric information contained in the eigenfunctions of Laplace-Beltrami Operator to obtain parcellations of the cortical surface based only on its description by triangular meshes. We proposed a framework adapted from spectral clustering, which is general in scope and suitable for the co-parcellation of a group of subjects. We applied it to a dataset of 62 adults, optimized it and revealed a striking agreement between parcels produced by this unsupervised clustering and Freesurfer lobes (Desikan atlas), which cannot be explained by chance. Constituting the first reported attempt of spectral-based fully unsupervised segmentation of neuroanatomical regions such as lobes, spectral analysis of lobes (Spanol) could conveniently be fitted into a multimodal pipeline to ease, optimize or speed-up lobar or sub-lobar segmentation. In addition, we showed promising results of Spanol on smoother brains and notably on a dataset of 15 fetuses, with an interest for both the understanding of cortical ontogeny and the applicative field of perinatal computational neuroanatomy.