ATLAS-BASED CLUSTERING OF SULCAL PATTERNS – APPLICATION TO THE LEFT INFERIOR FRONTAL SULCUS Olivier Coulon 1 , Vladimir Fonov 2 , Jean-Franc ¸ois Mangin 3 , D. Louis Collins 2 1 Aix-Marseille Univ, LSIS, UMR CNRS 6168, Marseille, France 2 McConnell Brain Imaging Center, Montreal Neurological Institute, McGill University, Montreal, Canada 3 CEA, I2BM, DSV, Neurospin, LNAO, Gif-sur-Yvette, France ABSTRACT We present an attempt at characterizing local patterns of cortical folding, an open problem in neuroscience. The tech- nique is applied to an extremely variable sulcus: the left infe- rior frontal sulcus (LIFS). Our approach is based on the use of an average template as a reference to define features that char- acterize the position, presence, and orientation of elementary sulcal elements of the LIFS. Clustering was performed on the resulting feature vectors after the optimal number of clusters was found using the Gap statistic. The technique was ap- plied to data from 151 subjects. Here, we present our results and discuss the nature of sulcal organization that exists within the apparently unstructured sulcal variability. Index Terms— Cortical folds, Clustering, Atlas 1. INTRODUCTION Cortical organization is an open problem in neuroscience. Lit- tle is known about the organization of the cortical folding pat- tern and its link with development and function. Nevertheless, it has been hypothesized that this pattern has a specific geo- metric structure [1][2], and models of organization have been presented and tested [3]. In particular, the hypothesis of an orthogonal organization of the cortical folds around specific sulcal entities has emerged. The existence of such an orga- nizational scheme is of great interest because of its potential relationship to brain function and because it is believed that many brain-related pathologies such as autism and epilepsy may be associated with disruptions of this scheme. Gaining an understanding of cortical variability and or- ganization first requires identification of the possible patterns for a specific fold or set of folds. The idea that a popula- tion can be divided into subgroups, with each member fitting a possible local pattern, has been used in the automatic la- beling of cortical structures in [4]. Essentially, each local cortical component can be identified in a catalogue, and the global cortical organization is a collection of these local con- figurations. To explicitly describe local libraries of cortical patterns, large groups of subjects must be clustered in con- sistent subgroups, similar to what was done manually in [7]. Such a clustering approach has been performed in [5][6] us- ing global shape descriptors or direct intersubject distances using the Iterative Closest Point (ICP) algorithm. Although the latter produces clear clusters for a few sulci, the global intersubject distance forces the use of either sulci with simple topological variations (e.g., the superior temporal sulcus) or larger sulcal sets with large-scale intergroup differences that can be captured by the global ICP-based distance (e.g., inter- mediate+marginal+orbital+inferior frontal region). In the following, we present an attempt to characterize the folding patterns and fine local variability of a specific sulcus with extremely variable topography, the left inferior frontal sulcus (LIFS). This sulcus is of particular interest as one of the key boundaries of Broca’s area, and existing cortical folding patterns could be related to functional specificities of the lan- guage system. Pattern characterization is achieved by cluster- ing typical configurations after comparison with an unbiased, statistically centered atlas template built from a large set of subjects. Inspired by the notion of orthogonal directions pre- sented in [1], namely, that variability occurs around two main orthogonal directions, our approach is based on the defini- tion of features that measure the distances and orientations of individual LIFSs with reference to the template. Indeed, two main orthogonal directions are observable on the tem- plate LIFS. In parallel to the clustering process, we also in- vestigated whether there was discernible structure within the variability, and in particular, whether clusters were present. 2. METHODS 2.1. Template extraction We used the publicly available asymmetric version of the ICBM 2009c nonlinear average template 1 [8], which was created from 152 subjects in the International Consortium for Brain Mapping (ICBM) database. The template was pro- cessed with the BrainVisa T1 segmentation pipeline 2 . The white matter surface was extracted, and a complete represen- tation of its folding pattern was constructed into a relational graph (Fig. 1). That it was possible to process this template through a pipeline dedicated to T1-weighted images is partic- ularly striking: despite being the average of a large number of subjects, this template has a very well-defined morphology, 1 http://www.bic.mni.mcgill.ca/ServicesAtlases 2 http://brainvisa.info