In classical textbooks (Duvernoy et al., 1991
; Ono et al., 1990
) and in many atlases used in the neuroimaging community (Desikan et al., 2006
; Lancaster et al., 1997
; Talairach and Tournoux, 1988
), gyri are defined as the cortex joining the bottom of two neighboring sulci, sulci being only consider as virtual landmarks between them. Nevertheless, the brain is more “sulcal” than “gyral”, since one half to two thirds of the cortical surface is hidden in the sulci and in the lateral fossa of the brain (Van Essen, 2005
; Zilles et al., 1997
). This deep anatomy only recently became clear thanks to medical imaging and computer engineering that allowed the development of inflated and flattened maps of the cortical surface of the human brain (Dale et al., 1999
; Dale and Sereno, 1993
; Fischl et al., 1999a
; Fischl et al., 1999b
; Van Essen, 2005
). Surface based cortical labeling methods have major advantages as compared to volume based methods; first, the complex folded anatomy of the human cerebral cortex, which makes the identification of sulco-gyral structures difficult, even by trained anatomists, is visually simplified by the inflation process. For instance, the anatomy of the occipital pole is usually not clearly described in classical textbooks (Carpenter, 1991
; Federative Committee on Anatomical Terminology, 1998
; Ono et al., 1990
), whereas cortical inflation clearly reveals a robust organization in 3 parallel gyri, similar to the one previously described by Duvernoy (Duvernoy et al., 1991
). The description of the lateral sulcus is also aided by the use of inflated maps as the entire -usually hidden- cortex of the insula and the opercula is exposed and parcellated on a single view. Second, interindividual differences in cortical anatomy are better taken in account in surface versus volume approaches. For instance, Talairach (Talairach and Szikla, 1967
) studied the location of the central sulcus in 20 hemispheres after they were registered in the AC-PC coordinate system: a variation of several centimeters in the antero-posterior location of the central sulcus was observed, though this sulcus is regarded as one of the most constant. Not surprisingly, since they use maps of cortical geometry to drive cross-subject registration, group average using surface based approaches give a markedly better alignment of sulco-gyral structures than volume-based methods (Fischl et al., 1999b
; Van Essen, 2005
). Third, compared to a classical orthogonal volume coordinate system, a surface coordinate system respects cortical topology: points with close surface coordinates are always close on the cortical surface, whereas points with similar Talairach coordinates may be widely separated on the cortical surface (Fischl et al., 1999b
; Van Essen et al., 1998
Despite these substantial advantages for analyzing, averaging and displaying data, cortical inflation software creates new representations of cortical anatomy that the users need to relearn. For this purpose, tools for automatically labeling the cortical surface should provide great help; for instance, the FreeSurfer package http://surfer.nmr.mgh.harvard.edu/
is a set of tools for fully automated volume and surface reconstruction and labeling. This paper presents the anatomical rules and nomenclature used to build the sulco-gyral atlas included in this package. It contains minor changes as compared to versions included in prior distributions. To date, the FreeSurfer package, including the current or a previous version of the atlas was individually licensed 6700 times.
Classical anatomical textbooks (Duvernoy et al., 1991
; Talairach and Szikla, 1967
; Talairach and Tournoux, 1988
) and most of the available atlases (Desikan et al., 2006
; Lancaster et al., 2000
; Shattuck et al., 2008
) use a gyral based parcellation of the brain. Since a large proportion of the brain is hidden in sulci and lateral fossa (Van Essen, 2005
; Zilles et al., 1997
), other authors proposed a sulcal-based parcellation (Rettmann et al., 2005
; Tosun et al., 2004
). In this paper we proposed a mixed, sulco-gyral-based parcellation: the gyral cortex was defined as the one seen on a 3D reconstruction before inflation (pial view), the remaining hidden part being conversely labeled sulcal. A similar sulco-gyral classification was proposed as a starting point for a sulcal segmentation using a deformable surface model (Rettmann et al., 2002
). In this approach, a deformable surface, similar to a flexible balloon, surrounded one hemisphere. This balloon was progressively deflated and the deflation was stopped as its surface matched the cortical surface. Due to this process the content of the lateral fossa of the brain was classified as sulcal although it also contained both gyral and sulcal elements.
In order to obtain a better sulco-gyral classification, including for the cerebral lateral fossa, we used two parameters computed from the cortical surface: the mean curvature and the average convexity. The latter gave a good sulco-gyral classification except for the lateral fossa whose deep location would have resulted in the labeling of all structures forming and bordering the insula as sulcal, including the insular gyri and deeper part of the opercula. Thus, the structures of the lateral fossa were classified based upon their mean curvature value, some being considered as sulcal (circular sulcus of the insula…), and others being labeled gyral (insular gyri, superior aspect of the superior temporal gyrus…). This process resulted in about 55.5% of hidden cortex for both hemispheres, which is slightly lower than previously published results estimating that about 60% of the cortex is buried (Van Essen, 2005
; Zilles et al., 1997
). This value is highly dependent upon the precise value of the average convexity used as a threshold to perform the sulco-gyral classification. Although anatomically relevant, since it gave a reasonable sulco-gyral classification on the pial views (), the threshold we used probably explains this “over representation” of gyral cortex in our parcellation.
Out method subparcellates the sulcal and gyral parts of the cortex into smaller entities based on classical anatomical descriptions. We mainly used Duvernoy s nomenclature (Duvernoy et al., 1991
) since it gives a simple but precise description of the entire cortical surface based on 18 brains, and because it is widely and internationally used. This nomenclature includes and completes terms provided by the TA (Federative Committee on Anatomical Terminology, 1998
). When necessary, the correspondence with other terminologies (Ono et al., 1990
) was indicated to help the reader who is familiar with them. The final result was a parcellation of the entire cortex 74 different structures. This high number of anatomical regions allows a more precise description of the cortical surface, with acceptable automated/manual concordance. By comparison, the Talairach Daemon (Lancaster et al., 1997
) and the surface-based parcellation proposed by Desikan (Desikan et al., 2006
) defined 48 and 34 gyral regions per hemisphere respectively.
While the subparcellation we performed was only driven by anatomical conventions, other methods have been proposed. For instance in a watershed based approach (Rettmann et al., 2002
; Rettmann et al., 2005
) the sulcal cortex is subparcellated in catchment basins defined by the geodesic distance from the bottom of the sulcus to the cortical surface. This approach allows a semi-automated segmentation of branching sulci, for instance superior frontal and precentral sulci. Nevertheless, this approach sometimes misses limits between sulci and a manual intervention is needed (Rettmann et al., 2005
In another approach (Cachia et al., 2003
), bottom lines at the sulcal fundi are first delineated using a contextual pattern recognition method. A pair of these bottom lines were then used as a starting point for a Voronoï diagram to localize the crown of the gyrus limited by these two lines. Finally, once the crowns of the gyri are located, another set of Voronoï diagrams is built to delineate the corresponding gyri. While this method is presented as automated, the definition of the pairs of sulci initially used as starting points remains manual.
Since we aimed to provide a familiar and standard parcellation scheme to the users, the limits between cortical structures we proposed were only based on anatomical conventions similar to the ones used in classical textbooks (Duvernoy et al., 1991
). The labeling of the 12 subjects used to build the atlas was completely manual, with the attendant possible variations of the boundaries between same structures across subjects. Nevertheless, the parcellation was performed by the same author, and cortical structures were labeled in the same order for each subject. Moreover, this manual labeling was not directly used to label “new” hemispheres, but was only one of the parameters used to train the parcellation program that also integrated other parameters, especially the geometry of the cortical surface, and the relative location of the cortical structures (Fischl et al., 2002
; Fischl et al., 2004
One major limitation to this surface atlas is that it only labels the cortex, ignoring subcortical structures and hippocampus. Nevertheless, in the FreeSurfer reconstruction stream, deep structures are labeled by a volume-based tool using a similar probabilistic algorithm (Fischl et al., 2002
), resulting in the labeling of cortical as well as subcortical and ventricular structures at the end of the process.
The automatic labeling algorithm used for this paper was previously published using the same set of brains (Fischl et al., 2004
) with similar results for the average CIO
: 80% for the left hemisphere and 79% for the right hemisphere. Because this paper is derived from the same dataset, and a very slightly modified parcellation scheme, we did not seek to extensively validate the technique again. Since the precision of the manual definition of boundaries on the cortical surface is obviously limited by the width of the lines drawn to limit contiguous labels, the CI were also computed without considering their boundaries (CIC
). This better evaluates the auto/manual concordance since a discordance located just at the border between two areas has negligible anatomical or functional significance. After correction for this “border effect”, CIC
(percentage of area identically labeled by the manual and automated procedure) was close to 85% for both hemispheres, with noteworthy differences between structures ( and ). Lower CIC
values were found for variable/inconstant structures (anterior occipital sulcus, sulcus intermedius primus, suborbital sulcus...), for structures without clear landmarks (subcallosal area), and for sulci not deep enough to be correctly and constantly recognized (subcentral and paracentral sulci, central sulcus of the insula…). The structure area s was another important factor determining its concordance (); the area of 27 out of 33 cortical structures with a CIC
lower than 0.75, was lower than 11 cm2
. Some structures with reproducibly low CIC
values across subjects were grouped to increase robustness of the labeling procedure. Conversely, other structures that are known to be less variable across individuals had noticeable high CIC
(e.g. central, pre and post central gyri, calcarine sulcus, orbital H-shaped sulcus, superior temporal sulcus).
To our knowledge, this paper is the first extensive description of the anatomical conventions used to build a probabilistic sulco-gyral atlas of the human cerebral cortex. Even for well trained neuroanatomists, manual labeling of the cortex reconstructed from one given MR scan remains challenging: it is time consuming, it implies a high degree of anatomical expertise and remains sensitive to labeling variations. Using a fully automated approach to label the same scan is more reproducible and practical for large datasets. Manual localization of sulco-gyral structure remains difficult even on inflated maps and is often a compromise between several possible labelings; the automated procedure usually selects one alternative labeling scheme that on visual examination by an expert anatomist is found to be acceptable. For this reason, and despite a lack of perfect concordance between automated versus manual labeling, the proposed atlas produces an acceptable, reproducible and rapid labeling of the entire cortical surface. It produces a detailed parcellation of the cortex into 74 different structures per hemisphere that may be used for morphological and functional analysis. This paper will also serve as a reference for users of this automated tool since it provides a precise description of each of the parcellations that are output from the FreeSurfer surface reconstruction stream.