Microstructural grey matter parcellation and its relevance for connectome analyses

Abstract

The human brain connectome is closely linked to the anatomical framework provided by the structural segregation of the cortex into distinct cortical areas. Therefore, a thorough anatomical reference for the analysis and interpretation of connectome data is indispensable to understand the structure and function of different regions of the cortex, the white matter fibre architecture connecting them, and the interplay between these different entities. This article focuses on parcellation schemes of the cerebral grey matter and their relevance for connectome analyses. In particular, benefits and limitations of using different available atlases and parcellation schemes are reviewed. It is furthermore discussed how atlas information is currently used in connectivity analyses with major focus on seed-based and seed-target analyses, connectivity-based parcellation results, and the robust anatomical interpretation of connectivity data. Particularly this last aspect opens the possibility of integrating connectivity information into given anatomical frameworks, paving the way to multi-modal atlases of the human brain for a thorough understanding of structure-function relationships.

Basic organizational principles in the brain — segregation and integration

Two fundamental principles govern the organization of the human brain: segregation and integration (Amunts et al., 2007; Eickhoff and Grefkes, 2011; Friston, 2002; Tononi et al., 1994, 1998; Zilles and Amunts, 2010). Functional segregation encompasses the subdivision of the cortex into distinct regions which particularly pertains to the cerebrum with its cortical areas and subcortical nuclei (Zilles and Amunts, 2012). That is, the cortex is not uniform, but modules can be delineated based on different structural and functional criteria, including intrinsic structural properties of the cortex (cellular and neurochemical architectonics, intracortical connections) as well as its functional response properties. Functional integration in turn emphasizes the idea that the execution of whatever kind of sensorimotor or cognitive task relies on the interplay and dynamic exchange of information between different brain regions. The presence of many distinguishable networks, each serving a given function or mental process such as specific aspects of language processing or spatial attention, is the predominant manifestation of such functional integration. It is, however, important to emphasize that these two concepts are not contradictory to each other, but rather should be integrated into a unifying view. This has recently also been highlighted by Sporns (2013) from the perspective of connectomics and large-scale brain connectivity. The interconnectedness of network communities as closely related sets of brain regions reflect the functional segregation of the brain. Their communication via hub regions integrates the information from different network communities. Each brain region is structurally and functionally optimized for a particular computation. The interaction of brain regions, each of which serving a distinct process within a given network, is necessary to fulfill a complex task (Friston, 1994, 2002).

It is particularly challenging to analyze how a structurally defined entity functionally specializes or, more general, what constitutes a particular functional entity. Importantly, not only the intrinsic properties (architecture, internal wiring) of such a module determine its functional role, but also its connections to other regions which enable the interplay in common structural and functional networks. Three major concepts of connectivity need to be considered since all of them focus on a different aspect of a region's input and output. However, all of them require the structural definition of a particular brain region as the starting point to which connectivity is ascribed to: (i) Anatomical connectivity refers to the presence of white matter fibres, which provide the physical connection between brain regions. It can be assessed in the living human brain using diffusion-weighted magnetic resonance imaging and tractography algorithms (Johansen-Berg and Rushworth, 2009; Le Bihan, 2003) and in post-mortem brains using dissections, various nerve fiber staining techniques or polarized light imaging (Axer et al., 2011; Bürgel et al., 1997, 1999, 2006; Flechsig, 1920; Rademacher et al., 2001, 2002). Recent approaches used structural correlation analyses to detect parts of the brain, of which anatomical features co-vary with each other, such as cortical thickness or volume (Chen et al., 2011; Lerch and Evans, 2005; Lerch et al., 2006; Wu et al., 2012). (ii) Functional connectivity is conceived as the temporal coincidence of spatially distant neurophysiological events (Friston, 1994), thus describing correlations (Buckner, 2010). Such events could be time-courses of resting-state or functional MRI experiments (Fox and Raichle, 2007), but also co-activation patterns in large-scale meta-analyses (Robinson et al., 2010) or magneto- and electrophysiological recordings. (iii) Effective connectivity introduces causality considerations since it assesses the influence which is exerted by one brain region over another or the contextual modulation thereof, as modeled by different techniques (Friston et al., 2003; Roebroeck et al., 2005). Thus, all three concepts of connectivity may pertain to different scales of analysis, from connectivity between single neurons to large brain regions. They describe different aspects of how a particular brain region is integrated in networks and complement the description of how a particular region is intrinsically structured in terms of its cellular and neurochemical organization.

To evaluate the relation between structure and function and to approach an integrated understanding of how the system ‘brain’ works, it is necessary to take these different pieces of information into account and integrate them in a common concept of brain organization. In the context of mapping the connectome, considering information about the brain regions themselves, i.e. grey matter parcellations, in addition to the information about their connectivity is thus of importance to maintain the reference frame given by the intrinsic anatomy of the brain's grey matter. This point is emphasized because assessing connectivity and network properties is also possible on a more abstract level, largely independent of the underlying anatomy, as has repeatedly been proposed in some fields of connectomics (Bullmore and Sporns, 2009; Rubinov and Sporns, 2010). The latter line of research encompasses computational methods such as graph theory or parcel-based approaches. Here, large-scale network measures are generated to describe the overall functioning of the entire network architecture. They abstract, however, from given spatial constraints and are thus largely independent of the underlying brain anatomy. Thus, these approaches mainly focus on quantitative measures of networks and their interconnectedness, e.g., in terms of strength or direction of a connection. However, different parts of the brain have different functions. Thus, the here discussed concepts of segregation and their related measures could complement data on overall network functioning.

Therefore, a concept of segregation and localization is indispensable to understand the system ‘brain’ not only conceptually, but as the organ with a given shape and texture of its constituting components. This article depicts how analyses on the structural organization of the cortex provide an anatomical reference frame for the analysis and interpretation of different types of connectivity data.

Macro- vs. microanatomy

The grey matter of the brain can be parcellated into distinct entities by two different approaches on different scales, as cortical areas or subcortical nuclei may be delineated based on either macroanatomical or microanatomical criteria. Referring to macroanatomy involves the description and delineation of gyri and sulci in the cortex. In this approach, different cortical regions are usually associated with gyri, using the surrounding sulci as their borders. Depending on the region, additional subdivisions are introduced by e.g. delineating rostral/caudal or ventral/dorsal parts of a gyrus separately. Contrarily, microanatomical parcellations use microscopical criteria for cortical mapping. The cytoarchitecture of the cortex is based on differences in cell size, packing density and distribution throughout the cortical layers (Amunts et al., 2007; Brodmann, 1909; von Economo and Koskinas, 1925; Zilles and Amunts, 2010,2012). Similarly, characteristics of the myeloarchitecture of the cortex, i.e. differences in myelination between the cortical layers, can be used to delineate cortical areas (Amunts et al., 2000; Flechsig, 1920; Nieuwenhuys, in press; Vogt and Vogt, 1919; Zilles and Palomero-Gallagher, 2001). Recent approaches demonstrated intracortical connectional structure in-vivo (Kleinnijenhuis et al., in press; Leuze et al., in press). Even considering the current limitations in resolution as compared to post-mortem studies, these new approaches provide the potential to delineate cortical areas in-vivo based on microstructural features, particularly when used at ultra-high field MRI (Koopmans et al., 2011). Moreover, microanatomical features also encompass differences in the distribution of neurotransmitter receptors (Zilles and Amunts, 2009) or gene expression data (Lein et al., 2007; Lipska et al., 2006). All these types of criteria are being used, separately or in combination, to obtain a parcellation of the cortex into distinct areas.

Applying either approach to the cerebral cortex led to the emergence of brain atlases which provide a parcellation of the cortex into distinct areas. Atlases based on microstructural criteria were already available in the early 20th century, the most well-known being the cytoarchitectonic map of Brodmann (1909). But these maps were only available as two-dimensional schematic drawings on a typical human hemisphere. Since about two third of the cortex are buried within the sulci (Zilles et al., 1988), such depiction misses a considerable amount of the cortex. Furthermore, no true correlation with functional imaging or connectome data is possible on the basis of two-dimensional schematic drawings. These shortcomings led to the development of modern computer-based brain atlases. They provide three-dimensional anatomical data within a common reference space (Roland and Zilles, 1994), one of the most widely used being the MNI space (Montreal Neurological Institute; Evans et al., 1992, 2012). Most of the data from functional magnetic resonance imaging (fMRI), positron emission tomography (PET) and magneto- and electroencephalographic (MEG/EEG) studies are registered to this reference space and thus have a common stereotaxic system. This allows describing locations in the brain by three-dimensional coordinates and assigning them to the underlying anatomical structure. The atlases therefore provide an anatomical reference for functional or connectome mapping.

Alternative brain atlases not only differ with regard to the criteria used for delineation of the cortical areas (macro-/microanatomy), but also regarding the underlying method used: The atlases could be either deterministic or probabilistic. Deterministic atlases are e.g. the Talairach atlas which manually transferred the two-dimensional map of Brodmann (1909) into three-dimensional space (Talairach and Tournoux, 1988), or the AAL atlas (Automated Anatomical Labeling), which describes a macroanatomical parcellation of the MNI single subject template (Tzourio-Mazoyer et al., 2002). The latter provides exact and fixed borders between adjacent regions and may thus be conceptualized as deterministic. While providing the approximate localization of each area, such an approach does not account for the interindividual variability between different brains. This might lead to false identification of a specific location in individual brains, particularly at the transition between regions. Probabilistic atlases, on the other hand, are based on a number of different brains where each brain is parcellated individually and the atlas is derived by averaging across all obtained individual parcellations. Thus, for each voxel in the reference space, the most likely anatomical structure is identified and probabilities are provided. They show how likely a specific anatomical structure is present within a voxel of the reference brain, considering the amount of interindividual variability present in the data. Probabilistic atlases are e.g. the Harvard-Oxford structural cortical and subcortical atlas (http://www.cma.mgh.harvard.edu/fsl_atlas.html; Desikan et al., 2006), which provides regions corresponding to gyri based on macroscopic criteria, and the JuBrain Cytoarchitectonic Atlas (Amunts et al., 2007; Zilles and Amunts, 2010), which is based on cytoarchitectonic delineations of brain areas (http://www.fz-juelich.de/inm/inm-1/EN/Forschung/JuBrain/_node.html). The multi-modal combination of probabilistic atlases with functional and connectivity data currently represents an important tool for the analysis of structure-function relationships in the brain (Toga et al., 2006).

The main difference between the macroanatomical and microanatomical parcellations relates to the level of detail and precision by which different regions and their borders may be differentiated. The delineation based on macroanatomical criteria may become arbitrary, if cortical landmarks are lacking, e.g. if a gyrus is further subdivided or sulci are extremely variable as visible in prefrontal and higher visual regions. If such macroanatomical subdivisions are expanded up to small equally sized blocks of cortex, these units are lacking any relevant anatomical information. Therefore, most macroanatomical parcellations generate subdivisions of the cerebral cortex mainly oriented at gyri and larger sulci.

Microanatomical criteria allow defining a border between two cortical areas precisely where the pattern of cyto- or myeloarchitecture, occurence of distinct cell types (e.g., Betz-cells in the primary motor cortex), neurotransmitter receptor or gene expression changes. These microstructural characteristics are the relevant structural basis for the function of an area, because the number and type of cells as well as their interconnections, their endowment with neurotransmitter receptors, the electrophysiological properties of cells and the interplay between different cells largely determines, together with the respective input and output, how this cortical area reacts to given stimuli and produces output.

Microscopically defined borders of cortical areas often vary independent of macroanatomical landmarks, although correspondences between microscopically identified areas and macroanatomical reference points are observed: The border between primary motor area 4 and primary somatosensory area 3a is always located in the depth of the anterior wall of the central sulcus (Brodmann, 1909; Geyer et al., 1999). The primary visual cortex area 17 is always found within the calcarine sulcus (Amunts et al., 2000; Fischl et al., 2008; Hinds et al., 2008). The position of Broca's region, consisting of cytoarchitectonic areas 44 and 45 (Amunts et al., 1999), is always on the caudal aspect of the inferior frontal gyrus, and the entorhinal cortex can be defined by superficially protruding cellular islands (Fischl et al., 2009). The exact borders of these cortical areas cannot precisely be delineated based on surrounding and consistently occurring sulci. Since the histological features of a cortical area are the crucial underpinning of its function, only the microstructurally defined border of an area is important. E.g., the borders of cytoarchitectonic area OBγ (von Economo and Koskinas, 1925), which represents the vertical meridian of the visual field, can solely be identified by microscopical inspection. Its very large pyramidal cells in deeper layer III are the structural basis for the interhemispheric connections of this region. Thus, only the microstructural delineation can provide more than topographical estimates of the localization of cortical areas (Amunts et al., 2007).

The difference between micro- and macroanatomically identified borders of cortical areas are indicators of a considerable interindividual variability between brains, particularly in higher order cortices which are located in regions of the brain with highly variable sulcal and gyral patterns. While primary and secondary cortices are located within early developing structures of the cortex such as the main sulci (central sulcus, calcarine sulcus), variability within these regions largely follows the variability of macroanatomical structures. Moving to higher order cortices increases variability in the extent of the microanatomically defined area, but also in the macroanatomical structures since these secondary, tertiary or higher grade smaller sulci develop later and are highly variable across individuals (Fischl et al. 2008; Rademacher et al., 1993). For mapping connections of functionally relevant patches of the cortex, it is important to use microscopically defined cortical probability maps, since those maps provide not only precise definitions of borders but also reflect the interindividual variability.

In the following section, the methods behind the JuBrain Cytoarchitectonic atlas are depicted in more detail, since this atlas combines both advantages: (i) being based on microscopic criteria, thus allowing precise delineations of cortical areas; and (ii) being probabilistic, thus adequately accounting for the considerable interindividual variability between different human brains.

Probabilistic cytoarchitectonic mapping — JuBrain

The probabilistic cytoarchitectonic mapping forming the heart of the JuBrain atlas project was developed in the C. and O. Vogt Institute for Brain Research of the Heinrich-Heine-University Dusseldorf (Germany) starting from the 1990s and further perpetuated since then in Dusseldorf and in the Institute of Neuroscience and Medicine of the Research Centre Juelich (Germany) (Zilles and Amunts, 2010). The main idea behind this approach was to overcome shortcomings, which were recognized since the emergence of the so called ‘classical’ cytoarchitectonic maps of the early 20th century, e.g. Brodmann’s map (1909), and to generate an atlas, which provides a thorough anatomical basis for the interpretation of functional imaging data and thus for the assessment of structure-function relationships.

Using the ‘classical’ maps, one has to face three major caveats (Zilles and Amunts, 2012): (i) they are only available as two-dimensional drawings, and therefore could not provide voxel-by-voxel information in 3-D space; (ii) they are usually based on delineations in one or only very few hemispheres while the results are schematically shown on a ‘typical’ human hemisphere; therefore, almost no information is available about the interindividual variability between different brains; and (iii) the delineations are based on subjective delineation criteria of the researcher, which might introduce an observer bias into the resulting parcellations.

The JuBrain atlas explicitly addresses these issues: (i) it is a 3-D atlas with cytoarchitectonic anatomical information at every voxel within the MNI reference space; (ii) areas are delineated in 10 post-mortem human brains and the respective variability is part of the atlas; and (iii) the delineations are based on an observer-independent, quantitative statistically testable mapping approach.

The post-mortem human brains (5 male, 5 female) are histologically processed, embedded in paraffin, sectioned into 20 μm thick sections (coronal, sagittal, or horizontal), and stained for cell bodies. This procedure reveals the cytoarchitecture of the cortex and subcortical nuclei. Different areas can be detected by differences in cell size and type, cell packing density, and thickness of cortical layers. To objectively capture these differences, an observer-independent mapping approach is used (Schleicher and Zilles, 1990; Schleicher et al., 1999, 2000, 2005, 2009; Zilles et al., 2002). Here, regions of interest within the cortex are digitized using a light microscope. The digitized images are transformed into GLI (grey level index) images where each pixel represents a measure for the volume fraction of cell bodies. Differences in GLI values between adjacent cortical regions represent relevant differences in cytoarchitecture. Therefore, GLI values along traverses running perpendicular to the cortex are extracted to generate GLI profiles across all cortical layers. By comparing the Mahalanobis distance between adjacent sets of profiles, dissimilarities in GLI distribution and thus in cytoarchitecture are detected. The Mahalanobis distance is a similarity measure (Mahalanobis et al. 1949) which allows comparing features, e.g. of cortical cytoarchitectonic profiles. Where the Mahalanobis distance function reaches a local maximum, the dissimilarity between adjacent parts of the cortex is highest, indicating a cytoarchitectonic border.

Using this procedure, a cortical area is delineated from surrounding parts of the cortex in each of the 10 post-mortem brains. The delineations on every histological section together with the whole histological dataset are 3-D reconstructed for each individual post-mortem brain. The reconstructed 3-D data of all 10 brains are spatially normalized and transferred to the MNI single-subject template within the MNI reference space to generate stereotaxic maps. Superimposing the normalized maps of each individual brain enables the calculation of probabilistic cytoarchitectonic maps of each delineated area in stereotaxic space. These probabilistic maps summarize and display the interindividual variability regarding the exact location and extent of an area in different brains. For each voxel of the reference space, the probabilistic map denotes the probability of finding this cytoarchitectonic area at that particular voxel. Thus, low probabilities (e.g. 10%-30%) indicate that this area is not very likely to be found at that position in the brain, i.e. this area is present in this particular voxel only in few individuals. High probabilities (e.g. 70%-100%) show that this area is typically to be found, over most subjects, in this particular voxel.

Due to the probabilistic nature, maps of adjacent cortical areas overlap. To obtain a continuous, non-overlapping parcellation of the cortex, maximum probability maps (MPM) are generated (Eickhoff et al., 2006b). Within these maps, each voxel is assigned to the cytoarchitectonic area with the highest probability.

The probabilistic maps of cortical areas and subcortical nuclei form the JuBrain Cytoarchitectonic atlas (Zilles and Amunts, 2010). The atlas can be assessed in 3-D using the JuBrain WebTool (http://www.fz-juelich.de/JuBrain/EN/_node.html), which provides different visualizations of each cytoarchitectonic area together with information about the probabilities in each voxel (Fig. 1). Furthermore, the atlas is implemented in the SPM Anatomy Toolbox (Eickhoff et al., 2005, 2006b, 2007; http://www.fz-juelich.de/inm/inm-1/EN/Forschung/_docs/SPMAnatomyToolbox/SPMAnatomyToolbox_node.html), which allows directly overlaying results from functional imaging and connectivity studies with the probabilistic cytoarchitectonic maps to identify the anatomical location of a functional activation (Fig. 2). Similarly, the JuBrain atlas is implemented in the widely used software packages FSL and AFNI and can be used within these frameworks for the interpretation of structural, functional and connectivity data.

Fig. 1.

The JuBrain Cytoarchitectonic Atlas (Zilles and Amunts, 2010) as accessible via the JuBrain WebTool (http://www.fz-juelich.de/luBrain/EN/_node.html), visualized on the MNI single subject template. (A) Cytoarchitectonic areas mapped until now (coloured) as maximum probability maps (MPMs) from left (top left), right (bottom left), occipital (top middle) and medial views (bottom middle). Note the pink and purple coloured areas as visible from the occipital and medial view, depicting primary visual area hOc1 (pink, probability map shown in B) and secondary visual area hOc2 (purple) (Amunts et al., 2000). (B) Probability map of area 17/hOc1 (Amunts et al., 2000) from occipital (top right) and medial right hemisphere view (bottom right), projected onto the grey matter/white matter interface for better visualization of the sulcal patterns. Colours from blue to red decode lower and higher probabilities, respectively, of finding this area in that voxel.

Fig. 2.

The JuBrain Cytoarchitectonic Atlas (Zilles and Amunts, 2010) as implemented in the SPM Anatomy Toolbox (Eickhoff et al., 2005; http://www.fz-juelich.de/inm/inm-1/EN/Forschung/_docs/SPMAnatomyToolbox/SPMAnatomyToolbox_node.html), depicted on sections of the MNI single subject template. Grey colours depict different cytoarchitectonic areas. (A) Cytoarchitectonic probabilities at defined MNI coordinates. Cross hair positioned within parietal opercular area OP4 (Eickhoff et al., 2006a,c). (B) Visualization of cytoarchitectonic probability maps, here for inferior parietal area PF (Caspers et al., 2006, 2008). (C) Example of assigning functional activations (here meta-analysis results) to cytoarchitectonic areas with different probabilities (Schilbach et al., 2012): Convergence of default-mode and affective processing within left amygdala (laterobasal nuclei group, LB; Amunts et al., 2005).

How microstructure can inform and supplement connectome analyses

In the context of the connectome, it could be asked how such atlases can be used to inform and complement the information on functional integration. In our view, there are three major fields in which brain atlases, in particular probabilistic microstructurally derived ones, may be of high value to the endeavor of mapping the human connectome: i) Setting up a framework for a particular connectome analysis by defining regions of interest for which connectivity could be assessed, i.e., providing seeds and targets in region-based analyses. ii) Providing independent information on cortical segregation that can be used as a comparison for studies on connectivity-based parcellation of the grey matter in order to test for converging evidence or discrepancies. iii) Allowing anatomical allocation and hence interpretation of results from whole-brain connectivity analyses.

Identifying seed and target regions for connectome analyses

Different approaches exist to study connectivity in the human brain, e.g., seed-based analyses. Seed regions are used in studies assessing functional connectivity, using either resting-state correlations (RS-FC; Fox and Raichle, 2007), meta-analytic connectivity modeling (MACM; Robinson et al., 2010) data, and in studies on structural connectivity, using diffusion tractography (Behrens et al. 2003; Le Bihan, 2003). Targets of these connections could be every single voxel in the rest of the brain. They can also be specified using prior knowledge to test a given hypothesis. In brain wide analysis, this may be expanded to test for the pairwise connectivity between any two regions of the human brain, thus allowing describing different complex network measures (for an overview, see Rubinov and Sporns, 2010). Whether analyzing seed to (voxel-wise) whole-brain connectivity or seed-to-target connectivity of the entire connectome as the connectional matrix between any two regions, there are several different approaches how the regions to be analyzed may be defined. One of them is to use functional definitions, e.g., from fMRI studies or meta-analyses thereof. This might be particularly worthwhile for the analysis of functional connectivity, as has recently been stressed (Smith, 2012). A definition of seed and target regions based on anatomical criteria might equally be useful. It is largely independent of prior assumptions about the functional involvement of any part of the cortex into a given functional network. Furthermore, it allows complementing information from different levels, i.e. anatomical grey matter parcellation and structural or functional connectivity (Fig. 3). Results of these analyses could intuitively be summarized in connectional fingerprints which depict a quantitative measure for the connectivity of a seed region, e.g. connection strength, to a given number of target regions within a polar plot (Passingham et al., 2002).

Fig. 3.

Seed-based probabilistic tractography (performed with FSL) on diffusion-imaging data of 39 subjects (for details: Caspers et al., 2011), visualized on the MNI single subject template. (A) Two seeds within the inferior parietal lobule, areas PFt (green) and PF (red), which are located next to each other on the supramarginal gyrus. (B) Tracts running from the seeds towards the external capsule. (C) Course of the tracts through the external capsule. The tract of PFt (green) is displayed in transparent fashion to allow for visualization of the tract of PF (red). (D) Transcallosal connections of the seeds, clearly distinct from each other.

In a seed-based regression analysis on resting-state data, Diederen etal. (2012) and Sommer etal. (2012) used macroanatomically defined seed regions within the inferior frontal, superior temporal, and parahippocampal gyrus. They could demonstrate aberrant functional connectivity between these regions in a group of non-psychotic and psychotic patients with auditory verbal hallucinations as compared to controls. Similarly, Zald et al. (in press) demonstrated that the macroanatomically defined medial and lateral parts of the orbitofrontal cortex showed differential co-activation patterns using meta-analytic connectivity modeling, another application for atlas-based definition of seed regions. The brain is readily dividable into macroanatomical regions, therefore allowing for connectome-wide analyses. However, this is at the expense of anatomical preciseness. The use of microanatomically defined brain areas increases specificity of the anatomical location. Respective definition of seeds and targets thus enables detection of subtle changes in connectivity which reflects differential involvement of adjacent brain areas in different functional networks and allows disentangling the functional and connectional variety observed in different parts of the brain. Identifying brain areas based on microanatomical criteria, such as change in neuronal structure, most likely reflects changes in the composition of functionally relevant features of the cortex. E.g. the term ‘fronto-parietal network’ might pertain to a considerable number of functional networks (e.g. language, attention, working memory, action processing), each involving different parts of the frontal and parietal lobes and thus providing the basis for different functional capabilities. Using microscopically defined seed regions, Caspers et al. (2011) found differential structural connectivity of five areas of the inferior parietal lobule, namely areas PFt, PF, PFm, PGa, and PGp (Caspers et al., 2006, 2008). As target regions, all other areas of the cortex were defined using either microscopically derived areas as available from the JuBrain atlas, or macroscopically defined regions using the Harvard-Oxford structural cortical atlas. Uddin et al. (2010) could show that inferior parietal areas PGa and PGp and intraparietal areas hIP1, hIP2 (Choi et al., 2006), and hIP3 (Scheperjans et al., 2008a,b) have differential resting-state functional connectivity. Within this resulting network of brain areas, they additionally showed that the functional connectivity was largely complemented by respective structural connections. Eickhoff et al. (2010) assessed the structural and functional connectivity of two cytoarchitectonically defined areas within the parietal operculum, OP1 and OP4 (Eickhoff et al., 2006a,c), using probabilistic fibre tracking and meta-analytic connectivity modeling. As target regions, combined areas of the JuBrain atlas were used, for which differential connectivity with the two seed regions was revealed.

These examples demonstrate that seed and target regions could usefully be defined using areas of anatomical atlases. The advantage is that the definition of seeds and targets is largely independent of functional a priori hypotheses, but anatomically informed. As compared to parcels, such atlas-based parcellations are not arbitrary, but based on relevant microanatomical criteria of brain organization, i.e. specific intrinsic properties of the cortex.

Validating data from connectivity-based parcellation

Besides mapping the connectivity between a set of regions of interest, another important aspect of connectome analysis is a data driven approach which uses connectivity information to parcellate a given cortical region: connectivity-based parcellation (CBP). Here, a connectivity profile summarizing the connectivity to every other voxel in the brain is obtained for every voxel within the region under investigation. The individual connectivity profiles between each seed voxel and every other voxel in the brain are summarized in a connectivity matrix. The seed voxels are then clustered into distinct groups based on the similarities in those whole-brain connectivity profiles between them. The idea is to derive sub-regions within the seed that have a homogeneous connectivity within each sub-region, but are maximally different in connectivity between them. This allows the mapping of distinct modules within the seed region, which differ with regard to their whole-brain connectivity. In this context, it should be mentioned that the underlying connectivity profiles may be obtained from different types of connectivity information, e.g. resting-state functional connectivity (Fox and Raichle, 2007), co-activation patterns (Cieslik et al., in press; Eickhoff et al., 2011) or diffusion-based structural connectivity using tractography algorithms (Klein et al., 2009).

Evidently, such parcellations of the grey matter are based on their connectivity profiles. Conceptually, however, a cortical area as a functional unit of the brain should feature not only a distinct set of connections but also a distinct set of intrinsic properties, i.e., architecture, as discussed in Section 1. Comparing connectivity-based parcellation with microstructural atlases thus allows investigations of intrinsic (architecture) and extrinsic (connectivity) properties of the cortex. Integrating these different pieces of information (atlas and CBP) allows cross-validation by two independent aspects of cortical organization, and additionally investigations into potential divergences pointing to the need for further research. Either way, both approaches complement each other, and in combination, generate novel information on brain organization (Cloutman and Lambon, 2012).

Within different regions of the brain, this combined approach has already been successfully applied. Anwander et al. (2007) performed CBP based on diffusion tensor imaging (DTI) data on Broca’s region in the caudal part of the inferior frontal gyrus. Cytoarchitectonically, this region can be subdivided into caudal area 44 and rostral area 45 (Amunts et al., 1999). Anwander et al. (2007) found three clusters within their seed region which differed in connectivity profiles, two of which corresponding to areas 44 and 45. The third cluster covered parts of the deep frontal operculum. Similarly, the medial part of the superior parietal lobule (precuneus) and adjacent posterior cingulate cortex have been parcellated (Zhang et al., in press). Three CBP-derived clusters were identified within the superior parietal lobule which covered areas 5L and 5M (cluster 1), 7A (cluster 2), and 7P and 7M (cluster 3) as defined cytoarchitectonically (Scheperjans et al., 2008a,b). A study by Mars et al. (2011) could identify five CBP-based subregions within the right inferior parietal lobule. Comparing the ensuing clusters with respective cytoarchitectonic data (Caspers et al., 2006, 2008) showed that the CBP clusters well corresponded to inferior parietal areas PFop, PFt, PFm, PGa, and PGp. Similar results were obtained for the left inferior parietal lobule (Wang et al., 2012). Additionally, Mars et al. (2011) found five clusters within the lateral superior parietal lobule and adjacent intraparietal sulcus, corresponding to cytoarchitectonic areas 7A, 7PC, 5L, and hIP3 (Scheperjans et al., 2008a,b). The combination of CBP and atlas information was also applied to subcortical nuclei such as the amygdala, for which a recent study (Bzdok et al., in press) found three CBP clusters corresponding to cytoarchitectonic parcellations into a laterobasal, centromedial, and superficial nuclei group (Amunts et al., 2005). It has to be noted that this latter study employed CBP based on co-activation patterns as a measure for functional connectivity instead of tractography results, showing that CBP can be performed on different types of connectivity information. The resulting clusters therefore differed in function as well as co-activation patterns. The study furthermore demonstrated that the subdivision of the amygdala into three groups of nuclei converge on all levels, i.e. structure, function, and connectivity.

All these studies revealed good correspondence between CBP and cytoarchitectonic atlas derived parcellation results. It should be noted, however, that in a recent study by Thiebaut de Schotten et al. (in press), CBP derived parcellation of the occipital cortex revealed only partial correspondence with known microstructurally derived subdivisions. Thus, while the general agreement in parcellation results between different methods is encouraging, differences might occur depending on the region and potential limitations of the methods used (Cloutman and Lambon, 2012).

Assigning results to the exact anatomical location

Results from whole-brain structural or functional connectivity analyses need to be interpreted within the anatomical framework of the brain. In particular, like in task-based functional neuroimaging, localized findings, i.e., clusters of significant effects defined by some form of analysis, should be allocated to the underlying cortical anatomy, not only for more precise labeling but also to allow an interpretation in terms of distinct cortical areas. In connectome analyses, such anatomical interpretation of the data is possible for every kind of analysis, being it seed-based or independent component analysis (ICA) approaches for resting state data (Beckmann and Smith, 2005; Smith et al., 2009), localizing hubs or pathologically disturbed nodes in parcel-based and graph theory approaches (Sporns, 2009), or identifying where fibre tracts from tractography terminate (Behrens et al. 2003). Anatomical atlases provide the relevant basis for such an interpretation. Using for example the SPM Anatomy Toolbox (Eickhoff et al., 2005), results from connectivity analyses can be overlaid with cytoarchitectonic maps of the JuBrain atlas (Zilles and Amunts, 2010). Peaks representing, e.g., the most pronounced differences in connectivity of a given seed between patients and controls, may be localized to a particular cytoarchitectonic area within the probabilistic atlas framework. Furthermore, the extent of an activation can be described by identifying what percentage of the activation covers which areas (Fig. 2C). It is also possible to identify how much of the whole volume of an area is activated or connected. Activations or connection patterns can thus be characterized anatomically in much more detail as possible by reference to macroanatomical landmarks. Such view might provide additional insights into the functional or connectional architecture of a brain area, on the way towards multi-modal atlases of the brain (Roland and Zilles, 1998; Toga et al., 2006). Since microstructural features not only relate to cytoarchitecture, but also encompass other features such as the myeloarchitecture, receptor architecture, or genetic expression profiles, the definition of brain areas based on the combined use of these features for delineation will generate an atlas which provides a comprehensive microanatomical framework for the interpretation and integration of functional and connectome data from different levels and scales, as available from functional MRI, PET, diffusion or resting-state analyses. This will open the possibility to link the regional intrinsic properties of a brain region with its input and output devices, i.e. the fibres, to understand how the different features might contribute to the function of the area and the role it plays within the system brain.

Summary

Due to the interdependence of the two fundamental concepts of brain organization, it is important to incorporate information about segregation, i.e. parcellation of the grey matter into distinct areas, into analyses on the functional integration, i.e. the connectivity between areas forming a specific brain network. To derive a functionally meaningful parcellation of the grey matter to inform connectome analyses, reference should ideally be made to microstructurally derived atlases. Microstructural maps contain information about those intrinsic features of the cortex, which determine, together with its respective connectivity, the functional role of this particular patch of grey matter. Such information cannot be obtained using solely macroanatomical criteria. Since interindividual variability is considerable in the brain, probabilistic mapping approaches, which integrate information on the variability, should be preferred. Probabilistic cytoarchitectonic maps combine both advantages of being based on microstructural analyses and accounting for the interindividual variability and thus provide a thorough structural basis for mapping the connectome within an anatomical reference frame.