The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture

Subjects and Data Acquisitions

Data from 40 healthy, unrelated adults (age: 22–35, 17 males) were obtained from the Q3 data release from the Human Connectome Project (HCP) database. The multimodal MRI data consisted of structural MRI, resting-state functional MRI (rfMRI), and diffusion MRI (dMRI), collected on a 3 T Skyra scanner (Siemens, Erlangen, Germany) using a 32-channel head coil. Because subjects 209 733 and 528 446 displayed structural brain abnormalities, they were replaced by 2 other subjects, 100 408 and 106 016, from the unrelated 80 subjects' group. All scanning parameters are detailed and motivated in Van Essen et al. (2013) and also provided in the supplement. Multimodal MRI data from the database were downloaded in a preprocessed form, that is, after the images had undergone the minimal preprocessing pipeline (v. 3.2). The details of this pipeline have been described previously (Jenkinson et al. 2002, 2012; Glasser et al. 2013; Smith et al. 2013) and are only summarized in the supplement for completeness.

In addition, another independent group of healthy subjects were included to do the repeatability validation. The dataset included 40 (20 males, age range, 17–20 years, age, 19.10 ± 0.80 years, mean ± SD) right-handed participants. The multimodal MRI data of 40 healthy adults were acquired using a 3.0 T GE MR Scanner (see Zhuo et al. (2016) for a full description of the data sample and acquisition parameters).

Initial Seed Masks Definition

First, each subject's T₁ image was parcellated into 34 cortical regions of interest (ROIs) per hemisphere and 14 subcortical ROIs based on the Desikan–Killiany (DK) atlas (Desikan et al. 2006). We then combined ROIs representing (arbitrary) subdivisions of a larger gyrus as well as those whose boundaries are determined by sulci that are highly variable (cf. Supplementary Table 1). In addition, we combined the basal ganglia into a single region of interest for subsequent parcellation (Tziortzi et al. 2014). The full name and abbreviation of each initial cortical and subcortical seed mask are listed in Supplementary Table 1. All the cortical and subcortical volumetric ROIs were extracted in MNI space based on the preprocessed individual structural data. These initial seed masks in each subject were then used to create population probability maps that were binarized using a threshold of 25% to obtain the volumetric ROIs. These ensuing masks were used as a starting point for the connectivity-based parcellation analysis (Fig. 1A).

Figure 1.

Framework of the Brainnetome Atlas construction based on connectivity-based parcellation. (A) Initial parcellation using automatic surface parcellation and subcortical segmentation. The FreeSurfer DK atlas produced the initial parcellations based on gyri and sulci. (B) Tractography-based parcellation with in vivo connectional architecture. Taking the parcellation of the human paracentral lobule by diffusion tensor imaging as an example, the paracentral lobule was first extracted from the DK atlas. The connectional architecture was then mapped with probabilistic tractography using diffusion MRI, after which, by calculating the similarity/dissimilarity between the connectivity architecture, the paracentral lobule was divided into subregions with distinguishing anatomical connectivity patterns. The stability across the population and the interhemispheric anatomic homology were evaluated to determine the final cluster number. (C) Subregional anatomical and functional connections and functional behavioral decoding. Diffusion MRI combined with tractography was used to reconstruct the major fiber bundles, while functional connectivity analysis of resting-state functional MRI was used to provide the in vivo large-scale connectivity in the human brain. We also mapped the functions to each paracentral lobule subregion via the behavioral domain and paradigm analysis using the BrainMap Database.

Probabilistic Diffusion Tractography

The (HCP minimally preprocessed) diffusion MRI data were further processed using the FMRIB Diffusion Toolbox. First, fiber orientation and associated uncertainties in each voxel were estimated using FSL's BEDPOSTX algorithm, estimating probability distributions for multiple fiber directions at each voxel (Behrens et al. 2007). Second, skull-stripped T₁-weighted images for each subject were co-registered to the subject's nondiffusion-weighted image (b = 0 s/mm²). On the basis of these aligned T₁ images, we derived (forward and inverse) nonlinear transformations between the diffusion space and the MNI 152 structural template, which were then used to transform the seed masks into the diffusion space for each subject. We checked the registration accuracy of each seed region in the coronal, axial, and sagittal planes slice-by-slice in native diffusion space and manually modified the masks following the DK protocol (Desikan et al. 2006) where necessary. Then, the probabilistic tractography was applied by sampling 5000 streamline fibers for each voxel in the seed region to estimate its whole-brain connectivity profile. A small threshold value was then used to remove connectivity information of voxels that were only reached by no more than 2/5000 samples and hence most likely represent stray connectivity, that is, noise (Heiervang et al. 2006; Johansen-Berg et al. 2007; Makuuchi et al. 2009).

Connectivity-Based Parcellation Through Tractography

To facilitate data storage and analysis, the whole-brain connectivity profiles for each voxel were down-sampled to 5 mm isotropic voxels. Cross-correlation matrices between the connectivity patterns of all the voxels in the seed mask were calculated and used for automatic parcellation (Johansen-Berg et al. 2004). The cross-correlation matrix was then permutated using spectral clustering to define distinct clusters (Liu et al. 2013; Fan et al. 2014). Importantly, the number of clusters must be defined by the experimenter when using this method. In the current study, we explored parcellations from 2 up to 12 (depending on the size of the seed region) clusters (Fig. 1B; see Supplementary Figs 2–7). All (per subject) solutions were transformed into the MNI template space using the nonlinear transformations described above.

Owing to the random labeling of clusters by clustering algorithm across subjects, we try to find the most consistent labeling scheme across subjects by the following steps. First, the labeling schemes of each subject's clusters were pooled into a thresholded group-level cross-correlation matrix where each entry represents the connectional similarity of any 2 voxels in the ROI (Zhang et al. 2015). Then, the spectral clustering algorithm was applied again on this similarity matrix, and a group-level labeling scheme was, thus, yielded. Last, the labeling pattern was propagated back to each subject's clusters by maximization of spatial overlap using Munkres’ assignment algorithm (Munkres 1957). We then calculated probabilistic maps representing the overlap of these clusters across the subjects and hence interindividual variability. Further, a maximum probability map (MPM) across all the subjects was created for each solution (number of clusters) by assigning each voxel to the most likely cluster at this position (Eickhoff et al. 2005). In addition, if 2 parcels representing the same region across hemispheres are given, the consistent labeling pattern across hemispheres is also achieved before propagation of the labeling pattern.

Determination of the Optimal Clustering Solution K

To avoid arbitrarily choosing the number of clusters, we used cross-validation to determine the number of clusters that yielded the optimal consistency across the subjects. We considered 2 general strategies for defining the optimized cluster number, that is, stability across the population and interhemispheric consistency of topographic arrangement (Fig. 1B; see Supplementary Figs 2–7).

The suitable number of subregions was first assessed by evaluating the reproducibility of the parcellation across random-split half sets as measured by Cramer's V (Liu et al. 2013; Fan et al. 2014). Participants were divided into 2 random groups, and the MPMs of the 2 randomly assembled groups were evaluated. The random-split half was then repeated 100 times to compute the average (and standard deviation) consistency. The optimal K was defined by the (local) peak of Cramer's V, indicating a better split-half reproducibility than the surrounding solutions.

We also evaluated the topological distance (TpD) to quantify the similarity in topological arrangement of putatively homologous areas on either hemisphere (Tungaraza et al. 2015). To this end, we computed a K × K topology matrix for a given parcellation. The (i, j) entry of this matrix was the number of voxels from region i that were spatially in contact (26-nearest neighbor) with voxels from region j. The TpD between the left and right given subregions in each hemisphere was defined as the cosine distance of the 2 matrices (normalized to a sum of 1 and vectorized), yielding TpD scores that range from 0 (identical topology) to 1 (completely incongruent topology).

Mapping Anatomical Connectivity Patterns

To map the whole-brain anatomical connectivity pattern for each subregion of the atlas, we performed probabilistic tractography by drawing 5000 samples from each voxel in each subregion (thresholded at 25% probability) to all the other voxels of the whole brain. To reduce the noise in fiber tracking, the raw trace counts for each subject were thresholded at a trace count of ≥2. Next, we obtained a population map of the major fiber bundles for each subregion by binarizing the obtained per-subject tractograms, normalizing them into standard space, computation of a probabilistic fiber-tract map, and thresholding the latter at 50% probability. In addition, we also computed the structural connectome between all identified subregions, again binarized them using a threshold for stray connectivity of ≥2 and then employed a nonparametric 1-tailed sign test to determine the connections that are consistent across subjects (Gong et al. 2009). To reduce the chances of obtaining false-positive connections, a Bonferroni correction was used to adjust for multiple comparisons (i.e., 246 × 245/2 = 30 135 pairs of subregions) at P < 0.001.

Mapping Resting-State Functional Connectivity Patterns

To map the whole-brain resting-state connectivity pattern for each atlas subregion, we resampled each subregion (thresholded at 50% probability) at the 2 mm resolution of the resting-state data and computed its average time series per subject. A functional connectivity map was then provided by the Pearson's correlation coefficient between the mean time series of each subregion and that of each voxel in the whole brain. The obtained correlation coefficients were normalized using Fisher's z-transformation and tested for statistical significance on the group level using a 1-sample t tests (FDR corrected inference at P < 0.05 with an additional extent-threshold of 50 voxels).

Function Decoding Using the BrainMap Database

The functional characterization of the connectivity-based parcellation-yielded subregions was based on behavioral domain and paradigm class meta data labels of the BrainMap database (cf.http://www.brainmap.org/taxonomy) using forward and reverse inferences (Eickhoff et al. 2011; Cieslik et al. 2013; Clos et al. 2013; Fox et al. 2014). Forward inference indicates the probability of observing activity in a brain region given knowledge of the psychological process, whereas reverse inference is the probability of a psychological process being present given information about activation in a particular brain region. For the forward inference approach, a subregion's functional profile was determined by identifying the taxonomic labels for which the probability of finding activation in a specific subregion was significantly higher than the overall chance (across the entire database) of finding activation in that particular subregion as established using a binomial test (P < 0.05 corrected). For the reverse inference, a subregion's functional profile was determined by identifying the most likely behavioral domains and paradigm classes given activation in this subregion using Bayes' rule. Significance (P < 0.05 corrected) was assessed by means of a χ² test.

Parcellation Scheme of the Human Brain in the Brainnetome Atlas

On the basis of the connectivity architecture derived from probabilistic tractography using in vivo dMRI data, we subdivided the brain into a total of 210 cortical areas and 36 subcortical regions based on the reproducibility of the parcellation between the subjects, characterized using Cramer's V, and the interhemispheric consistency of topological relationships between clusters, characterized by the topographic distance (TpD).

Given that the anatomical connectivity maps of each identified subregion at whole-brain level were generated, we used the remaining 245 subregions as seed targets to construct a 246 × 246 connectivity matrix in which each item represents the structural connectivity of all subregions, and each row represents the fingerprint of a subregion. To supplement the validity of topological distance (TpD), here, we compared the similarity of connectivity fingerprint of each putative pair across the 2 hemispheres. Given a specified ROI, a K × K matrix was derived from the connectivity matrix of each subject. In this matrix, K denotes the label number of subregions, row and column represent the label number of subregions in ascending order in the left and right hemisphere, respectively, and each entry represents the similarity (1 − cosine distance) between the fingerprints of 2 subregions in this ROI (see details in Supplementary Figs 2–7). From the nomenclature relating label number and the corresponding name, we can conclude that TpD is an effective index that is able to find the contralateral part with the most similar connectivity profile of a subregion. For the final atlas, the MPM including all of these subregions was created in a standard MNI space (Fig. 2; see Supplementary Fig. 1). The details of the parcellation results for each initial region (and the repeatability validation results using an independent dataset) are listed in Supplementary Figures 2–7.

Figure 2.

Parcellation scheme of the human brain in the Brainnetome Atlas. The MPM for each of the cortical subregions was created in standard MNI space ((A) lateral view, (B) medial view, (C) ventral view) and visualized using ITK-SNAP (www.itksnap.org). The atlas primarily combines ontological and nomenclature information from 2 sources, that is, anatomical and modified cytoarchitectonic descriptions. For convenience, these 2 types of descriptions are separately displayed in the left and right hemispheres. The details of the parcellation results for each subregion are listed in Table 1, and the online version is available athttp://atlas.brainnetome.org/bnatlas.php. See also Supplementary Figures 1–5.

One of the challenges that arise in the context of new brain parcellation schemes is the naming of the ensuing subregions in the Brainnetome Atlas. To denominate the identified subregions at the Brainnetome Atlas, we hence employed 2 kinds of nomenclature (cf. Supplementary Table 1). In particular, we followed the primary strategy of using the DK atlas labels providing the initial macroanatomic parcellation, followed by a numeric labeling of the clusters. While preferable for not implicating potentially unwarranted correspondence with previous (microstructural) brain maps, this solution has the disadvantage of being rather hard to follow. Acknowledging that most researchers will be more familiar with labeling schemes (broadly) based on the Brodmann atlas, we thus provide tentative labels following this nomenclature, incorporating more recent refinements where available (Vogt et al. 1995; Petrides and Pandya 1999, 2002; Ongur et al. 2003; Caspers et al. 2008; Scheperjans et al. 2008; Amunts et al. 2010; Morel et al. 2013). To facilitate using the current atlas and comparing it with other atlases, we have taken our parcellation results into careful consideration for maximum consistency with existing cyto-, myelo-, or receptor-based architectonic parcellations as follows. The naming of subregions in frontal lobe is similar to that of Ongur et al. (2003), Petrides and Pandya (1999), Petrides and Pandya (2002), and Amunts et al. (2010). The naming of subregions in insular lobe is similar to that of Morel et al. (2013). The naming of subregions in parietal lobe is similar to that of Caspers et al. (2008) and Scheperjans et al. (2008). The naming of subregions in cingulate gyrus is similar to that of Vogt et al. (1995). The naming of subregions in temporal lobe is similar to that of Brodmann's descriptions. Because of the mismatch with the existing architectonic maps, we have kept the macroanatomical descriptions to name the subregions in occipital lobe.

Taking the parcellation results from the right middle frontal gyrus (MFG) and right insular cortex as an example, we identified 7 subregions in the right MFG and 6 subregions in the right insula based on the assessment of split-half reliability (Cramer's V shown in Figs 3B and 4B). MPMs for the ensuing subregions (Figs 3A and 4A) across the 40 subjects indicate the topography of the defined subregions. The MFG was found to contain MFG-1(dorsal division of area 9/46), MFG-2 (inferior frontal junction, IFJ), MFG-3 (area 46), MFG-4 (ventral division of area 9/46), MFG-5 (ventrolateral area 8, A8vl), MFG-6 (ventrolateral area 6, A6vl), and MFG-7 (lateral area, 10l). The insular cortex was composed of INS-1 (the hypergranular insula, G), INS-2 (ventral agranular insula, vIa), INS-3 (dorsal agranular insula, dIa), INS-4 (ventral dysgranular and granular insula, vId/vIg), INS-5 (dorsal granular insula, dIg), and INS-6 (dorsal dysgranular insula, dId). In addition, the probability distribution for each subregion of the MFG and insula was calculated to characterize the individual variance (Figs 3C and 4C).

Figure 3.

Brainnetome Atlas of the right middle frontal gyrus. (A) The right MFG ROI (on the left) and the MPM of the 7 subregions using a connectivity-based parcellation (on the right). We identified area 10l, area 46, dorsal and ventral divisions of area 9/46, area 6vl, area 8vl, and the IFJ subregion in the MFG. (B) The mean Cramer's V for each cluster number from 2 to 12. Cramer's V shows that 7 was the most stable solution for the right MFG. The TpD showed the similarity of the 7 solution for the topological arrangement between the 2 hemispheres. (C) The population probability maps for each MFG subregion. (D) Using the MFG-5, that is, A8vl, as an example: This figure shows the resting-state functional connectivity patterns (left), the tractographic signatures of the A8vl (middle), and the functional behavioral decoding (right).

Figure 4.

Brainnetome Atlas of the right insular cortex. (A) The right INS ROI (on the left) and the MPM of the 6 subregions using connectivity-based parcellation (on the right). Six subregions were identified in the insular cortex, including areas G, vIa, dIa, vId/vIg, dIg, and dId. (B) Cramer's V indicated that 6 was a local peak compared with the nearby solutions for the right INS. The TpD showed the similarity of the 6 solution for the topological arrangement between the 2 hemispheres. (C) The population probability maps for each INS subregion. (D) Using the INS-3, that is, area dId, as an example, this figure shows the resting-state functional connectivity patterns (left), the tractographic signatures of the area dId (middle), and the functional behavioral decoding (right).

Anatomical and Functional Connectivity Patterns in the Brainnetome Atlas

To reveal the connectivity patterns of the identified subregions in detail, we first delineated the major fiber bundles connecting the respective subregion with the rest of the brain using the whole-brain probabilistic fiber tractography (Figs 3D and 4D). All anatomical connectivity maps will be released with the Brainnetome Atlas Viewer (Fig. 6C4,7) and presented on the interactive website (Fig. 6B2). Second, we obtained a connectivity matrix representing the structural connectivity between all identified subregions as shown in Figure 5A,B for intrahemispheric and Figure 5C for interhemispheric connections. The anatomical and functional connectivity maps as well as the detailed connectogram for each brain subregion are shown on the website (http://atlas.brainnetome.org; Fig. 6B2,4). Taking the right MFG-5, that is, area 8vl, as an example, we found that this area showed connections with the major frontal subregions (including areas 44v, 45c, inferior frontal sulcus (IFS), 6vl, 8vl, 9/46d, 9m, 6m, 6dl, 9l, 8m, and 8dl), the limbic area 24rv, the parietal subregions (including areas 7pc, 39rv, and 40rd), and the subcortical connections with the thalamus and basal ganglia subregions (Fig. 5D).

Figure 5.

Connection matrices and connectogram of the Brainnetome Atlas. (A) The intrahemispheric connection matrix of the left hemisphere. (B) The intrahemispheric connection matrix of the right hemisphere. (C) The interhemispheric connection matrix across the 2 hemispheres. (D) Examples of the subregional connectograms for areas of the right MFG-5, that is, A8vl. The connectograms are represented using the Circos data visualization tool, with the left half depicting the left hemisphere and the right half depicting the right hemisphere. The hemispheres are divided into the frontal lobe, insular cortex, limbic lobe, temporal lobe, parietal lobe, occipital lobe, and subcortical structures.

Figure 6.

Resources of the Brainnetome Atlas: pipeline, Brainnetome Atlas Viewer, and interactive website. (A) Automatic Tractography-based Parcellation Pipeline (ATPP): The GUI version is single-ROI oriented; thus, it is a user friendly method that can modify some parameters to parcellate a specific brain region. The command line version is multi-ROI oriented, which can be used to parcellate many brain regions simultaneously. (B) The Brainnetome Atlas interactive website: The website makes all the information in the Brainnetome Atlas available to researchers. It provides a hierarchical tree of brain structures in the left panel (B1). The right panel contains an atlas viewer (B2 and B3) that shows slice views together with a connectogram viewer (B4). (C) The Brainnetome Atlas Viewer (V1.0): The main window of the software contains push buttons, pull-down lists, and checkboxes for different function modules (C1), such as subregion selection, template/surface selection, and connectivity visualization. The entire Brainnetome Atlas contains 246 subregions and can be viewed as a maximum probabilistic map in a tri-planar view (C2). The subregion can be viewed as a 2D overlay on the selected structural template (C3) and a 3D patch can be rendered on the cortical surface (C6). Using the checkboxes, the structural and functional connectivities of the selected subregion can be viewed in both 2D tri-planar views (C4 and C5) and 3D renderings (C7 and C8) as probabilistic maps.

Furthermore, we acquired whole-brain rfMRI connectivity maps for all the subregions in the Brainnetome Atlas. The functional connectivity maps are shared online (Fig. 6B2) and will be released for use with the Brainnetome Atlas Viewer (Fig. 6C5,8). For example, Figs 3D and 4D illustrate the functional connectivity maps of the right MFG subregion (MFG-5, A8vl) and the right insular subregion (INS-3, dIa).

Mapping the Region-to-Tasks Associations of the Brainnetome Atlas

The functional characterizations of each subregion in the Brainnetome Atlas are illustrated based on their behavioral domain and paradigm class meta data labels following the BrainMap taxonomy (http://www.brainmap.org/taxonomy). For the 2 macroanatomic brain regions used to illustrate the Brainnetome Atlas process (MFG and INS, cf. above), the functional characterization of one of the ensuing subregions is shown in Figures 3D and 4D, respectively. It showed that A8vl was significantly associated with cognitive, for example, reasoning, explicit memory, working memory, and action-related processes, such as inhibition. In terms of paradigm classes, this area was significantly associated with flanker tasks, n-back tasks, and cued explicit recognition. The dIa was involved in perception, for example, pain somesthesis, action, such as inhibition, and cognition. This insular subregion was observed to be significantly activated by paradigms associated with pain monitoring/discrimination, reward tasks, and Sternberg tasks. The functional characterization for all Brainnetome Atlas subregions will again be available through the atlas viewer on our website.

Automatic Tractography-Based Parcellation Pipeline

As part of this work, we developed an integrated “Automatic Tractography-based Parcellation Pipeline (ATPP)” to realize the parcellation using automatic processing and massive parallel computing (Fig. 6A) that we share with the atlas. ATPP is a platform that combines tractography using FMRIB's diffusion toolbox with in-house MATLAB scripts for parcellation. It uses Oracle Grid Engine and MATLAB's Parallel Computing Toolbox for parallel computing across and within machines. Both a command line version and a graphical user interface (GUI) version are available. The GUI version is single-ROI oriented and therefore a user friendly method that allows the targeted analysis of any brain region defined, for example, by functional or structural findings (Cieslik et al. 2013; Muhle-Karbe et al. 2015). That is, while the Brainnetome Atlas provides a whole-brain parcellation at a (necessarily) coarser level, the distribution of the ATPP will allow researchers to use the same framework to address specific questions on local brain organization.

Interactive Website Viewer

The Brainnetome Atlas website, available athttp://atlas.brainnetome.org, allows to fully explore the atlas and the various information associated with each subregion (Fig. 6B). To improve user interaction, the following functionalities were included: 1) Clicking a structure name in the tree will reveal that subregion in the slice viewer. At the same time, the behavioral domain and paradigm class analysis results, together with the anatomical and functional connectivity patterns, are updated in the atlas viewer. 2) Structural and functional connectivity maps are displayed with the possibility of full angle rotation. 3) By picking a bar that indicates another structure in the connectogram viewer, the user can navigate the connected structure and visualize the related information in the atlas viewer. 4) Users can search for a structure of interest. All branches with related keywords are highlighted so that the user can click to visualize the node.

StandAlone Brainnetome Atlas Viewer

We likewise provide a standalone Brainnetome Atlas Viewer running under MATLAB with a user friendly GUI (Fig. 6C), featuring subregion selection, template/surface selection, and connectivity visualization. The entire Brainnetome Atlas can be viewed as a maximum probabilistic map in a triplanar view. Navigation through the Brainnetome Atlas can be synchronized with a collection of widely used canonical templates/atlases. Once a subregion is selected, various information about this structure are displayed. Publications related to the selected subregion, if any, can be visited on PubMed. Subregions can be viewed as a 2D overlay of the MPM representation or the probabilistic map on the selected structural template and a 3D patch can be rendered on the cortical surface. Probably the key feature, however, is the possibility to generate ROI mask for further analyses by selecting atlas structures and probability thresholds.

Brainnetome Atlas: Towards a Framework for Multimodal Information Integration

Because of the variety of needs of different fields within neuroscience and the shortcomings of existing brain atlases, a new human brain atlas with a framework for integrating multimodal information is urgently needed (Evans et al. 2012; Amunts et al. 2014). Consequently, many studies have used different MRI modalities to identify individual brain region or provide more comprehensive maps of the cerebral cortex (Tzourio-Mazoyer et al. 2002; Desikan et al. 2006; Cohen, Fair, et al. 2008; Cohen, Lombardo, et al. 2008; Eickhoff et al. 2011; Wang et al. 2012; Fan et al. 2014; Wig et al. 2014; Laumann Timothy et al. 2015; Liu et al. 2015; Yang et al. 2015). While acknowledging that there is no consensus on which modality or aspect of brain organization may be most reflective of the brains’ “true” organization (and in fact, there may be no single answer to this question), brain atlases are crucial to advance understanding of the human brain given that macroanatomical landmarks or coordinate systems are not valid indicators of regional specialization (Brett et al. 2002; Bohland et al. 2009; Evans et al. 2012). The Brainnetome Atlas addresses this need by providing a whole-brain parcellation of the human brain into distinct subregions based on the local structural connectivity architecture, that is, by identifying subregions that are maximally different from each other and maximally homogeneous internally in terms of their white matter connections. It thus provides an objective and stable starting point from which to explore the complex relationships between structure, connectivity, and function featuring the following advantages.

• Robust and biologically plausible anatomical parcels: An increasing number of in vivo neuroimaging studies have demonstrated that fine-grained parcellations of large regions of the human brain are needed. The Brainnetome Atlas not only confirmed several differentiations from earlier cytoarchitectonic maps but also revealed numerous anatomical subdivisions that were not previously described (Liu et al. 2013; Fan et al. 2014). For example, in the atlas, the insular cortex has been subdivided into 6 subregions (Fig. 2; see Supplementary Fig. 5A). Although the correct parcellation of the human insular cortex has been disputed (Cauda and Vercelli 2013), the current parcellation, showing the highest split-half reliability of all cluster solutions, relates well to previous functional and histological maps of the insular cortex (Kurth, Eickhoff, et al. 2010; Kurth, Zilles, et al. 2010; Kelly et al. 2012; Chang et al. 2013; Morel et al. 2013), identifying a dorsal and ventral aspect of the anterior insula, a central cluster, a more ventral component, and (2) posterior subregions. As another example, the inferior frontal gyrus contains 6 subregions that were robustly identified across subjects: the dorsal/ventral portions of area 44, the rostral and caudal subdivisions of area 45, the inferior fontal sulcus, and 1 cluster on the frontal operculum, which correspond to a combination of areas op8 and op9 (Fig. 2; see Supplementary Fig. 2C). This parcellation compares well with a parcellation scheme that was proposed based on the differential distribution of transmitter receptors (Amunts et al. 2010) and the differential inclusion in task activation networks (Clos et al. 2013).

• Detailed description of anatomical and functional connections: One of the major drawbacks of several previous atlases is that they present rather isolated information, that is, are only reflective or brain organization from a single perspective. The Brainnetome Atlas in turn has been designed to enrich our understanding of human brain organization from a multimodal perspective. Combining diffusion MRI with tractography could allow reconstruction of the major fiber bundles, while functional connectivity analyses of BOLD functional MRI data could also provide a noninvasive way to assess in vivo the large-scale connectivity of the human brain. These approaches should enable researchers to identify connectivity patterns and relate this information to the parcellations in the Brainnetome Atlas. This should be useful in detailed investigations of particular subsystems.

• Functional interpretation of the Brainnetome Atlas: One of the key challenges in human brain research is to reconcile the regional segregation of the cortex into distinct modules with the representation of mental functions as provided by task-based neuroimaging data. Because it was designed as a framework for a variety of resources, the Brainnetome Atlas is suitable for integrating a wealth of information from existing neuroimaging studies. In this study, we used BrainMap (Laird et al. 2009, 2011; Fox et al. 2014), currently the largest database of brain activation studies, to provide an initial assessment of the mental processes that may be sustained by each subregion of the Brainnetome Atlas. These descriptions, which are shared with the community, thus provide an objective guideline to the functional interpretations of any effects observed within a given Brainnetome Atlas subregion.

• Data sharing: The Brainnetome Atlas can be flexibly wrapped into common reference spaces, such as the volumetric MNI space, the vertex-based FreeSurfer, or the Caret surface template. The atlas together with its related software is available for download to serve as a shared community resource (http://atlas.brainnetome.org, Fig. 6B). The pipeline software is open to the community to facilitate the parcellation of specific brain regions of interest (Fig. 6A). The Brainnetome Atlas Viewer was coded in MATLAB so that it can easily be implemented into commonly used brain MRI processing pipelines (Fig. 6C). In addition, the atlas will be useful for the definition of masks for seeding specific a priori cortical regions or networks of interest in prospective neuroimaging studies.

Relationship Between Connectivity Architecture and Microstructural Organization

Animal studies have indicated that connectivity architecture and microstructural organization are 2 complementary anatomical properties of the brain (Scannell et al. 1995; Barbas and Rempel-Clower 1997; Passingham et al. 2002; Hilgetag and Grant 2010). For example, by observing the laminar organization of areas in the monkey prefrontal cortex, Barbas and Rempel-Clower (1997) found that the cortical structure could predict the pattern and relative laminar distribution of cortico-cortical connections. Furthermore, it has been shown that cortico-cortical connectivity patterns exert strong effects on both the anatomical and functional characteristics of a specific brain region (Rempel-Clower and Barbas 2000; Glickfeld et al. 2013). Assessing the relationship between microstructure and connectivity is important to advance our understanding of brain organization, but a direct comparison between functional in vivo and anatomical postmortem studies is possible only in experimental animals. In turn, studies on the relationship between microstructure and function or connectivity in humans need to rely on across-subject analysis by means of probabilistic cytoarchitectonic maps in standard space (Eickhoff et al. 2005; Zilles and Amunts 2010). Such analyses have provided evidence for distinctions between (neighboring) histologically defined areas in terms of their structural and functional connectivity patterns (Eickhoff et al. 2010; Bludau et al. 2014; Palomero-Gallagher et al. 2015).

The importance of connectivity in determining functional specialization, however, suggests that parcellating brain regions based on their connectional architecture may provide important complementary information on human brain organization. For example, a recent study predicted which parts of the fusiform gyrus are involved in face recognition based on structural connectivity patterns (Saygin et al. 2012). In addition, it has been shown that connectivity-based parcellation may closely follow histological subdivisions (Anwander et al. 2007; Klein et al. 2007) but also reveal additional subdivisions (Liu et al. 2013; Fan et al. 2014). The latter case is actually not particularly surprising, considering that intrinsic (cyto- or receptorarchitecture) and extrinsic (long-range connectivity) conjointly define the functional property of a given brain location (Eickhoff and Grefkes 2011). That is, it may be the intersection between microstructural and connectivity maps that defines specialization.

Closing the gap between these 2 approaches to parcellation still needs to be addressed. Specifically, additional work is needed to untangle the complex relationship between the brain's microstructural organization and its connectivity architecture as well as their roles in determining the brain's functional organization. In a recently published study (van den Heuvel et al. 2015), researchers performed a cross-scale examination and found that the organization of macroscale connectivity derived from diffusion MRI data correlates with cortical variation in cytoarchitectonics, in particular with the size of neurons in cortical layer 3. Other studies have also provided first experimental evidence that brain function arises from a combination of local infrastructure (microstructural/molecular features) and connectivity (Passingham et al. 2002; Scholtens et al. 2014; Barbas 2015). Therefore, combining connectivity-based parcellation results with probabilistic maps of microstructure could hold great promise for relating brain structure to function at the macroscopic scale.

Connectivity-Based Parcellation Using Multimodal Connectivity Information

Many approaches to parcellating the brain into subregions using different connectivity features have recently become available. These include tractography-based anatomical connectivity (Behrens et al. 2003; Johansen-Berg et al. 2004), resting-state functional connectivity (Cohen, Fair, et al. 2008; Cohen, Lombardo, et al. 2008; Nelson et al. 2010), structural covariance (Cohen, Fair, et al. 2008; Cohen, Lombardo, et al. 2008; Kelly et al. 2012), and meta-analysis-based functional coactivation (Eickhoff et al. 2011). However, the biological basis of these methods and what we can infer from these imaging modalities are not yet fully understood (Eickhoff et al. 2015). The current version of the Brainnetome Atlas was created using structural connectivity patterns as estimated by diffusion-weighted imaging and hence primarily relies on direct connections as opposed to indirect, multisynaptic interactions as revealed, for example, by resting-state or task-based functional connectivity. Several studies have focused on the relationship between anatomical and functional connectivity indicating that resting-state connectivity (Honey et al. 2009; van den Heuvel et al. 2009) and meta-analytic coactivations (Eickhoff et al. 2010) at least to some degree reflect the underlying anatomical connectivity architecture of the human brain. Finally, it has been shown that structural covariance patterns are largely in agreement with the aforementioned methods (Kelly et al. 2012). Although its biological significance remains controversial, it is hence assumed that functional connectivity should at least contribute to the patterns of structural covariance (Mechelli et al. 2005; Alexander-Bloch et al. 2013).

The other important aspect to consider in the construction of brain atlases, or more general, the definition of brain regions is the distinction between methods focused on either the clustering or the detection of borders. While similar in their final appearance, the former are driven by aggregating locations (voxels) with similar properties, whereas the later are aimed at identifying abrupt changes in the respective feature. The second relates to the nature of the features being either locally (such as cyto- or receptorarchitecture) or globally (such as connectivity profiles). To synthesize the existing brain parcellation studies, we may hence distinguish a “2 × 2” matrix of ways that have been used to generate brain atlases: regionally versus globally and by clustering or by border detection. The JuBrain cytoarchitectonic atlas and other histology-based atlases are examples of regional border detection (Brodmann 1909; Von Economo and Koskinas 1925; Zilles and Amunts 2010), while whole-brain parcellation studies based on rfMRI or dMRI may be realized by either global border detection (Wig et al. 2014; Gordon et al. 2016) or global clustering (Craddock et al. 2012; Moreno-Dominguez et al. 2014). In this framework, the Brainnetome Atlas would represent an example of clustering based on a global feature (connectivity).

Given the heterogeneity of both features and methods that may be used to parcellate the brain into distinct subregions, one of the particular challenges will be to examine the consistency or inconsistency of the ensuing parcellations and to evaluate different brain parcellation schemes. Using resting-state connectivity, meta-analytic coactivation and structural covariance, but not fiber tracking, Kelly et al. (2012) found a consistent pattern in the parcellations of the insula. In our recent work, we consistently identified 5 subregions in the superior parietal lobule of each hemisphere based on its anatomical connections as well as its resting-state connectivity and coactivation patterns (Yang et al. 2015). Further systematic comparison—across modalities, features, and methods—of the maps that may be computed using connectivity-based parcellation is still needed. Such integration would not only be crucial to obtain a more comprehensive picture of human brain organization, but in particular also to better understand the relationship between the different approaches and to arrive at a mechanistic relationship between the different aspects of brain organization (Amunts et al. 2014).

Methodological Considerations

The Brainnetome Atlas provides a cross-validated, robust group-level parcellation of the human brain, but ultimately individual, subject-level parcellations will be required to reflect interindividual variability in the location of brain modules (Barnes et al. 2011; Fonov et al. 2011; Laumann Timothy et al. 2015). In that context, we note that the parcellation current scheme is consistent with our previous parcellations of specific regions, including the frontal pole (Liu et al. 2013), temporal pole (Fan et al. 2014), parahippocampal region (Zhuo et al. 2016), and superior (Wang et al. 2015) and inferior parietal lobules (Wang et al. 2012) based on a different set of subjects. While reassuring, it still remains to be tested, how reliable individual brains may be parcellated using structural connectivity information and how the ensuing maps relate to other aspects of interindividual variability such as age and gender.

In forming the Brainnetome Atlas, we identified the following key issues for further consideration: 1) Development of reliable clustering algorithms and effective measures for validating the quality of parcellations needs to be further explored. Here, we used across-subject consistency as the key cluster-validity criterion, which is in line with previous work (Beckmann et al. 2009; Liu et al. 2013; Fan et al. 2014; Neubert et al. 2014), but may introduce a bias against more (spatially) variable patterns. 2) Maintaining the macroscopically visible sulcal and gyral anatomy so that we could provide an intuitive description of the location of the activations motivated by the use of the DK atlas as the initial parcellation. We have attempted to follow the widely accepted anatomical conventions and then to further divide the brain regions into subregions using tractography-based parcellation. It should be acknowledged, however, that the DK atlas boundaries provide an a priori parcellation that may not be related to the actual differentiation of the cortex into distinct areas based on pure connectivity profiles. Additionally, the macroanatomy is generally a rather poor predictor of microstructural boundaries, and hence, further work may be warranted on the relationship between the Brainnetome Atlas and functional or microstructural parcellations at various scales. 3) To calculate the population-based probabilistic maps, it is vital to register the individual parcellation results to the commonly used MNI space. Recently, few groups have begun trying to work with the structural connectivity-based registration or fusing the connectivity information with the cortical anatomy (Gutman et al. 2014; Wang et al. 2014). However, such kind of registration methods based on structural connectivity is still very experimental and not well developed yet. The macro-anatomy-based registration is still the standard of the field, and we are keeping with the established approach in forming the Brainnetome Atlas. 4) The other important issue for this framework is the validation of the connectivity-based brain atlas (Klein et al. 2007; Gordon et al. 2016; Lefranc et al. 2016). The connectivity-based boundaries within the initial macroanatomical cortical areas are validated using another independent dataset with different scan parameters. The reproducible results from the independent dataset show good consistency with the parcellation results based on the HCP data.

Conclusions and Outlook

The long-term aim of the Brainnetome Project is to understand the organizing principles of the brain. The current version of the Brainnetome Atlas will facilitate investigations into structure–function relationships and holds promise for the identification of new biomarkers for diagnostic and clinical studies. In the future, however, novel methodologies and brain-mapping techniques should evolve and allow an enhanced assessment of the structure, function, and spatiotemporal changes in the human brain on different spatial and temporal scales. The Brainnetome Atlas should thus be regarded as a starting point, which will enable the generation of future brain atlases that are even more fine-grained and advance from single anatomical descriptions to an integrated atlas that includes structure, function, and connectivity, together with other potential sources of information (Amunts et al. 2014). The next stage of the Brainnetome Atlas will be multimodal instead of unimodal and dynamic instead of static by including information on spatiotemporal changes during normal development or aging as well as disease-related effects. Finally, integration with gene expression data should provide entirely novel insights into human brain organization. A recent study that analyzed brain imaging and gene expression data found that large-scale resting-state functional brain networks correlated with the expression of genes that code for ion channels and other synaptic functions (Richiardi et al. 2015). In a preliminary study of the relationship between genetics and brain parcellation, we noninvasively investigated the genetic influences on a fine-grained topological arrangement of the human cerebral cortex using MRI data from twins (Cui et al. 2015). While far from being understood and readily used, such genetic information will be crucial for the next-generation human brain atlas by linking phenotypically observed effects to genetic causes. While human brain atlasing is thus not only an endeavor that has been ongoing for more than a century but also one that will see constant changes and refinement, the current Brainnetome Atlas represents an important step in this development by providing the first whole-brain parcellation based on structural (connectivity) information on the basis of a robust cross-validation in a high-quality in vivo dataset.