1. Home
  2. Brain Parcellation Survey

Brain Parcellation Survey

Overview

This page provides the surface-based parcellations evaluated in the brain parcellation survey, Human Brain Mapping: A Systematic Comparison of Parcellation Methods for the Human Cerebral Cortex, in which a large-scale and systematic comparison is carried out between the state-of-the-art connectivity-driven, anatomical, and random parcellation methods. Please click here to download a copy of the accepted manuscript.

Using resting-state functional MRI (rs-fMRI) data and several quantitative evaluation techniques, 10 subject-level and 24 groupwise parcellation methods are evaluated at different resolutions. The accuracy of parcellations is assessed from four different aspects: (1) reproducibility across different acquisitions and groups, (2) fidelity to the underlying connectivity data, (3) agreement with fMRI task activation, myelin maps, and cytoarchitectural areas, and (4) network analysis.

This extensive evaluation of different parcellations generated at the subject and group levels highlights the strengths and shortcomings of the various methods and aims to provide a guideline for the choice of parcellation technique and resolution according to the task at hand.

Data

This study is carried out using data from the publicly available Human Connectome Project (HCP) S900 release. All connectivity-driven parcellations are derived from the rs-fMRI acquisitions of 100 unrelated subjects (54 female, 46 male adults, aged 22-35). This dataset is publicly available as the “Unrelated 100” in the HCP database and is referred to as “Dataset 1” in the paper. For evaluation purposes, we gather a different set of 100 unrelated subject from the HCP database (i.e. Dataset 2) comprising randomly selected 50 male and 50 female adults of age 22-35.

Parcellations

You can find all subject-level and groupwise parcellations here. Once clicked, you will be redirected to the Imperial College London data storage (powered by Box) and be able to navigate between different methods/subjects. Simply click the ‘Download‘ button as shown in the following figure, to start downloading all the parcellations and additional files, as well as several scripts to read and visualise the parcellation data in MATLAB and Python.

In order to walk you through the parcellation files, we have provided a reference manual alongside the parcellation data, which can be accessed here (or can be found in the extracted zip, if you have proceeded with the download by following the instructions above). Your are strongly advised to have a look at the manual for a better understanding of the parcellation data (including the directory structure and file names) as well as learning more about the implementation details and processing decisions made throughout this study.

Alternatively, you can download any of the parcellations from the following tables (again via Box). We have grouped 10 subject-level and 24 groupwise parcellations into three tables, namely, Single Subject Parcellations (Table 1), Computed Groupwise Parcellations (Table 2), and Publicly Available Groupwise Parcellations (Table 3).

Each table is divided into five columns with the following attributes: [Name] Name of the method, as referred to in the paper. [Reference] Reference to the parcellation method. It may also include links to online resources, from which the source codes and/or original parcellations can be downloaded. [Resolution] The number of cortical regions a parcellation is comprised of. Some parcellations are provided at varying resolutions (generally ranging from 25 to 250 regions in increments of 25, per hemisphere). [Description] Very brief information about the parcellation method and pre/post processing steps. [Download] Links for downloading parcellations individually.

Single Subject Parcellations

All parcellations included in the subject-level analysis are generated from the individual rs-fMRI data in Dataset 1. Each subject is represented by two sets of 30-minute BOLD fMRI timeseries (rs-fMRI 1 and rs-fMRI 2), which allows the production of two data-driven parcellations per subject.

Name Reference Resolution Description Download
 

Arslan

 

  

Arslan and Rueckert (2015); source codes available here.

 

  

Varying

 

  

A two-level approach that combines k-means and hierarchical clustering. Ward’s linkage rule (Ward, 1963) with Euclidean distance is applied to an initial finer parcellation (i.e. supervertices) of 1000 regions per hemisphere.

 

  

 
 

Blumensath

 

  

Blumensath et al. (2013); re-implemented as described in the original paper.

 

  

Varying

 

  

A two-level method that combines region growing and hierarchical clustering. Ward’s linkage rule (Ward, 1963) with Euclidean distance is applied to an initial finer parcellation of 1000 regions per hemisphere.

 

  

 
 

Bellec

 

  

Bellec et al. (2006); source codes available here.

 

  

Varying

 

  

A competitive region growing approach driven by parcel homogeneity. A size threshold is applied to avoid over-growing of parcels.

 

  

 
 

Ward

 

  

Ward (1963); in-house implementation, featuring scikit-learn’s AgglomerativeClustering function.

 

  

Varying

 

  

A hierarchical tree is built by merging pairs of clusters, if their similarity is the maximal among the other pairing clusters. Only adjacent clusters are joined into a higher level in order to ensure spatial contiguity. Clustering is driven by Ward’s linkage rule (Ward, 1963) and Euclidean distance.

 

  

 
 

K-Means

 

  

k-means clustering as described in Thirion et al. (2014); in-house implementation, featuring scikit-learn’s KMeans and PCA functions.

 

  

Varying

 

  

PCA is applied to BOLD timeseries for feature reduction. PCA components that explain 50% of the variance combined with spatial vertex coordinates to improve the spatial connectedness of parcellations.

 

  

 
 

N-Cuts

 

  

Craddock et al. (2012); in-house implementation of spectral clustering featuring discretisation (Yu and Malik 2003).

 

  

Varying

 

  

Spectral clustering with normalised cuts. An affinity matrix is built by cross-correlating the adjacent vertices with each other. Spectral decomposition is applied to the normalised graph Laplacian. The final parcellations are obtained by discretising the eigenvectors.

 

  

 
 

Destrieux

 

  

Fischl et al. (2004); provided for each subject in the HCP database.

 

  

150                    (75 L, 75 R)

 

  

An anatomical surface-based parcellation that subdivides the cortex with respect to the limit between the gyral and sulcal regions.

 

  

 
 

Desikan

 

  

Desikan et al. (2006); provided for each subject in the HCP database.

 

  

70                       (35 L, 35 R)

 

  

A surface-based parcellation that subdivides the cortex with respect to anatomical landmarks based on the gyri.

 

  

 
 

Geometric

 

  

Geometric parcellations as described in Thirion et al. (2014); in-house implementation, featuring scikit-learn’s KMeans function.

 

  

Varying

 

  

k-means clustering is applied to the spatial vertex coordinates. No connectivity information is accounted for.

 

  

 
 

Random

 

  

Random parcellations as described in Schirmer et al. (2015).

 

  

Varying

 

  

Poisson disk sampling is used to generate regions of approximately equal size by ensuring that two region centres are not closer than a given threshold that controls the number of parcels.

 

  

 

Computed Groupwise Parcellations

We provide two group-level parcellations for each data-driven method, one obtained from Dataset 1 and the other derived from Dataset 2 (this primarily allows to evaluate group-to-group reproducibility).

Name Reference Resolution Description Download
Arslan
2-level group-wise parcellations obtained from subject-level Arslan.
Varying
2-level approach is applied to the subject-level Arslan parcellations to obtain group-level parcellations. 
Blumensath
2-level groupwise parcellations obtained from subject-level Blumensath.
Varying
2-level approach is applied to the subject-level Blumensath parcellations to obtain group-level parcellations. 
Bellec
2-level groupwise parcellations obtained from subject-level Bellec.
Varying
2-level approach is applied to the subject-level Bellec parcellations to derive group-level parcellations. 
Ward-2
2-level groupwise parcellations obtained from subject-level Ward.
Varying
2-level approach is applied to the subject-level Ward parcellations to derive group-level parcellations. 
K-Means-2
2-level groupwise parcellations obtained from subject-level K-Means.
Varying
2-level approach is applied to the subject-level K-Means parcellations to acquire group-level parcellations. 
N-Cuts-2
2-level groupwise parcellations obtained from subject-level N-Cuts.
Varying
2-level approach is applied to the subject-level N-Cuts parcellations to acquire group-level parcellations. 
Ward-AVR
Ward (1963); in-house implemen-tation featuring scikit-learn’s AgglomerativeClustering function.
Varying
The group average matrix is fed into the Ward’s agglomerative hierarchical clustering algorithm using the same setting as for the subject-level Ward parcellations.
K-Means-AVR
k-means clustering as described in Thirion et al. (2014); in-house implementation featuring scikit-learn’s KMeans and PCA functions.
Varying
The group average matrix is fed into k-means clustering after being concatenated with the average spatial coordinates to improve the spatial contiguity of parcellations.
N-Cuts-AVR
Craddock et al. (2012); in-house implementation of spectral decomposition featuring discretisation (Yu and Malik 2003).
Varying
A temporal correlation matrix is derived from the group average matrix and transformed into a spatially constrained affinity matrix. Spectral clustering with normalised cuts is used as in the same setting as the subject-level N-Cuts parcellations.
JOINT
Arslan et al. (2015); source codes available here.
Varying
A surface-based parcellation method based on a joint spectral decomposition of individual subjects. An initial finer parcellation of 2000 regions per hemisphere is used for spatial feature reduction in order to compensate for the computational cost.
GRASP
Honnorat et al. (2015); parcellations computed based on the publicly available codes.
Varying
A method based on Markov Random Field (MRF) that subdivides the cortex into spatially contiguous parcels by using shape priors. The group average matrix is parcellated into 10000 initial clusters by running the method in the hierarchical clustering mode. Final parcellations are derived from this low-dimensional matrix.
GRAMPA
Parisot et al. (2016); parcellations computed from an in-house implementation of the method. Codes will be available here.
Varying
An MRF model that iteratively updates parcel centres and parcel assignments based on modality specific costs (we utilise the unimodal approach based on rs-fMRI). The parcellation is computed using the group average matrix.
Geometric
Geometric parcellations computed as described in Thirion et al. (2014); in-house implementation featuring scikit-learn’s KMeans function.
Varying
k-means clustering is applied to the average spatial vertex coordinates. No connectivity information is accounted for.

Publicly Available Groupwise Parcellations 

Publicly available parcellations included in this study can be downloaded from the following table. Alternatively, you can visit the external websites that host the original parcellations via the links provided in Reference. It is important to note that, we applied additional processing to certain parcellations in order to make them comparable on a more standard basis. These alterations are summarised in Description and further explained in the data reference manual. Not to mention, some of the parcellations may be different from what is originally proposed.

Name Reference Resolution Description Download
Gordon
Gordon et al. (2016); original surface-based parcellation available here.
333                   (161 L, 172 R)
A surface-based parcellation computed from the average gradients of resting-state functional connectivity networks. Provided parcellation is iteratively dilated to cover the entire cortical surface.
Power
Power et al. (2011), original surface-based parcellation available here.
130                   (65 L, 65 R)
Resting-state communities originally identified in volume space are projected onto the cortical surface and made publicly available by Van Essen et al. (2016). Connected components within each parcel are relabelled to ensure spatial contiguity.
Yeo
Yeo et al. (2011); original surface-based parcellation available here.
96                       (49 L, 47 R)
17-cluster resting-state networks originally derived in volume space from average resting-state functional connectivity data using a GMM-based clustering algorithm are projected onto the cortical surface and made publicly available by Van Essen et al. (2016). Connected components in each parcel are relabeled to ensure spatial contiguity.
ICA
Beckmann et al. (2004), Smith et al. (2014); original surface-based parcellations available here.
Varying
Group-average parcellations by means of group-ICA (Beckmann et al. 2004) are obtained at several different dimensionalities (25, 50, 100, 200, 300), using a group-PCA output (Smith et al. 2014) from the HCP S500 subjects. Connected components within each parcel are relabeled.
Baldassano
Baldassano et al. (2015); original surface-based parcellations and source code available here.
171                     (84 L, 87 R)
A multi-purpose clustering algorithm based on nonparametric Bayesian modelling is applied to dense connectome derived from the HCP S500 group PCA output (Smith et al. 2014) in order to compute a surface-based parcellation.
Glasser
Glasser et al. (2016); original surface-based parcellation (HCP_MMP1.0) available here.
360                 (180 L, 180 R)
A cortical parcellation generated from multi-modal images of 210 adults from the Human Connectome Project database, using a semi-automated approach. Cortical regions are delineated with respect to function, connectivity, cortical architecture, and topography, as well as, expert knowledge and meta-analysis results from the literature.
Fan
Fan et al. (2016); original volumetric parcellation available here.
210                     (105 L, 105 R)
A volumetric brain parcellation is obtained using both anatomical landmarks and connectivity-driven information. Anatomical regions defined by Desikan et al. (2006) are parcellated into subregions using functional and structural connectivity data from 40 adults provided by the HCP. Cortical parcels are projected from volume to surface.
Shen
Shen et al. (2013); original volumetric parcellation available here.
200                     (102 L, 98 R)
A spectral clustering approach is used to compute a volumetric groupwise parcellation based on an optimisation process that guarantees functional homogeneity within each parcel and ensures that computed parcels are consistent across subjects. Volumetric parcels from the provided 1 mm sampled 268-parcel atlas are projected to cortical surface.
AAL
Tzourio-Mazoyer et al. (2002); original volumetric parcellation available here.
82                     (41 L, 41 R)
The popular Automated Anatomical Labelling atlas, which is manually delineated with respect to anatomical landmarks, in particular, by following the sulci course in the brain. Cortical parcels are projected from volume to surface.
Destrieux
Fischl et al. (2004); original surface-based parcellations available here.
150                     (75 L, 75 R)
Majority voting is applied across the subject-level Destrieux parcellations to obtain a group-level parcellation.
Desikan
Desikan et al. (2006); original surface-based parcellations available here.
70                       (35 L, 35 R)
Majority voting is applied across the subject-level Desikan parcellations to obtain a group-level parcellation.

Evaluation Code

Scripts/functions used to carry out the experiments for the evaluation of parcellations with respect to reproducibility, clustering validity, and agreement with other modalities are made publicly available. If you have downloaded the parcellation data by following the instructions above (i.e. via Box), you can also find a copy of each script in the “eval” folder under “Scripts”. Please see in-file instructions/comments in order to run them.

References

If you use any of the parcellations (or evaluation code) for your own research, you are kindly asked to cite the following paper (along with the original paper of the parcellation) in your publications.

Arslan, S., Ktena, S.I., Makropoulos, A., Robinson, EC., Rueckert, D., Parisot, S., 2017, Human brain mapping: A systematic comparison of parcellation methods for the human cerebral cortex, NeuroImage, doi: 10.1016/j.neuroimage.2017.04.014

Arslan, S., Parisot, S., Rueckert, D., 2015. Joint spectral decomposition for the parcellation of the human cerebral cortex using resting-state fMRI, In International Conference on Information Processing in Medical Imaging (pp. 85-97).
Arslan, S., Rueckert, D., 2015. Multi-level parcellation of the cerebral cortex using resting-state fMRI, In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 47-54).
Baldassano, C., Beck, D.M., Fei-Fei, L., 2015. Parcellating connectivity in spatial maps. PeerJ 3, p. e784.
Beckmann, C.F., Smith, S.M., 2004. Probabilistic independent component analysis for functional magnetic resonance imaging. IEEE Trans Med Imag 23, pp. 137-152.
Bellec, P., Perlbarg, V., Jbabdi, S., Pelegrini-Issac, M., Anton, J.L., Doyon, J., Benali, H., 2006. Identification of large-scale networks in the brain using fMRI. NeuroImage 29(4), pp.1231-1243.
Blumensath, T., Jbabdi, S., Glasser, M.F., Van Essen, D.C., Ugurbil, K., Behrens, T.E., Smith, S.M., 2013. Spatially constrained hierarchical parcellation of the brain with resting-state fMRI. NeuroImage 76, pp. 313-324.
Craddock, R.C., James, G.A., Holtzheimer, P.E., Hu, X.P., Mayberg, H.S., 2012. A whole brain fMRI atlas generated via spatially constrained spectral clustering. Hum Brain Mapp 33, pp. 1914-1928.
Desikan, R.S., Segonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker, D., Buckner, R.L., Dale, A.M., Maguire, R.P., Hyman, B.T., et al., 2006. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. NeuroImage 31, pp. 968-980.
Fan, L., Li, H., Zhuo, J., Zhang, Y., Wang, J., Chen, L., Yang, Z., Chu, C., Xie, S., Laird, A.R., et al., 2016. The human brainnetome atlas: A new brain atlas based on connectional architecture. Cerebral Cortex 26(8), pp. 3508-3526.
Fischl, B., van der Kouwe, A., Destrieux, C., Halgren, E., Segonne, F., Salat, D.H., Busa, E., Seidman, L.J., Goldstein, J., Kennedy, D., et al., 2004. Automatically parcellating the human cerebral cortex. Cerebral Cortex 14, pp. 11-22.
Glasser, M., Coalson, T., Robinson, E., Hacker, C., Harwell, J., Yacoub, E., Ugurbil, K., Anderson, J., Beckmann, C., Jenkinson, M., et al., 2016. A multi-modal parcellation of human cerebral cortex. Nature 536(7615), pp. 171-178.
Gordon, E.M., Laumann, T.O., Adeyemo, B., Huckins, J.F., Kelley, W.M., Petersen, S.E., 2016. Generation and evaluation of a cortical area parcellation from resting-state correlations. Cerebral Cortex 26(1), pp. 288-303.
Honnorat, N., Eavani, H., Satterthwaite, T., Gur, R., Gur, R., Davatzikos, C., 2015. Grasp: geodesic graph-based segmentation with shape priors for the functional parcellation of the cortex. NeuroImage 106, pp. 207-221.
Parisot, S., Glocker, B., Schirmer, M.D., Rueckert, D., 2016. GraMPa: Graph-based multi-modal parcellation of the cortex using fusion moves,In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 148-156).
Power, J.D., Cohen, A.L., Nelson, S.M., Wig, G.S., Barnes, K.A., Church, J.A., Vogel, A.C., Laumann, T.O., Miezin, F.M., Schlaggar, B.L., Petersen, S.E., 2011. Functional network organization of the human brain. Neuron 72, pp. 665-678.
Schirmer, M.D., 2015. Developing brain connectivity: Effects of parcellation scale on network analysis in neonates. Ph.D. thesis. King’s College London.
Shen, X., Tokoglu, F., Papademetris, X., Constable, R.T., 2013. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. NeuroImage 82, pp. 403-415.
Smith, S.M., Hyvarinen, A., Varoquaux, G., Miller, K.L., Beckmann, C.F., 2014. Group-PCA for very large fMRI datasets. NeuroImage 101, pp. 738-749.
Thirion, B., Varoquaux, G., Dohmatob, E., Poline, J.B., 2014. Which fMRI clustering gives good brain parcellations? Front Neurosci 8, p. 167.
Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., Joliot, M., 2002. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. NeuroImage 15, pp. 273-289.
Van Essen, D.C., Smith, J., Glasser, M.F., Elam, J., Donahue, C.J., Dierker, D.L., Reid, E.K., Coalson, T. and Harwell, J., 2017. The brain analysis library of spatial maps and atlases (BALSA) database. NeuroImage 144, pp. 270-274.
Ward, J.H., 1963. Hierarchical grouping to optimize an objective function. J Amer Statist Assoc 58, pp. 236-244.
Yeo, B.T.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D., Hollinshead, M., Roman, J.L., Smoller, J.W., Zollei, L., Polimeni, J.R., Fischl, B., Liu, H., Buckner, R.L., 2011. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 106, pp. 1125-1165. 
Yu, S.X., Shi, J., 2003. Multiclass spectral clustering, In ICCV (pp. 313-319).

Acknowledgments

The data has been provided by the Human Connectome Project.

Contact

The content on this page is maintained by Salim Arslan. Let him know if you have any questions.