Skip to the content.


Within this project we consider two main sources for surface parcellations, Existing and Random. We also considered two additional sources: Downsampled / Icosahedron and FreeSurfer Extracted.

Breakdown by type:

Existing Parcellations

In total, we assessed 82 existing parcellations. Parcellations available at multiple scales were assessed at every scale; in some cases where multiple versions of the same parcellation were available (e.g., from different re-sampling procedures or with different post processing applied), both versions were tested. 68 of the 82 parcellations were static or “hard” parcellations, in which each vertex is labelled as a part of exactly one parcel. We additionally considered 14 probabilistic or “soft” parcellations, where each parcel is represented by a set of probabilities or weightings across the whole surface or volume.

As we were interested in having the parcellations match the space of the data, all parcellations if not already is fs LR 32k space were re-sampled accordingly. For a detailed look at how resampling parcellations between different spaces see resampling.

The existing parcellations used are listed below:

Name # of Parcellations Type Reference
Schaefer Local-global 10 (scales 100-1000) Hard Schaefer 2018
Gordon 3 (different sources) Hard Gordon 2016
Brodmann Areas 1 Hard Brodmann 1909
VDG11b 1 Hard Van Essen 2012
HCP-MMP 3 (different sources) Hard Glasser 2016
Automatic Anatomical Labeling (AAL) 2 (different sources) Hard Tzourio-Mazoyer 2002
Baldassano 1 Hard Baldassano 2015
Desikan 2 (different sources) Hard Desikan 2006
Destrieux 2 (different sources) Hard Destrieux 2010
Brainnetome 2 (different sources) Hard Fan 2016
Power 2 (different sources) Hard Power 2011
Shen 268 Parcels 2 (different sources) Hard Shen 2013
Shen 368 Parcels 1 Hard Salehi 2020
Yeo 3 (7 Networks, 17 Networks and parcel level version) Hard Yeo 2014
DiFuMo 5 (scales 64-1028) Soft Dadi 2020
MIST 9 (scales 7-444) Hard Urchs 2019
AICHA 1 Hard Joliot 2015
Economo 1 Hard von Economo 2915
NSPN500 1 Hard Whitaker 2016
Oasis 1 Hard Sabuncu 2011
SJH 1 Hard Harrison 2015
Allen 1 Soft Allen 2011
BASC 9 (scales 9-444) Hard Bellec 2013
MSDL 1 Soft Varoquaux 2011
Harvard Oxford 4 (different versions) Hard / Soft Jenkinson 2012
Craddock 4 (different versions) Soft Craddock 2012
Smith ICA 2 (different versions) Soft Smith 2009
CPAC 1 Hard Craddock 2013
Hammersmith 1 Hard Hammers 2003
JuBrain 1 Hard Eickhoff 2005
MICCAI 1 Hard 2012 MICCAI Challenge_Data
Slab 2 (907 and 1068) Hard Sripada 2014
Princeton Visual 1 Hard Wang 2015

Click the parcellation name link to see that parcellation plotted

See also the folder raw/ which contains the ‘raw’ existing parcellations, before any preprocessing or re-sampling conducted by this project, also included are information on how they can be downloaded. See also the script setup/ which includes the specific code used to process the data from the raw/ folder into the Final Parcellations Used.

Random Parcellations

This project uses the idea of random surface parcellations extensively. We generated 5 random parcellations for of the following sizes in the base experiment:

10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1500, 2000, 3000, 4000, 5000, 6000

Click a parcellation size to see those random parcellations

Random parcellations are generated as follows: For a random parcellation of size N, N random points are first selected at random across both hemisphere’s 59,412 vertices (medial wall vertices excluded). Each selected point is then assigned as the seed of a new region and is randomly assigned a size probability between 0 and 1. Next, a region is randomly selected according to a weighted random choice between all regions (e.g., if a region was assigned an initial probability of .5 it would be picked on average twice as often as a region assigned .25). A random vertex is then added to the selected region from the list of valid neighboring unassigned vertices. This sequence, of selecting a region and adding one valid vertex, is repeated until all regions have no unassigned neighbors and therefore all non-medial wall vertices are assigned to a region.

Example of generating a random parcellation:

Random Parc Gif

Note: The above example, in contrast to the random parcellations generated in this project, is in fsaverage5 space (vs. fs_LR_32k) and doesn’t mask the medial wall (the medial wall is masked in this project).

Source code for generating random parcellations is implemented and available through the Brain Predictability toolbox (BPt), specifically here.

Random parcellations within this project are generated in the setup/ script.

Icosahedron Parcellations

We test 6 different downsampled icosahedron parcellations (with medial wall removed). These spanned sizes: 42, 162, 362, 642, 1002, 1442. See parcellations plotted.

FreeSurfer ROIs

The last parcellation we tested was actually using the FreeSurfer extracted ROIs directly. These differ from the other tested parcellations both in how values are generated (FreeSurfer extracts values in an individual’s native space whereas we extract values from data warped to a common space) in addition to the surface modalities used (only average thickness, surface area and mean curvature are employed, which differs from the features used in the base analyses). These were extracted in order to provide a comparison with with other common a ML neuroimaging analyses.

The code for generating the dataset of FreeSurfer ROIs is found here.


To access the final saved numpy array representations of the parcellations used, see Final Parcellations Used. These are saved as numpy arrays corresponding the fs_LR_32K space, in the case of probabilistic parcellation are saved with shape (64984, number of parcels), otherwise are saved as a flat array of 64984 vertex (left hemi first).

These parcellations are also all visualized here.