A Description of a Complete Dataset

The primary purpose of The Virtual Brain platform is to simulate whole brain dynamics. A simulation pipeline has different stages. The most fundamental stage is building a realization of a computational model which we call a large-scale brain network model. This model is constituted by a set of structural and functional components linked together, completely creating a particular model of the brain.

The following document is a generic description of what we often call a “minimal structural dataset for TVB” and “a complete dataset”. The aim of this document is to specify the different pieces of data required to derive the structural skeleton/substrate of a BNM. Hopefully, experts in the field of data acquisition will help us completing and improving our description.

A complete dataset should:

  • provide the history of the acquisition/processing protocols (traceability);

  • be in a standard format to perform analysis with different toolboxes;

  • have the Minimal Structural Dataset to derive a self-consistent set of files used in TheVirtualBrain. This minimal dataset will permit users to build large-scale brain network models and save their simulated data under different output modalities (eg, EEG, MEG, fMRI).

  • increase reproducibility of the results.

General Considerations

A dataset should be made available on one single place (eg, through XNAT) to ensure traceability and if required for clinical/privacy reasons, restricted access.

Note

the following definitions are not definitive, if you come across with a better categorization, please let us know.

Raw Structural Dataset

a collection of files describing a subject’s head (eg, structural MRI AND DTI Data)

Structural Dataset

(a collection of T1/T2 weighted MRI scans + DTI data + parcellation )

Minimal (Preprocessed) Structural Dataset

(surface mesh, parcellation, region centres, region mapping, connectivity matrix)

Complete Structural Dataset

in addition to the structures mentioned above, a complete structural dataset also contains head model surfaces, information about the units (areas, lengths, connectivity strengths), information about the EEG/MEG/iEEG sensors if users wish to compare simulated data to empirical data. Complete Dataset all of the above + functional data (eeg, rsfMRI, meg)

In general, all the steps in the processing pipeline should be documented, so it’s possible to apply the same treatment to subsequent datasets. Sending pieces of information to different files without descriptions from where they came and how they were processed is a bad practice and only detrimental for your own research project (takes a lot of time and it’s not reproducible). We can’t provide any meaningful help to integrate/check or validate incomplete datasets.

Ideally, any volumetric data (e.g., in NIFTI format), surface data (e.g., GIFTI format) or combination thereof (e.g., CIFTI format) should be provided in their RAW format, and if any pre-processing was performed on the raw data, associated data such as the region centres and parcellation mask should be provided in the same coordinate system as the cortical mesh (i.e., self-consistent dataset). The meaning of the (x,y,z) coordinates depends entirely on how the volumetric file was generated. It is possible to set any coordinate system you want (“native”, “mni”, “talaraich”) depending on the processing you apply to your data. A region centre, for example, would be a single spatial location in 3D. This location is specified by three numbers (x,y,z), these numbers should ideally represent mm and must be relative to an origin (x=0, y=0, z=0). The “same coordinate system” means that the origin is in the same location relative to the head, and that the axis(x,y,z) points in the same direction with the same orientation.

Parcellation Mask

What is the purpose of the mask?

A brain mask basically covers the standard brain. The mask needs to partition/parcellate the volume of space containing the head/brain into regions. It can be used to partition the cortical mesh into regions, consistently with the derived parcellated connectivity matrices. So, for each row in a connectivity matrix the mask needs to specify a region of space (region of interest). The mask can then be used to map the thousands of vertices which make up the surface mesh to the hundred or so regions in the connectivity matrix, that is, each region/row- in-a-connectivity-matrix is associated with hundreds of vertices in the cortical mesh. So the mask provides a way for us to generate a one to many mapping (region mapping) – from a row in the connectivity matrix to the many vertices of the surface mesh which lie in that region. This anatomical parcellation is also used to obtain finer parcellations and further divide the cortical surface into small regions of interest (REF to Zalesky) and distinguish subcortical structures.

What should be the format?

NIFTI is a standard format for volumetric time-series, and it is widely used in the neuroimaging community. Originally, NIFTI-1 file format was based on 16-bit signed integers and was limited to 32,767 in each dimension. The NIFTI-2 format is based on 64-bit integers and can handle very large volumes and matrices. The more recent CIFTI format is compatible with the NIFTI-2 format and it also has the extension .nii. We encourage to use the NIFTI-1 format (.nii and .nii.gz).I

Note

Add references to the libraries and software that are available for NIFTI-1 TVB also has a reader. Not the same case for NIFTI-2 and CIFTI. FieldTrip is the only one providing CIFTI i/o functionality.

In the case of a parcellation mask, each voxel contains an integer corresponding to a specific region (numeric labeling). This means, assuming voxels of 2x2x2 mm, the mask would consist of roughly 100x100x100 (ie, 1 million) voxels. The range of the integer number in each of these voxels should correspond to the number of regions (or rows ) in the parcellated connectivity matrix. Some voxels may contain no number or say -1, to specify that that piece of space doesn’t belong to a region, for example if it lies outside of the head. These type of conventions should be specified and documented. A list with the region names/labels and corresponding integer index should be provided.

How should region labels and names look like provided with the mask?

A region label is a short unique identifier, a region or area name usually refers to a human readable description. Examples for one region/name would be something like ‘label: RM-TCpol_R’ / ‘name: right temporal polar cortex’. Ideally, a reference to the original atlas/template should be provided as well. Notice that the correspondence between integer values in the parcellation mask and anatomical/human readable labels should be provided if they are not specified in the volume file.

Are region labels essential?

From view point of the implementation of The Virtual Brain the labels are essential.

Are region names essential?

The region names on the other hand are primarily a matter of usability, though a very valuable one, when you want to identify an area that you wish to modify in a simulation (e.g., modeling lesions). Unless a user is an anatomist and acquainted with the labels, then the names are much clearer.

Why is information on cortical vs. subcortical regions needed?

We need a means of distinguishing cortical from subcortical regions within the mask, so that when we apply the mask to a cortical mesh we don’t inadvertently associate parts of the cortex with subcortical regions in the connectivity matrix. Ultimately a vector of the length of the number of regions is needed, specifying whether each region is part of the cortex or not. If the labels or names clearly include this information, that is they clearly state whether they are cortical regions or not, then the vector could be generated on this basis.

Is the parcellation mask unique?

No. Currently, there are several parcellation masks being used in the community. NOTE: REF parcellation papers. One of the main problems is that parcellation masks are often custom made and subsequently modified, so it becomes very difficult to track the origins. To begin with, we suggest to use parcellation masks provided by neuroimaging software tools like FSL AAL 90. If you want to use a custom made parcellation mask, then it should have the characteristics mentioned above. Also, having the structural raw data it is possible to derive connectivity matrices from the same dataset, but at different resolutions. NOTE: (reference to Hagmann and Zalesky).

What is the coordinate system of the parcellation mask?

It depends on how the parcellation mask was obtained. In principle, it should be registered to a standard space such as MNI. These coordinate systems should be consistent with the surface’s coordinate systems.

Connectivity and path length data

What is required for building a connectivity matrix (parcellated connectome)?

Diffusion data, a parcellation mask and probably the white matter surface (in the same space, aligned). In TVB, we are not providing the tractography tools to create structural connectivity matrices.

Are the tract lengths essential for using TVB?

Yes. The simulations in TVB take into account time delays, and their magnitude is given by the distance between pairs of regions scaled by the conduction speed.

Are the region centres important?

Yes! If for a reason unbeknown to you, you happen to not have the white matter fibre lengths, then TVB uses the region centres to compute a tract lengths matrix based on the Euclidean distance between region pairs. The region centres are merely a list of Cartesian triplets (x,y,z) that specify the spatial location relative to the consistent coordinate system mentioned above. Each region centre is just a single point in space, corresponding to the centre of the region. The region itself might be spatially extended (if we have the cortical surface), and thus not a single point.

What is the parcellated connectome?

This term was introduced by the HCP, and it refers to the connectivity matrix. For TVB a Connectivity refers to a set of two matrices (of size “anatomical regions x anatomical regions “), one with weights giving the strength of the connections between anatomical regions and a second matrix with the white matter fibre lengths between regions.

Cortical Mesh

We encourage to use the MNI brain template (eg, MNI152) to register your subjects data and extract the corresponding cortical surface.

Is the cortical surface essential?

Yes! Strictly speaking, TVB can perform simulations using only a parcellated connectome as spatial support. From a scientific point of view MODELING THE ELECTRICAL ACTIVITY ON THE FOLDED CORTICAL SURFACE is one of the most interesting capabilities to exploit in TVB. Modeling work where different output modalities (like EEG and BOLD) are compared need a certain level of geometrical detail that is not provided by a coarse-grained connectome. While in the field of macroconnectomics, the parcellated connectome is sufficient (debatable subject, see the paper by Zalesky), the cortical surface is necessary to work with neural field modeling and to account for spatial inhomogeneities.

The cortical surface represents the outer surface of the gray matter. It’s often called ‘pial surface’.

How is a surface represented?

A way of representing 2D meshes embedded in 3D space is by storing two arrays, one for vertices, and one for triangles. Tha latter is an array with triplets of indices into the first array of vertices. So, basically a surface mesh is given by a set of vertices (triplets (x,y,z) defining the location of those vertices). And alternatively, the mesh can be represented by triangle arrays which are indices into the vertex arrays; three indices for each triangle.

Then there are other ‘attributes’ that can be derived from these two main arrays, for instance ‘normals’. A normal determines the orientation of a vertex.

All vertex-related/derived information is calculated and stored in separate arrays, although bound to the surface instance they were derived from. Read more about normals here: http://user.xmission.com/~nate/smooth.html

Region Mapping

What is the Region Mapping?

The region mapping is just a relationship between the two pieces of data, mapping regions of a connectivity onto the nodes of a surface simulation. We are talking about a one to many relationship for the vertices of the cortical surface and one to one relationship for the remaining non-cortical regions. NOTE: A region mapping could be between two connectomes of different resolution (eg, the connectomes presented in Hagmann 998 to 66 regions).

How is the Region Mapping obtained?

It is obtained by combining the lh.aparc.annot and rh.aparc.annot files that FreeSurfer generates for a particular subject. The details of the resulted Region Mapping depend on the preprocessing pipeline used, some of them cut out the regions that do not have any vertices assigned, others keep them etc.

Head model

What is the purpose of the head model

Head: the bucket that contains the brain. The head is often represented as a set of concentric spheres, in order to compute the electric field or potential on the skin surface (eg, as recorded with EEG electrodes). The concentric spheres (surfaces) represent the boundaries between the brain and the skull; the skull and the skin; and, the skin and the air mesh.

What should be the format?

A surface format like GIFTI, or in the same format used for the cortical mesh.

Is the head model essential?

From a scientific point of view, it is essential to compute the lead-field matrices which will project the neural activity time-series into sensor space (eg EEG). The boundary surfaces are then required to assist Open MEEG (or any other similar tool like FieldTrip) to generate good forward models for EEG/MEG)

The surfaces describing a subject’s head: skin, skull, cortical surface. See the description below.

A Minimal Structural Dataset For TVB:

All 3D coordinates should be consistent, i.e., vertices, parcellation masks, and region centres should be in the same units, axis orientations, alignment, etc.

A minimally-complete connectivity data set for TVB

should include the following:

  • Mesh surface for the cortex (regularised, continuous and complete per hemisphere, that is, there should be no holes in the surface and it should be possible to unambiguously define an inside and an outside, in other words, each hemisphere should be topologically spherical):

    • vertices (Cartesian (x,y,z))

    • triangles (triplets of indices into the vertices array, TRIANGLES, but not

      generalised polygons)

  • Parcellation:
    • Spatial mask, 3D, PROPERLY ALIGNED WITH THE SURFACE, i.e. coordinates, orientation should be IN THE SAME SPACE.

    • Labels for all regions composing the parcellation/connectivity data.

    • A clear delineation, if not explicit in the labels, between cortical regions and subcortical structures.

  • Region centres (Cartesian (x,y,z), consistent with surface, mask, etc.), for all regions composing the parcellation/connectivity data.

  • Connectivity (DSI):
    • Connection strength/s between regions.

    • Tract length between regions.

Ideally

For a complete structural dataset, we should also have:

  • Connectivity: mainly Connection strength between regions.
    • This should include information specifying the directionality. That is, if the data is provided as a matrix rather than a file format including meta-data such as graphml, directionality should be clearly and unambiguously specified.

  • Mesh surfaces for:
    • inner-skull: boundary between the brain and the skull,

    • outer-skull: the boundary of between the skull and the skin

    • outer-skin: boundary surface between the skin and the air (for EEG/MEG monitors)

  • Basic additional information:
    • Units: tract lengths, coordinates etc (mm).

    • Units: strength/weights units, (au) if none.

    • additional relevant information…

Guidelines to import the data into TVB

Currently we have some guidelines describing what data fields and in which format users can import different components of a complete dataset (connectome, surface, sensors, gain matrix for eeg, etc…).

Note

Check the DataExchange chapter of the User Guide manual.

The TVB demonstration dataset

Note

DISCLAIMER: This dataset was custom made and built to serve the purpose of numerically testing the simulator, as well as for theoretical exploration. It does have, however, certain issues with regard to biophysical realism and so it shouldn’t be used/relied-upon for that purpose. References, where appropriate, are given. Also, this is an open source project and contributions are greatly appreciated. If you see an error, please leave a comment or make corresponding modifications (please give proper references and argument your corrections).

  • The parcellation was chosen to be as homologous as possible between Macaque and Human. (See the [scalable brain atlas interactive tool] (http://scalablebrainatlas.incf.org/main/coronal3d.php?template=PHT00&plugin=CoCoMac))

  • Weights are primarily CoCoMac – exceptions are colossal connections. These are DSI fibre bundle widths scaled to fill the 0-3 of CoCoMac.

  • Most colossal connection are missing. Tract-lengths are actual DSI tracts where possible and Euclidean distance used where explicit DSI/DTI tract- lengths weren’t available.

  • Region centres were generated to be consistent with the demo cortical surface.

  • In the current parcellated connectome all the non-cortical regions were stripped.

  • The CoCoMac connectivity belongs to a single hemisphere, so the weights matrix is symmetric (weighted undirected graph), but the DSI was “whole” brain and so there is probably hemispheric asymmetry in tract lengths and the cortical surface is hemispherically asymmetric so region centres aren’t the same for both hemispheres. (this item is maybe deprecated…)

The default TVB connectivity is a bi-hemispheric hybrid CoCoMac/DSI matrix. Subcortical regions (e.g. thalamus and other subcortical nuclei) are not included in this matrix.

Anatomical labels and names:
  • A1: Primary auditory cortex

  • A2: Secondary auditory cortex

  • Amyg: Amygdala

  • CCa: Gyrus cinguli anterior

  • CCp: Gyrus cinguli posterior

  • CCr: Gyrus cinguli retrosplenialis

  • CCs: Gyrus cinguli subgenualis

  • FEF: Frontal eye field

  • G: Gustatory cortex

  • HC: Hippocampal cortex

  • IA: Anterior insula

  • IP: Posterior insula

  • M1: Primary motor area

  • PCi: Inferior parietal cortex

  • PCip: Cortex of the intraparietal sulcus

  • PCm: Medial parietal cortex (Precuneus)

  • PCs: Superior parietal cortex

  • PFCcl: Centrolateral prefrontal cortex

  • PFCdl: Dorsolateral prefrontal cortex

  • PFCdm: Dorsomedial prefrontal cortex

  • PFCm: Medial prefrontal cortex

  • PFCorb: Orbital prefrontal cortex

  • PFCpol: Pole of prefrontal cortex

And more:
  • PFCvl: Ventrolateral prefrontal cortex

  • PHC: Parahippocampal cortex

  • PMCdl: Dorsolateral premotor cortex

  • PMCm: Medial premotor cortex (supplementary motor cortex)

  • PMCvl: Ventrolateral premotor cortex

  • S1: Primary somatosensory cortex

  • S2: Secondary somatosensory cortex

  • TCc: Central temporal cortex

  • TCi: Inferior temporal cortex

  • TCpol: Pole of temporal cortex

  • TCs: superior temporal cortex

  • TCv: ventral temporal cortex

  • V1: Primary visual cortex

  • V2: Secondary visual cortex

We have:
  • An importer for RegionMapping (externally computed);

We need:
  • At least one, preferably multiple, complete dataset to serve as a default dataset available to users who can’t or aren’t interested in providing their own. Of specific importance here is the Connectivity Parcellation Mask, as well as a specification of hemisphere and cortical vs non-cortical regions. If you are interested in contributing to a dataset, please contact the tvb google group.

  • Algorithm for calculating the region mapping, given a coregistered Cortex and Parcellation Mask, including an “island” removal/correction mechanism to deal with the imperfect alignment that will exist, even with coregistered data, between an individual’s cortical surface and the “generic” parcellation mask.

    Note

    Demo data as described in this chapter, can be found on Zenodo: https://zenodo.org/record/10128131, or inside TVB_Distribution, under the following path: TVB_Distribution/tvb_data/Lib/site-packages/tvb_data/ on Windows, TVB_Distribution/tvb_data/lib/python3.x/site-packages/tvb_data/ on Linux, or TVB_Distribution/tvb.app/Contents/Resources/lib/python3.x/tvb_data/ on Mac. These demo files can be used together with the GUI and/or the script interfaces, or taken as reference for you, when creating TVB compatible dataset.

Other datasets

Hagmann

What has been provided/shown :

  • A 998 ROIs connectome (weights + resampled distances)

  • A mapping to the parcellated connectome of 66 regions

  • Label and anatomical names

  • Info about the coordinate system: Talaraich

What’s missing:

  • The parcellation mask file

  • The cortical surface

  • The head model

Permissions:

  • On request to the authors

The Human Connectome Project

So far, it contains the most complete datasets available. We aim to integrate some of the datasets provided by the HCP. Structural connectivity is the fundamental substrate for building large-scale brain network models, and being able to use these high quality, standardized and equally pre-processed data would be ideal.

However, “advanced” HCP datasets will be hopefully released next year. The HCP data release does not include extensively processed connectivity data for individual subjects, but mainly “an average dataset”. In the current release, Q3, there are dense (“grayordinate-to-grayordinate”) functional connectivity datasets based on resting state fMRI from individual subjects. However, HCP people are still working on improving many of the steps for generating structural connectivity datasets, based on diffusion imaging and probabilistic tractography. In the future, they will release probabilistic tractography and “dense structural connectome” datasets (perhaps with the Q4 release, Q3 release was made available on September 20th, 2013).

There are ongoing efforts both within and outside the HCP consortium to generate improved methods of brain parcellation, especially cerebral cortex. “HCP- sanctioned” parcellated connectome datasets (based on improved cortical parcellations) will be made publicly available in the future (no target date announced yet). Once these (plus the dense connectome datasets) are released, users will be able to generate parcellated connectomes based on their own preferred parcellation scheme.

They do plan to make a (FieldTrip-compatible) head model available for each subject scanned using MEG.

What they have:

  • Almost everything: raw, minimally processed and processed data.

What’s missing:

  • Preprocessed diffusion data (e.g., fiber orientation, fiber tracts) and derived structural connectomes and individual based parcellations.

Permissions:

  • available after agreeing with the privacy and sharing conditions. In principle, datasets can be distributed as long as we make users sign the terms required by the HCP. I would suggest, once the dense and some parcellated connectomes are available, to buy the connectome in a box and have a copy in a centralized storage server so TVB can read these data in.

Brain-mapping softwares:
MRI Processing/Analysis/Modeling platforms:
Data exchange/db platforms:

Glossary

Space Coordinate systems:
Atlases:
Structural Anatomical Parcellations:
  • Kotter (macaque)

  • Broadmann

  • FSL AAL 90

  • Hagmann (based on Desikan)