Structural connectome construction

Included in this new version of MRtrix are some useful tools for generating structural connectomes based on streamlines tractography. Here I will describe the steps taken to produce a connectome, and some issues that should be taken into consideration. Note that I will not be going into appropriate parcellations or network measures or anything like that; once you’ve generated your connectomes, you’re on your own.

Preparing a parcellation image for connectome generation

Parcellations are typically provided as an integer image, where each integer corresponds to a particular node, and voxels where there is no parcellation node have a value of 0. However, for all of the parcellation schemes I’ve looked at thus far, the values used for the nodes do not increase monotonically from 1, but rather have some non-linear distribution; a text file (or ‘lookup table’) is then provided that links node indices to structure names. This is however undesirable for connectome construction; it would be preferable for the node indices to increase monotonically from 1, so that each integer value corresponds to a row/column position in the connectome matrix.

This functionality is provided in the command labelconvert. It takes as its input a parcellation image that has been provided by some other software package, and converts the label indices; this is done so that the code that actually generates the connectome can be ‘dumb and blind’, i.e. the integer values at the streamline endpoints correspond to the row & column of the connectome matrix that should be incremented. In addition, this processing chain design provides flexibility in terms of both the source of the parcellation data, and the way in which the user wishes to customise the layout of their connectome.

Please consult the tutorial labelconvert: Explanation & demonstration for a guide on how to use the labelconvert command.

Generating the connectome

The command tck2connectome is responsible for converting the tractogram into a connectome matrix, based on the provided parcellation image. By default, the streamline count is used as the connectivity metric; run tck2connectome -help to see alternative heuristics / measures.

A factor in structural connectome production commonly overlooked or not reported in the literature is the mechanism used to assign streamlines to grey matter parcels. If done incorrectly this can have a large influence on the resulting connectomes. This is one aspect where Anatomically-Constrained Tractography (ACT) really shines; because streamlines can only terminate precisely at the grey matter - white matter interface, within sub-cortical grey matter, or at the inferior edge of the image, this assignment becomes relatively trivial. The default assignment mechanism is a radial search outwards from the streamline termination point, out to a maximum radius of 2mm; and the streamline endpoint is only assigned to the first non-zero node index. If you do not have the image data necessary to use the ACT framework, see the ‘No DWI distortion correction available’ section below.

SIFT and the structural connectome

If you are generating structural connectomes, you should be using Spherical-deconvolution Informed Filtering of Tractograms (SIFT).

Extracting pathways of interest from a connectome

The command connectome2tck can be used to extract specific connections of interest from a connectome for further interrogation or visualisation. Note that since the resulting connectome matrix does not encode precisely which parcellation node pair each streamline was assigned to, the streamlines are re-assigned to parcellation nodes as part of this command. Run connectome2tck -help to see the various ways in which streamlines may be selected from the connectome.

Also: Beware of running this command on systems with distributed network file storage. This particular command uses an un-buffered file output when writing the streamlines files, which re-opens the output file and writes data for individual streamlines at a time (necessary as many files may be generated at once); such systems tend to be optimised for large-throughput writes, so this command may cause performance issues.

No DWI distortion correction available

If you can’t perform DWI susceptibility distortion correction, it severely limits how accurately you can estimate the structural connectome. If this is the case for you, below is a few points that are worth considering.

Non-linear registration

Rather than actually correcting the DWI geometric distortions, some people try to do a non-linear registration between DWI and T1 images. In general I’m against this: the registration is fairly ill-posed due to the differing contrasts, and an off-the-shelf non-linear registration will have too many degrees of freedom. Pursue at your own risk.

Grey matter parcellation

With good spatial alignment, parcellations that highlight only the cortial ribbon (e.g. FreeSurfer) are highly accurate and effective, and the assignment of streamlines to those parcellations will also be robust if ACT is used. But without these, residual registration errors may have a large influence, and assigning streamlines to parcellations only as thick as the cortex may also be erroneous (streamlines may terminate prior to the parcel, or travel through and extend well beyond it). A parcellation with large-volume nodes that is based on atlas registration (e.g. AAL) is likely more appropriate in this case.

Assignment of streamlines to parcellation nodes

Without ACT, streamlines will terminate pretty much anywhere within the DWI brain mask. Not only this, but they may traverse multiple parcellation nodes, turn around within a node and traverse elsewhere, terminate just prior to entering a node, all sorts of weirdness. I have provided a few assignment mechanisms that you can experiment with - run tck2connectome -help to see the list and parameters for each. Alternatively if anyone has a better idea for how this could potentially be done, I’d love to hear it.