labelconvert: Explanation & demonstration

The labelconvert (previously labelconfig) step in Structural connectome construction has proven to be a hurdle for many. It may be a ‘unique’ step in so far as that other software packages probably deal with this step implicitly, but in MRtrix we prefer things to be explicit and modular. So here I’ll go through an example to demonstrate exactly what this command does.

Worked example

For this example, let’s imagine that we’re going to generate a structural connectome for Bert, the quintessential FreeSurfer subject. Also, we’re going to generate the connectome based on the Desikan-Killiany atlas. The default FreeSurfer pipeline provides the volumetric image aparc+aseg.mgz; this is the file that will be used to define the nodes of our connectome.

labelconvert\_before

Looking at the raw image itself, each node possesses a particular intensity, corresponding to a particular integer value. If we focus on the superior frontal gyrus in the right hemisphere, we can see that the image intensity is 2028 for this structure.

This immediately presents a problem for constructing a connectome: if any streamline encountering this region were written to row/column 2028, our connectome would be enormous, and consist mostly of zeroes (as most indices between 1 and 2028 do not correspond to any structure). Therefore, what we’d prefer is to map the unique integer index of this structure to a particular row/column index of the connectome; this should be done in such a way that all structures of interest have a unique integer value between 1 and N, where N is the number of nodes in the connectome.

Now looking at the file FreeSurferColorLUT.txt provided with FreeSurfer, we see the following:

...
2026    ctx-rh-rostralanteriorcingulate     80  20  140 0
2027    ctx-rh-rostralmiddlefrontal         75  50  125 0
2028    ctx-rh-superiorfrontal              20  220 160 0
2029    ctx-rh-superiorparietal             20  180 140 0
2030    ctx-rh-superiortemporal             140 220 220 0
...

This gives us a meaningful name for this structure based on the integer index. It also gives us some colour information, but let’s not worry about that for now.

Our goal then is to determine a new integer index for this structure, that will determine the row/column of our connectome matrix that this structure corresponds to. This is dealt with by mapping the structure indices of this lookup table to a new lookup table. For this example, let’s imagine that we’re using the default MRtrix lookup table for the FreeSurfer Desikan-Killiany atlas segmentation: this is provided at shared/mrtrix3/labelconvert/fs_default.txt.Examining this file in detail, we see the following:

...
74    R.RACG  ctx-rh-rostralanteriorcingulate 80  20  140 255
75    R.RMFG  ctx-rh-rostralmiddlefrontal     75  50  125 255
76    R.SFG   ctx-rh-superiorfrontal          20  220 160 255
77    R.SPG   ctx-rh-superiorparietal         20  180 140 255
78    R.STG   ctx-rh-superiortemporal         140 220 220 255
...

(This file is in a slightly different format to FreeSurferColorLUT.txt; don’t worry about this for the time being)

This file contains the same structure name as the FreeSurfer look-up table, but it is assigned a different integer index (76)! What’s going on?

The following is what the labelconvert command is actually going to do under the bonnet, using these two lookup table files:

  1. Read the integer value at each voxel of the input image
  2. Convert the integer value into a string, based on the input lookup table file (FreeSurferColorLUT.txt)
  3. Find this string in the output lookup table file (fs_default.txt)
  4. Write the integer index stored in the output lookup table file for this structure to the voxel in the output image

This is what the actual command call looks like:

labelconvert $FREESURFER_HOME/subjects/bert/mri/aparc+aseg.mgz $FREESURFER_HOME/FreeSurferColorLUT.txt ~/mrtrix3/src/connectome/config/fs_default.txt bert_parcels.mif

And this is what the resulting image looks like:

labelconvert\_after

The integer labels of the underlying grey matter parcels have been converted from the input lookup table to the output lookup table (hence the name labelconvert). They now increase monotonically from 1 to the maximum index, with no ‘gaps’ (i.e. ununsed integer values) in between. Therefore, when you construct your connectome using tck2connectome, the connectome matrix will only be as big as it needs to be to store all of the node-node connectivity information.

Design rationale

Making this step of re-indexing parcels explicit in connectome construction has a few distinct advantages:

  • You can use parcellations from any software / atlas: just provide the structure index / name lookup table that comes with whatever software / atlas provides the parcellation, and define an appropriate target lookup table that defines which index you want each structure to map to.
  • tck2connectome can be ‘dumb and blind’: it reads the integer indices at either end of the streamline, and that’s the row/column of the connectome matrix that needs to be incremented.
  • You can have your grey matter parcels appear in any order in your matrices: just define a new lookup table file. Doing this prior to connectome construction is less likely to lead to heartache than re-ordering the rows and columns in e.g. Matlab, where you may lose track of which matrices have been re-ordered and which have not.
  • You can remove structures from the connectome, or merge multiple structures into a single parcel, just by omitting or duplicating indices appropriately in the target lookup table file.
  • Looking at your matrices and need to find out what structure corresponds to a particular row/column? Just look at the config file!

Obviously if your parcellation image already has node indices that increase monotonically from 1, and you’re happy enough with the numerical order of the nodes, you don’t actually need to use the labelconvert step at all.

Custom design connectomes

Some notes for anybody that wishes to define their own configuration files (either for re-ordering nodes, changing selection of nodes, or using parcellations from alternative sources):

  • If you wish to omit nodes from your connectome (e.g. the cerebellar hemispheres), you may be better off making these nodes the largest indices in your connectome, but then cropping them from the connectome matrices retrospectively, rather than omitting them from the parcellation image entirely: If you were to do the latter, streamlines that would otherwise be assigned to your unwanted nodes may instead be erroneously assigned to the nearest node that is part of your connectome (exactly what happens here will depend on the streamline-node assignment mechanism used).
  • The command labelconvert is capable of reading in look-up tables in a number of formats. If you wish to define your own lookup table, you will need to conform to one of these formats in order for MRtrix commands to be able to import it. If you are using an atlas where the look-up table does not conform to any of these formats (and hence MRtrix refuses to import it), you can either manually manipulate it into a recognized format, or if it is likely that multiple users will be using that parcellation scheme, we may choose to add a parser to the MRtrix code: contact the developers directly if this is the case.