… because brain midline structures are not segmented well in freesurfer.
This statement is not only in reference to sub-cortical grey matter structures, but also many other brain structures on or near the midline within the cerebrum for which prior-driven tissue segmentation is difficult;
labelsgmfix only replaces the SGM structures with estimates from FIRST. I would suggest simply trying it and visually comparing the results with
5ttgen fsl. My personal judgement was that there is a lot of midbrain volume containing complex structure that is simply globally designated as white matter.
How does 5tt (not freesurfer) behave for ACT in conjunction with cortical ROIs defined using freesurfer’s ribbon-segmentation? I guess that singe the GM/WM interfaces of the cortical parcellation and the 5tt segmentation are not identical, there will be some problems there?
Firstly, can I change this to:
How does ACT behave in conjunction with cortical ROIs defined using freesurfer’s ribbon-segmentation?
I think this is the question you’re asking.
Answer: Not ideally. For two reasons (primarily the latter):
Potential resolution mismatch between a 5TT image generated from
5ttgen and FreeSurfer’s output image parcellations (if the native T1 used as input to
5ttgen fsl is not 1.0mm, this will be used as-is; but
recon_all immediately regrids to 1mm). MRtrix3 handles this seamlessly, but the precise behaviours may change slightly due to image interpolation during the re-gridding.
The 5TT format permits tissue partial volume estimates and may terminate a streamline as soon as the GM fraction exceeds the WM fraction. However FreeSurfer’s parcellation image outputs only label voxels that contain more GM than WM, and don’t include partial volume fractions at all. Therefore ACT may terminate a streamline due to “entering the cortex”, but the voxel in which that termination occurs is not labelled by FreeSurfer.
This is an effect that was known during the development of the connectome tools in MRtrix3 way back when, and was reported in more detail in this abstract (0118). It’s also why
tck2connectome by default uses a 2mm radial search to find the nearest parcel. However this mechanism only really “works” for connectome construction, not for Regions Of Interest.
If wanting to do tractography between cortical parcellations obtained from freesurfer, would it then not make more sense to use freesurfer’s tissue segmentation?
5ttgen freesurfer to derive the 5TT image would indeed alleviate this issue; streamlines would only be terminated if entering a voxel containing a GM label as designated by FreeSurfer, and thus using those same parcels as ROIs in
tckedit would behave as expected. You then however inherit the detriments of
5ttgen freesurfer, e.g. poor midbrain segmentation, “voxelised” / “jagged” cortical surface rather than smooth.
Applying a slight dilation to the ROIs in order to encapsulate the GM-WM interface is another alternative; but this may obviously capture some adjacent streamlines.
Longer term (i.e. still in development), the issue is being approached in a number of ways:
Using FreeSurfer’s surface-based segmentation to drive the 5TT image derivation, incorporating partial volume fraction estimation, e.g. as discussed in this thread.
Using “track extensions” to associate each streamline with a “column” of cortical grey matter, allowing ROIs (both from e.g. FreeSurfer, and hand-drawn) to be defined within the GM, and streamlines terminating at the GM-WM interface will be appropriately assigned to them; in
tck2connectome (i.e. superseding the radial search mechanism),
tckedit and others. Similar concept to this article.
Performing ACT using mesh-based representations of tissue surfaces, rather than a partial volume image; see this abstract (0058).
Enabling use of mesh representations for streamline assignment to parcellations / ROIs (e.g.
These are all probably still a fair way off though. We’re only human after all.