Could you please explain a bit on -sgm_amyg_hipp option of 5ttgen and labelsgmfix. In which scenario one should use this option?
I am a bit confused. If I use 5ttgen fsl -sgm_amyg_hipp and down the road use freesurfer to generate the grey matter parcellation, do I have to also set this option in labelsgmfix?
-sgm_amyg_hipp option simply includes the amygdalae and hippocampi in the lists of sub-cortical grey matter structures to be processed. So the effect depends on the script being used:
5ttgen fsl, the segmentations of these structures by FSL FIRST is used to over-write the underlying GM/WM segmentation from FSL FAST.
5ttgen freesurfer, the mapping from FreeSurfer segmentation to tissue index is simply altered such that these two structures are mapped to “sub-cortical” grey matter rather than “cortical” grey matter (really, they’d be better described as “GM that streamlines terminate as soon as they touch” v.s. “GM that streamlines can project into”).
labelsgmfix, the segmentations of these structures from FreeSurfer are removed from the parcellation image, and replaced by the estimates from FIRST.
The reason these structures were initially omitted from the “sub-cortical grey matter” list for ACT is that FAST does a pretty reasonable job of delineating these structures, so streamlines basically terminate at the outer edge of the structure. Conversely, if you do explicitly segment these structures with FIRST, quite often there remains a “strip” of GM tissue from FAST that lies outside the FIRST segmentation; in this instance, streamlines will still terminate as soon as they hit that outer strip, and will not “project” into the voxels labelled as sub-cortical grey matter as one might expect. But you’re more than welcome to experiment with the option and see the differences for yourself.
Technical aside regarding 5TT generation, the fsl_anat script, and ACT in general
As an aside, I tried using FSL’s
fsl_anat script to derive a 5TT image, as I think has been brought up on a couple of occasions around here. In this script the B1 inhomogeneity correction is so strong that FAST ends up doing a much better job of labelling all sub-cortical grey matter structures as GM tissue. Hence, inserting the FIRST estimates basically doesn’t work for any sub-cortical grey matter structures, because each still has this outer “skin” of voxels labelled as “cortical” GM, so streamlines don’t project into them as intended.
Fundamentally, “overriding” an image-based segmentation with a mesh-based segmentation, as is done in
5ttgen fsl, only works if the mesh-based segmentation fully encapsulates the image-based segmentation. In my experience when first developing ACT, this was the case for the five sub-cortical structures segmented by FIRST in
5ttgen fsl, but not really for the hippocampi and amygdalae. Hence why they were initially omitted from the list; the
-sgm_amyg_hipp command-line option was later added simply to remove a sort of “artificial imposed constraint” on 5TT generation (even though people are very much invited to experiment with
5ttgen, this was an easy variable to provide to non-technical users).
A better solution would be to use only mesh-based segmentation to define tissue boundaries, rather than mixing image-based and mesh-based. This could be either an alternative
5ttgen script algorithm that only uses mesh-based anatomical segmentations to derive the 5TT image; or a more advanced solution is to natively use a five-tissue-type mesh database rather than a 5TT image for tracking, as shown by @chunhungyeh at ISMRM 2017 (abstract 58).
Using the option at one step of processing does not make its use at another step compulsory; though it would probably be recommended. It’s affecting two different aspects of processing in different ways.
Thank you so much for the explanation.