Pyafq ARC and SLF segmentation issue

[nwellman13]

Hi there,

We use qsirecon with the mrtrix_multishell_msmt_pyafq_tractometry option for processing our subjects (all kids under 18) and have noticed a fairly consistent (apparent) problem with pyafq’s segmentation, particularly for the ARC and SLF. The tracts often appear clipped prematurely. This seems to be occurring at the cleaning stage and seems to be at a point where multiple tracts intersect. It’s difficult for us to tell when this is accurate, due to an anatomical abnormality, or due to a failure in the segmentation. Any help would be appreciated!

Here's an example of the ARC before and after cleaning:

[arokem]

Hi @nwellman13 : thanks for reporting! It would be helpful to know what version of pyAFQ you are running, and if you are not using current main, it would be helpful to know whether you still see this when using the version that is currently on main.

[nwellman13]

We're using qsiprep version 0.18. I ran a single subject that we noticed this issue for yesterday with pyafq version 1.3.6 using the tck produced by qsiprep/recon 0.18 and this clipping was replicated.

clean bundles output from pyafq version 1.3.6:

Unclean bundles output from qsiprep version 0.18:

clean bundles from qsiprep v 0.18:

[arokem]

Any chance you can update to current main and try this out? To be honest, I am also not sure whether this should be considered an error. In some cases, users may need to adjust the cleaning parameters for their data, and this might be the case here.

[36000]

I agree, to further clarify, you have already done the majority of the work in identifying that the problem is with cleaning. I believe the uncleaned version looks better than the cleaned one, but the uncleaned one does have a few aberrant streamlines. You can tune distance_threshold, length_threshold, and n_rounds to get the best cleaned result. Increasing the thresholds and decreasing the number of rounds will reduce the aggressiveness of the cleaning. You may also want to try current pyAFQ main and DIPY tractography if you are still unsatisfied.

[36000]

I will add the on pyAFQ main we have also subdivided the SLF into 3 parts, which should help with this

[nwellman13]

Thank you both for your inputs! I tried a few different things. I updated to version 2.1, which in itself didn't change the output for the arcuate, but I then began tweaking some of the cleaning parameters which helped. Below was the best I could get. This was with clean rounds set to 2, distance_threshold = 6, length_threshold = 3.

Anything more strict began removing anterior streamlines:

[arokem]

Heads up that I made this into a discussion, because I believe this is not an issue that needs to be addressed in changes to the software, but is still worth keeping around for future reference.