AFQ - streamlines and ROI filtering

[emiliazms19]

Hi,

I built a pipeline importing whole-brain tractograms from MRtrix (downsampled to ~200k streamlines to manage RAM constraints) along the preprocessed dmri using "PARICIPANTAFQ" to extract dki_fa.
The pipeline runs and exports csv profiles, but I'd like to get opinions in three specific points:

• During profile calculation, I frequently encounter "WARNING:AFQ:Not enough streamlines for weight calculation, weighting everything evenly", how much noise does unweighted averaging introduce into the resulting profile compared to core weighting?.
• ⁠Across subjects, the number of recognized bundles varies between 13 and 19 (some pathways like CST or Corpus Callosum segments drop streamlines during ROI filtering). Given that these are parkinson's brains and ~200k tractograms, what segmentation parameters and how would you suggest tuning to improve bundle recovery?.
• ⁠what is the recommended way to call the native 3D visualization utilities directly using the outputs from a ParticipantAFQ run?.

Thank you!!!

[36000]

Hello,
Points 1 and 2 are a consequence of the low number of streamlines. I think it may be better to find a way to increase the number of streamlines, if possible.

By default, as of 3.0, of pyAFQ will generate tractograms with 1e7 streamlines, but it will generate them in batches, reducing RAM requirements. Have you tried this? It will also run bundle recognition in batches as of version 3.1, given the input is a .trx file. If you would prefer to use MRtrix tractography, you could also try writing a script to convert the .tck file into a .trx file in batches, then pass that into pyAFQ, allowing pyAFQ to perform bundle recognition in batches.

For point 3, pyAFQ does generate an html file where plotly is used to visualize the recognized bundles. This can be useful for quick QC checks, but for paper quality figures, see our fury example: https://github.com/tractometry/pyAFQ/blob/main/docs/source/examples/tutorial_examples/plot_005_viz.md

Best,
John