I want to do an analysis, comparing fixel-based, NODDI and DTI wholebrain statistic in a population of preterm 9-year old children. For this I use the msmt FBA recommended on the readthedocs.
However, I notice 2 things in my fod/fixel templates.
For starters, my Cerebellum is showing very bright in the fod template, is this normal?
Secondly, if I use the recommended pipeline I get a very wide fixeltemplate, covering the whole cerebellum, some GM and making some false positive conection over the edge of the ventricles. Also leading to almost 80000 fixels in the mask (reducing statistical power).
Plotting, the fmls_peak_values of the wmfod I think raising the threshold from 0.06 to 0.15. However this is not recommended in the pipeline. What would you suggest as a more proper way to restrict the mask? I would like to hold the benefit of the Dhollander msmt, and not getting my T1 images involved.
For starters, my Cerebellum is showing very bright in the fod template, is this normal?
This has been observed in various instances in individual subjects following B1 bias field correction (see eg. here; given it’s visible in your population template, I would suggest that the bias field correction may be going awry in a decent fraction of your subjects, and it’s worth going back for a look.
Also leading to almost 80000 fixels in the mask (reducing statistical power).
What would you suggest as a more proper way to restrict the mask?
Personally I’m a strong advocate of restricting statistical analysis to those fixels that are traversed by an adequate number of streamlines in the template-based whole-brain tractogram (more to come on this in the future).
Bear in mind that you can provide a fixel mask to fixelcfestats, and it’s the number of fixels in that mask that affect statistical power / RAM usage, not the total number of fixels in the template. So you can be more “liberal” with the template voxel masking / fixel segmentation, and then subsequently use multiple sources of information (eg. fixel streamline count, fixel FOD peak amplitude, fixel FD, voxel AFDtotal, …) to derive a fixel processing mask that you deem appropriate.
Is the spatial resolution of your template the same as suggested in the documentation? That strongly influences the number of fixels also.
Thanks for your advice and sorry for coming back at you so late.
I indeed already had to exclude one participant due to the overcorrection of N4 Biasfield correction.
Going back to this participant using the parameters suggested by Lucius in the thread you suggested did indeed solve the issue. I guess that mainly the initially high shrink factor is not doing good for me.
I guess this with the fact that children tend to move more than adults and the fairly low resolution of my data (2.5 x 2.5 x 2.5) is making my data more prone to these issues.
As a result I went even a bit more stringent than Lucius: shrink-factor to 1 (so off) and optimization to 500x500 which seems to result in quite a good correction.
I do resonate with the idea of restricting your CFE stats based on the number of streamlines passing through a fixel. I can calculate this using tck2fixel, right? Do you have any recommendations on how to decide on a threshold for this, or would you rely on visual inspection?
Regarding your concern of the spatial resolution, I have indeed followed the recommended msmt fixel pipeline.
I guess this with the fact that children tend to move more than adults and the fairly low resolution of my data (2.5 x 2.5 x 2.5) is making my data more prone to these issues.
I’m not sure that motion should affect bias field estimation too much; if anything it should make the field slightly more smooth. We’ve been acquiring the same spatial resolution here for a long time, so I wouldn’t think that would be a cause either. Maybe if you don’t have correction for B1 field on the scanner (eg. “Prescan normalise” on Siemens) that could make the bias field estimation harder. Beyond that though the N4 parameters we have are based on experience across a lot of data but not all data, so shouldn’t be considered divine word.
I do resonate with the idea of restricting your CFE stats based on the number of streamlines passing through a fixel. I can calculate this using tck2fixel, right? Do you have any recommendations on how to decide on a threshold for this, or would you rely on visual inspection?
Yep; eventually this will be part of tckmap, but for now it’s in tck2fixel. Thus far I’ve relied on manual inspection to set a threshold.
Regarding your concern of the spatial resolution, I have indeed followed the recommended msmt fixel pipeline.
It’s unusual for the number of fixels to vary so much from a standard case then. Though I should confirm: your initial quote was in fact 80k fixels, whereas I erroneously read it as 800k. For the standard pipeline (1.25 / 1.3mm), about 400k fixels would be considered normal. If you have not upsampled your data, then 80k fixels is probably about right.
Just to come back on some of the points you raised, we do have prescan on our machine and indeed I had 800k fixels (which was a typo on my account). The bright cerebellum might maybe have contributed to these large amount of fixels found. Anyway, I will rerun the whole analysis with the improved settings on the preprocessing.
