SIFT preparation in lesioned brains

Dear MRtrix experts,

I am currently analyzing my DWI data of a group of brain tumor patients, on which I would like to do whole-brain tractography to generate structural connectomes. However, I’m not entirely sure about the optimal way to prepare my data for this purpose.

In particular, I have 2 separate datasets. The first one (acquired by myself) is a multi-shell DWI protocol, including PA scan for optimal distortion corrections. Using this sequence, I have data for 11 control subjects, 13 meningioma brain tumors (pushing the brain aside, but brain remains intact) and 10 glioma brain tumors (infiltrating -“eating away”- the brain). Then, I have another dataset of 5 glioma patients, with single-shell acquisition without PA scan.

In my preprocessing (using the latest MRtrix version 3.15) I included the following:

  • dwidenoise
  • distortion correction using eddy & topup + bias field correction using fast
  • coregistration of T1 to DWI space
  • 5tt segmentation of coregistered T1
  • intensity normalization DWI images (using dwinormalise)
  • estimating msmt response function per subject

My first question is whether or not to average my response function across subjects. If I should, should I then average across both studies (multi + single shell) or within study? I guess averaging across studies wouldn’t make much sense, but on the other hand if I average per study, I assume my response function will be biased because in one study I include controls + patients, whereas in the other study I only have 5 patients.

Secondly, in the documentation it says that before you average your response function, you should do inter-subject intensity normalisation. Is this step accomplished using the dwinormalise command?

In case averaging wouldn’t be appropriate, I found on the documentation that the simplest and most common solution is to use an identical number of streamlines for every subject in connectome construction. This is also what I tried in the first place: generating 10M tracts and “sift”-ing to 1M. However, some of the glioma patients have very large lesions - some with almost an entire lobe affected/resected. For these patients, you would suspect a priori to find less tracts. So I’m looking for a better termination criterion for tract generation and sifting as well.

All suggestions and advice is welcome!
Thanks!

Best,
Hannelore

Hi Hannelore,

Apologies for the delay; was dreading trying to address this question, then went on leave and forgot about it :disappointed:

Then, I have another dataset of 5 glioma patients, with single-shell acquisition without PA scan.

I would be slightly concerned about inclusion of the 5 glioma patients due to the absence of PA scan over and above the issue of intensity normalisation / response functions. Your FOD template image will be principally based on EPI-distortion-corrected images, and hence non-linear registration of those subjects to the template will include both subject variance and EPI distortions. We’ve never tried this to my knowledge, so I don’t know just how well / poorly it might perform.

My first question is whether or not to average my response function across subjects. If I should, should I then average across both studies (multi + single shell) or within study? I guess averaging across studies wouldn’t make much sense, but on the other hand if I average per study, I assume my response function will be biased because in one study I include controls + patients, whereas in the other study I only have 5 patients.

You likely still want to average response functions across subjects. If you perform intensity normalisation based on the b=0 images, but a cohort has a difference in macroscopic fibre density or T2 in those regions used to derive the response function, failing to use an average response function will mask this effect. Accepting this limitation and going with subject-specific response functions would actually simplify the issue in a way since you would no longer need to worry about acquisition differences, but you would need to accept and divulge these limitations.

Note however that averaging response functions is predicated on the b-values being equivalent. Different b-values result in different response function shapes over and above estimation variance, so averaging is not appropriate. Even then, if your single-shell protocol corresponded to one of the shells in your multi-shell protocol, I don’t think I’d be averaging that one shell across all acquisitions and averaging the “unique” shells only within the multi-shell acquisition…

If you can get control subjects with the single-shell acquisition, then you could average response functions within protocols, and include protocol as a nuisance regressor. However I’m pretty sure this won’t work if you have only subjects from one of the two groups in that protocol: the model fit would simply set the intercept for that protocol such that the mean value across those subjects lies wherever the group-wise regression predicts that the subject group mean will be, and the extra subjects from that protocol would therefore not be contributing any information to your hypothesis.

Secondly, in the documentation it says that before you average your response function, you should do inter-subject intensity normalisation. Is this step accomplished using the dwinormalise command?

Yes, this is the step provided by dwinormalise. More recently we also have the mtbin command, which may have provided an alternative compromise if you had differing multi-shell acquisitions (since it additionally uses the total tissue volume within a voxel as a kind of ‘soft’ normalisation), but can’t currently be used to combine single-shell and multi-shell acquisitions.

In case averaging wouldn’t be appropriate, I found on the documentation that the simplest and most common solution is to use an identical number of streamlines for every subject in connectome construction. This is also what I tried in the first place: generating 10M tracts and “sift”-ing to 1M. However, some of the glioma patients have very large lesions - some with almost an entire lobe affected/resected. For these patients, you would suspect a priori to find less tracts. So I’m looking for a better termination criterion for tract generation and sifting as well.

If you’re performing connectome analysis rather than FBA, then yes: You need to normalise your quantification of “connection density” over and above AFD-style intensity normalisation. See this response…

Hope I didn’t miss anything!
Rob

Hi Rob,

Thanks for your elaborate feedback! I left out the single-shell glioma data since it complicated matters too much (no PA scan and no control subjects available with that protocol). I also kept the subject-specific response functions for now, and will examine the effects of averaging the RF across subjects later on.

Best,
Hannelore