Longitudinal Fixel-Based Analysis in Paediatric Temporal Lobe Epilepsy surgery

[luislacerda]

Dear MRtrix team,

I hope all of you are well.

I am currently investigating changes in a paediatric cohort that underwent temporal lobe epilepsy surgery, both at a global connectivity level and individual white matter pathways. In particular, I have multi-shell data (13 interspersed b=0; b=1000 / 60 directions and b=2200 / 60 directions) before and after surgery, and I would like to use Fixel Based Analysis so that I could disentangle differences in specific white matter pathways. After having done all the required preprocessing (denoising, Gibbs ringing removal, Motion and Eddy/Susceptibility distortions and bias field correction) I still have a few remaining questions before proceeding and would really welcome your comments:

• First of all, I should note, that as I expect to find different changes depending if the operation was on the left or on the right (according to previous literature) I believe performing the FBA independently for left and right sided patients is the way to go; however I believe the response function should be fairly similar before surgery, so would I was thinking of using both left and right sided patient data to create a group response function (before surgery) prior to continuing the FBA analysis individually for left and right sided patients. I have also read in another post that I could use all of the data to create an group average response, but I have some reservations because of the different changes that can occur after surgery. Any comments?

• Secondly, I understand that one of the advantages of the FBA analysis is to benefit from FOD based registration which we can also use to register other data, independently for this analysis. However, I am concerned that attempting to create a population template including pre and post-surgical data will result in poor registration – do you have any experience in that? Alternatively, I was thinking of creating a template based (with ‘’‘population_template’’’) on the “intact” pre-surgical data (no resections) and then use transformations that I have derived from ANTS registration between pre and post surgical FA (using resection segmentation as a mask) to propagate individual FODs to the “intact” template space. Does it sound like the right way to go or would you try to do the initial template generation with all of the data (pre and post surgical together)?

• Thirdly, having seen first uses of longitudinal FBA analysis which applies to my case – timepoint 1 (before surgery) and time point 2 (after surgery) - I am still undecided whereas the best approach is just to do a “standard” FBA with pre and post surgical data in one template vs creating a “midway-point” template onto which I would project the individual FODs.

• Finally, I was wondering whether the data corresponding to resected areas should be kept or removed from the analysis, prior to running tractography on the fixel template (template and white matter masks). In voxel-based analysis I am expecting to see immediate changes in the regions surrounding the resection (as I have with TBSS) and so I removed the resected area, but with FBA, those changes might be more pronounced in the whole extent of the tracts and if I remove data I may change interpretation of the results…?

Thank you very much for all the work you put in these tools and your time spent in helping others!

Luís

[rsmith]

Hi Luis,

1. With this sort of question, I’ve become that annoying person that keeps reminding people to go all the way back to their fundamental hypothesis rather than thinking about what they “can” do. If your questions are “how does left temporal lobectomy change the brain white matter” and “how does right temporal lobectomy change the brain white matter”, then having them as two separate cohorts makes sense. Whereas if you wanted to ask the question “how does temporal lobectomy change the brain white matter”, you may instead want to perform your analysis considering ipsilateral vs. contralateral hemispheres, which would involve axis flipping.

   More specifically with respect to response functions, I don’t personally see a mechanism by which the response functions estimated from LHS temporal lobectomy and those estimated from RHS temporal lobectomy would differ substantially. Indeed if that were the case, it would suggest that lobectomy itself changes response function estimation so much that it may be preferable to instead use only pre-surgical data for RF estimation.

   If the two cohorts are to be handled entirely independently of one another, then there’s no issue with having separate response functions for each; they’re basically two entirely independent experiments. However doing so precludes many future possibilities for collating and comparing data across the two cohorts. So personally I’d be using a common response function across the two, as the alternative may be the source of limitations at a later date.

2. There’s actually a number of different things all bundled up within this one question:

   1. Use of all data in population template generation versus using pre-surgical data only. While the template may look “nicer” if it were generated from pre-surgical data only, it may mean that the alignment of pre-sugical data to the template is accurate where the alignment of post-surgical data to the template is not. A benefit of generating the template image based on your actual experimental population (or a representative subset of such) is that the inadequacies associated with registration to the template is distributed among all inputs. I.e. If your template was generated only from pre-surgical data, and then you observe a difference between pre-surgical and post-surgical data, is the difference due to the surgery, or is it because the registration of pre-surgical data was superior to post-surgical data? The image alignment is never going to be perfect here as part of the brain has literally been removed, but my own intuition is that you want your template space to be the midway between pre-surgical and post-surgical data, just like we use a midway space for all image registration to remove preferential biases.

   2. If you did indeed generate the template using only pre-surgical data, that doesn’t preclude the use of FOD-based registration. You could still register post-surgical FOD data to that template using mrregister (and you would also need to explicitly run mrregister on the pre-surgical data rather than utilising the warps provided by population_template so that data processing would be identical across the input data). You can also utilise masking within mrregister to prevent certain areas from driving the registration.

   3. Perhaps within the expression of your question is the idea that the combination of ANTs registration and the FA metric would be superior in the case of temporal lobectomy to MRtrix3 registration with the FOD coefficient sum-of-squares difference metric. While I can’t predict naively which of these two combinations may work “better” in this case, the fact that masking is a capability within mrregister means that you’ve not actually expressed an adequate justification for switching to ANTs for this specific purpose. You would need some form of evidence that for your specific data, deviating from FOD-based registration because of the presence of resection provides sufficient benefit to justify doing so.

3. I’m failing to see the distinction between the two cases you describe here: both involve a single FOD template image, pre- and post-surgical data, and transformation of FOD data onto that template. What is the intended difference between “do a standard FBA … in one template” versus “create a midway template and project the FODs to that”?

4. This again comes back to point 1 RE: stating the hypothesis. The fact that you expect a priori for the resected area to show an effect doesn’t necessarily mean that you want to omit such from the analysis, it just means that you’re looking for statistically significant effects that are above and beyond what appears to be a direct consequence of the resection.

   I think what you’re trying to get at here is the spatially extended effects of data smoothing and statistical enhancement. Ignoring the nature of temporal lobectomy for a moment, if you were to synthetically introduce a statistical effect in one specific region of the template, then the data “connected” to that region (whether spatially proximal in the case of VBA, or fixel-fixel connectivity in the case of CFE) may be shown as statistically significant as a kind of “propagation” of that effect. Here, I would firstly ask about the time period between resection and scan. If you were to scan the subject immediately after resection, then in comparison to the pre-surgical scan, one would expect the resected area to “disappear” and for the “connected” regions to not show any difference. In this case, if the resected region were included in the processing mask, then observation of a statistically significant effect external to the resected area would essentially be an artifact of the smoothing / statistical enhancement. Conversely, if the scan occurs many years after the surgery, then one would expect any white matter pathways that once connected to the resected region to have degenerated. In this case, observation of a statistically significant effect there makes sense; but whether or not such an observation is reflective of the underlying biology or an artefact of smoothing / statistical enhancement would depend on the relative magnitudes of the contributing factors (and hence the underlying parameters of smoothing and statistical enhancement).

   There are further technical possibilities here with respect to characterising the connectivity to the resected region and looking for white matter differences “over and above that caused by the resection alone”. But that gets very technical, and starts using capabilities that I’ve not yet had the chance to publish, or potentially even capabilities that are not yet implemented. If you want to go even further down this rabbit hole, that probably becomes a private conversation, as I’m now in an Epilepsy lab and would need to resolve such against my own plans.

Cheers
Rob

[luislacerda]

Hi Rob,

First of all, thank you so much for the time you spent getting back to me - it is really appreciated.

@rsmith

If your questions are “how does left temporal lobectomy change the brain white matter” and “how does right temporal lobectomy change the brain white matter”, then having them as two separate cohorts makes sense

I might end up considering the more general question of whether temporal lobectomy changes the brain white matter, but my main question is whether those changes will be different depending on the operated hemisphere.

@rsmith

So personally I’d be using a common response function across the two, as the alternative may be the source of limitations at a later date.

Understood.

@rsmith

You can also utilise masking within mrregister to prevent certain areas from driving the registration.

@rsmith

You would need some form of evidence that for your specific data, deviating from FOD-based registration because of the presence of resection provides sufficient benefit to justify doing so.

Ahh, I didn’t know this - Yep I will definitely try FOD-based registration and compare to what I already have and draw my conclusions then.

@rsmith

I’m failing to see the distinction between the two cases you describe here: both involve a single FOD template image, pre- and post-surgical data, and transformation of FOD data onto that template. What is the intended difference between “do a standard FBA … in one template” versus “create a midway template and project the FODs to that”?

I didn’t express myself very well here - I was probably just thinking of the design that I would apply and whether I should consider any difference images before creating the template. But I think it is better to keep things simple for now.

@rsmith

Here, I would firstly ask about the time period between resection and scan. If you were to scan the subject immediately after resection, then in comparison to the pre-surgical scan, one would expect the resected area to “disappear” and for the “connected” regions to not show any difference. In this case, if the resected region were included in the processing mask, then observation of a statistically significant effect external to the resected area would essentially be an artifact of the smoothing / statistical enhancement.

That was my intuition, but it makes more sense now!

@rsmith

If you want to go even further down this rabbit hole, that probably becomes a private conversation, as I’m now in an Epilepsy lab and would need to resolve such against my own plans.

It sounds good! At the moment, I will give it a go with all your useful suggestions and mature my research hypothesis, but I will be sure to reach out if/when it all starts making more sense.

Thank you very much on again.

Cheers

Luís