Using CSD Dhollander for hippocampal gray matter imaging?

Hello,

I am interested in looking at gray matter (and a bit of white matter) diffusion within the hippocampus. I have just recently gotten into mrtrix and noticed some of your response functions and CSD methods that work with only 1 b-shell collected (+b0). So far I have predominantly looked at measures like FA and MD withing hippocampal gray matter using DTI, however, I was wondering if using CSD through the Dhollander response function might improve my measure by eliminating CSF contributions to my hippocampal diffusion measures.

Really interested in anyone’s thoughts.

Yes, I think it can be done and would be sensible – in fact, that’s pretty much the primary motivation behind this pull request

Just as an aside: what you’re talking about here is multi-tissue CSD – the dhollandher algorithm refers specifically to the estimation of the response functions, which is one step in the processing, and which can be done other ways (although it typically works great out of the box, so that’s what we’d normally recommend).

Hi @CallowBrainProject,

Happy New Year!
You might be interested in this work: https://www.biorxiv.org/content/10.1101/764506v1.full , where we showed that the CSF-like (or maybe rather, “free water”) signal fractions as well as all 3 tissue types can be estimated quite reliably, even for single-shell data, and in particular for the hippocampus. Note that, to estimate the free water signal fraction accurately in e.g. the hippocampal gray matter, it’s quite important to have all 3 tissues (WM, GM, CSF) in the model. In the absence of GM in the model, both WM and CSF will be overestimated.
The above preprint should soon be published (in the final minor revision stages now); in the final work, the experiments for the hippocampus are included for all datasets in the study. Just drop me a message, email, … if you want to be kept in the loop.

All the best,
Thijs

Yes. This is great. Does the output dive you some specific estimate of the average diffusion or shape of diffusion of each type of tissue within each video. I will definitely try this out and would appreciate it if you were to keep me in the loop with respect to your work/paper on it!

2 Likes

I’m not 100% sure what you’re after, but the direct output provides you with a WM FOD, and GM and CSF “compartments”. After normalising to sum to unity, you can get the tissue signal fractions; the latter is something I know several others are starting to look closer into now.

You could also use the estimated CSF compartment from a 3-tissue model to “remove” that compartment from the original data again, essentially free-water (signal) removal. It’s a bit more involved, even when the facilities would be there. In any case, once free-water signal is removed, you could e.g. estimate your existing metrics (FA / MD / …) again without free-water signal affecting them.
But that doesn’t remove the core problems with those metrics (FA / MD) in the first place though. You’d be better off going directly with the 3-tissue information itself: this indirectly also encodes the main “kinds” of properties of the data (anisotropy and signal decay / diffusivity). There’s a series of diagrams / visualisations in my ISMRM 2019 talk that might help you better understand some of these concepts (note this talk was mainly on the topic of the new response function estimation; but regardless of that it shows how response functions and the 3-tissue space of the 3-tissue CSD results relate to each other).

No worries at all; just connected on Twitter. :+1:

Just to clarify a couple of points:

  • the pull request I was talking about provides functionality to perform the forward convolution and mapping to amplitude (i.e. the inverse of multi-tissue CSD), which you can use to re-generate the predicted dMRI data, given an existing multi-tissue CSD decomposition. In your case, you can use this to predict your dMRI data without the contribution of the CSF signal, simply by omitting the CSF density and corresponding response function from the shconv call. You can then process these data as-is using whatever tools you were using (e.g. a tensor fit to get measures with little to no CSF contribution). But that relies on the availability of a multi-tissue CSD decomposition, which brings me to the next point.

  • I’d overlooked the fact you wanted to apply this to single-shell data specifically, in which case you’ll be restricted to a 2-tissue (CSF/WM) multi-tissue CSD (with the issues that @ThijsDhollander has highlighted here) or his single-shell 3 tissue CSD approach – my apologies for getting the wrong end of the stick here, I’d assumed you were talking about the response function estimation specifically. But the code I pointed to above can be used in either case to generate the predicted data.

Thank you both for the replies and input on the topic. After I am done calculating the response functions and WM FOD and compartments I will look into how I can go about estimating these metrics (such as FA and MD with “CSF-like signal removed”. I apologize that I am misunderstanding a few of the aspects of your tools. I just came across this work and thing it would be excellent to try and implement in some of our data so I am currently processing and trying to consume as much information about it as possible right now. Thanks again for the advice and input! I am sure it won’t be the last question I ask.

2 Likes

No worries, I figured that was the case. It’s good to emphasise the difference between response function estimation and CSD indeed though; this confusion seems to be quite common and often leads on to other confusions, including what kinds of data certain response function estimation methods as well as certain CSD methods can operate on, and under which assumptions.

In this context of free-water removal specifically, the difference between a 2-tissue and 3-tissue CSD model becomes quite interesting and consequential actually. As the 2-tissue model only has WM and CSF compartments, represented by fixed response functions, free-water removal by evaluating only the “remaining” WM compartment to predict dRMI signal means using only a single unique diffusivity (of the WM response) in that model. That’ll in turn lead to some worries about e.g. the MD from such data. Free-water removal is in that regard a very sensitive operation, because the bit that’s removed from the signal has a very large contrast between b=0 and other b-value(s).

All good. :+1: It might be a bit hard without the above mentioned functionality to conveniently remove the CSF-like signal (even though you can estimate it) at this stage. But in any case, see how you go with the CSD modelling itself. Note that the WM-like, GM-like and CSF-like compartments (and signal fractions, when normalised explicitly) themselves can directly serve though as useful quantification. The free-water fraction (or “CSF-like signal fraction”, as I would sometimes call it in this context) is a common marker of interest. It might also be more easy to interpret (e.g. biologically) compared to FA and MD.

… functionality to perform the forward convolution and mapping to amplitude (i.e. the inverse of multi-tissue CSD), which you can use to re-generate the predicted dMRI data, given an existing multi-tissue CSD decomposition. In your case, you can use this to predict your dMRI data without the contribution of the CSF signal, simply by omitting the CSF density and corresponding response function from the shconv call. You can then process these data as-is using whatever tools you were using (e.g. a tensor fit to get measures with little to no CSF contribution).

I would think that the better approach would be to perform the forward convolution of the CSF ODF and subtract the resulting estimated CSF signal from the DWI series; that way any residuals of the CSD fit are not removed prior to the fitting of some other model?

1 Like