has someone already tried to read the generalized fractional anisotropy (GFA) or multi-directional anisotropy (MDA) out of the ODFs produced by “dwi2response dhollander” and “dwi2fod msmt_csd” (DOI:10.1002/jmri.24589) ?
The method described in Raffelt 2012 is, in my opinion, unsuitable for my specific question (lesions in different locations). Using the tensor model (FA,MD,AD,RD) to describe regions, which I have segmented with the magnificent MRtrix methods ,I consider to be unsuitable. If so, are there recommendations or reasons not to do so.
I haven’t personally, but I did have a go at GFA with the standard CSD many years ago: it’s no use at all, you get a more or less homogeneous output from that. Contrast can be re-introduced with different regularisation schemes: for instance, I’d expect that Flavio Dell’Acqua’s damped RL approach, which dampens the fODF in GM regions, would show reduced GFA in cortical regions, simply because the ODF is smoother in these regions, with a reduced range of values. I have no idea what that might look like with MSMT-CSD, but I’d expect it to show broadly similar behaviour.
In general, I don’t find the GFA particularly informative, and certainly not easily interpreted from a biological point of view. There are many ways to influence these values that have nothing to do with the pathology: just changing the lmax value used in the reconstruction will change these values drastically. And there is still no way to decouple the effect of crossing fibres from these metrics, as is clear from the maps shown in the relevant papers: there is still lots of contrast within deep white matter, where you’d expect a uniform value if these measures genuinely were reporting on the ‘intrinsic’ anisotropy with the effect of crossing fibres factored out…
Another point is that the MDA measure you mention is suitable only for q-space based methods, i.e. those that provide the diffusion ODF - the whole derivation depends on this. Moreover, it requires that the ODF be computed using a radial integral with solid angle consideration (the _r_² term in the radial integral), so even Q-ball wouldn’t be appropriate here (although CSA-QBI arguably might). On the other had, CSD and its variants are fundamentally not based on q-space, but on mixture models, and provide the fibre ODF. The MDA measure is simply not applicable to these ODFs.
I’m a bit confused here - why would it be unsuitable? More to the point, you seem to suggest that other scalar metrics like GFA or MDA might be more suited, which seems contradictory. AFD would at least give you the ability to be more specific about which directions were affected within crossing fibre regions, whereas the other scalar metrics would not (although arguable MDA might, but then you’d be using it more like Flavio Dell’ Acqua’s HMOA, which bears strong similarities with AFD anyway). But using these measures would certainly not avoid the issues related to lesions in different locations. In my opinion, the way the statistics are performed in the more recent CFE pipeline would be more likely to pull out entire tracks affected by focal lesions, assuming the lesions in your pathology group did consistently affect that specific pathway.
Not sure that’s all that helpful to you? Maybe you need to provide a bit more context about what it is you’re trying to do, and why you think approaches such as AFD might be unsuitable?
Might add a couple of questions / ideas on top of @jdtournier’s response, in case it helps clarify the question:
The method described in Raffelt 2012 is, in my opinion, unsuitable for my specific question (lesions in different locations).
Are you referring to multiple lesions in different locations within the same subject, or lesions differing in location between subjects? The former does not preclude any particular quantitative metric more than any other; for the latter, group analysis will be problematic regardless of whether or not the analysis technique has crossing fibre specificity.
Was the intended message here that the observed effect is focal, and does not extend along the relevant tract? If so, consider this: Although there may not be a spatially extended region of small FODs in subject native space, the total density of the pathway displaced by the lesion has to be smaller compared to if the lesion were not there; this is (ideally) captured in the registration component of the analysis framework. This paper does an excellent job (if I do say so myself) of explaining how we can both combine, and distinguish between, the two.
Using the tensor model (FA,MD,AD,RD) to describe regions … I consider to be unsuitable.
If you are trying to interrogate the local tissue composition within lesions, you may want to consider simply looking at the relative tissue intensities that come out of multi-tissue CSD, rather than limiting yourself to just looking at the WM component.
The latter would be challenging if your focus is specifically on investigating the nature of the lesions themselves; but certainly not impossible. However, if you’re just looking to perform fixel-based analysis in the presence of lesions (be they in different locations across subjects or not; but realistically they are for a lot of pathologies), there’s no real problem at all. Depending on the kinds of lesions you’re talking about (and hence, the population you’re investigating), the WM component (“apparent fibre density”) may still work very well, and still tell you a lot about what happened to the actual white matter (i.e. axons). As @rsmith hinted at, there may also be information in the relative tissue intensities from multi-tissue CSD; but no worries @Marc , I’ve got an accepted ISMRM abstract exactly on that topic. I’ll make it available as soon as the ISMRM proceedings themselves are available (which should be pretty soon now).
But in any way, @Marc, it would be helpful if you’d specify what kind of lesions you’re talking about (or which population). Are you talking about white matter hyperintensities (of the kind that would show brightly on a FLAIR)?
Hello! Sorry for my late answer, it’s a very complex issue for me. Thank you very much for your fast and detailed answer.
The lesions differing in location between subjects (FLAIR hyperintense MS lesions).
The 2017 Raffelt et. al. paper is actually methodically very interesting and elegant, but I would have to process all data again. My idea is to use a “total” voxel based AFD map and compare mean AFD values.
Firstly I use my preprocessed DW data (eddy, motion corrected …) and do:
group response funktion
After that I compute new FOD files for each subject and read out the AFD from the first volume ( fod_wm.mif ) and divide by 0.282. The masks for my ROIs are already there.
If i would have the resources I would do FBA, but do you think that is a reasonable compromise?
I don’t really understand what you mean by relative tissue intensities that come out of multi-tissue CSD or how can i do that
The 2017 Raffelt et. al. paper is actually methodically very interesting and elegant, but I would have to process all data again.
Given you’re already performing the necessary pre-processing steps, I’m not sure why a full FBA pipeline would require re-processing of all data, compared to what you’re proposing?
My idea is to use a “total” voxel based AFD map and compare mean AFD values … If i would have the resources I would do FBA, but do you think that is a reasonable compromise?
We wrote a paper related to this. It can be done; but in addition to the loss of fibre specificity, if you calculate the total AFD before registration, you lose the ability to modulate fibre density appropriately based on fibre orientation during spatial normalisation, and can hence only investigate differences in microscopic fibre density without modulation. Conversely, if you’re performing FOD-based registration and modulation, why then discard that crossing-fibre information just for the final analysis step?
I don’t really understand what you mean by relative tissue intensities that come out of multi-tissue CSD
The total AFD within a voxel is equivalent to the “white matter tissue intensity” that comes out of CSD. When performing multi-tissue CSD, you obtain “tissue intensities” for all tissues for which multi-shell response functions are provided. Note that we prefer to refer to “tissue intensities” rather than “volume fractions” or even “tissue volumes”, since we don’t normalise the tissue volumes to sum to 1.0 in each voxel, and are prone to T2 effects. All I was suggesting here was that if you don’t have any prior information about the tissue content of those lesions, looking at the WM component (“total AFD”) in isolation may be hiding some information.
It’s also worth noting that if you’re using single-shell CSD, with a single WM response function, then extracting the mean AFD from the resulting FODs, that’s almost equivalent to simply calculating the mean DWI signal.
I suppose @Marc probably meant “go back to the data to do some more processing again”, not per se redoing all pre-processing (as that’s indeed not necessary). Also maybe…
FBA does require some extra processing power/time, and unavoidably quite a bit of memory.
That said, we would of course always advise to aim for a full FBA. The question…
…is hard to answer in general. It depends on what you want to look into, and most importantly with how much precision. If you want to compromise, it’s just important to know exactly what FBA buys you that a voxel-wise analysis would not.
Fibre density and cross-section (FDC), the combined effect
…and that we were able to assess all of these properties fixel-wise, i.e. specific to single fixels, even if there’s more than one in given voxels.
In line with what @rsmith mentioned, a voxel-wise analysis will be limited to only assessing FD (not FC and FDC, since they require orientationally dependent modulation of specific fixels). Furthermore, that assessment of FD will not be fixel-specific. This may further mean that you will detect less (significant) differences: if a difference only exists for a given fixel in a voxel, the variance introduced by FD of the other crossing fixels may reduce the power to detect the difference. And of course, if you do find a difference, you’ll not be able to attribute it to a specific bundle for most voxels in crossing areas.
Finally note from that poster how we didn’t find differences specifically in lesions, due to their spatial location varying too much across subjects. But nonetheless, if using a multi-tissue CSD, one can account for the presence of these lesions quite well (and in single subjects, note the very clear reduction in FD in the actual lesions).
That’s of course always the case, since we only assess FD (and potentially FC/FDC as well in a FBA, but the extra information only comes from the spatial warps here). The more important point is that the FD can (probably) still be assessed quite well, even in the presence of other tissue types, since the latter may be modelled quite well by other tissue types. So you already win by applying multi-tissue CSD in and of itself; there’s no need to actually investigate the other types if the properties of the axons are what you’re after. If you’re after a group-wise whole brain analysis (be it voxel- or fixel-wise), you probably won’t even gain anything by looking at the other tissue types; as the spatial locations of the lesions may vary too much.
…and that’s why single-tissue CSD is almost certainly not specific enough in many cases. The WM FODs from multi-tissue CSD are already providing more specific out of the box, due to the sheer fact that they came from a fit that accounted for other tissue types.
It may indeed clarify quite a bit… and make you (want to) look a second time at those multi-tissue results you’ve got sitting there. Especially in subjects with white matter hyperintensities, such as the MS subjects you’re studying. Don’t hesitate to contact me privately, if you need some more info!