[Mrtrix-discussion] SIFT: normalise to the b0 image

Wed Jan 20 02:14:14 PST 2016

Hi Scott,

Yes, the T2 shine-through is one of those issues that also puzzles me. I'm
not sure how best to handle it. If the high T2 could be assumed to arise
purely from increased free water, it would have CSF-like properties, and
hence contribute very little to the high b-value image, so all would be
good. But genuine shine-through still shows up with high signal in the DWI,
so there must be some other explanation. Unfortunately I've not had a
chance to really dig into the histology literature to figure out what might
actually be causing this.

In terms of how to handle this within MRtrix, you can use the multi-tissue
CSD algorithm that's recently been made available
<https://plus.google.com/101826858124420446685/posts/QhYaHLGcdmN> within
MRtrix3 <https://github.com/MRtrix3/mrtrix3/wiki> (although we need to tidy
a few things up on that front) to decompose your data into CSF and tissue
components (it won't be perfect with only 2 compartments, but it'll do a
decent job of removing the CSF signal from the FOD). However, this won't
necessarily help you since the areas you're concerned about probably don't
have a lot of CSF-like signal, and the output of the multi-tissue CSD is
also not normalised to the b=0, so you'd probably get very similar results
to what you already have...

Daan Christiaens showed recently that multi-tissue CSD running with 4
compartments
<https://lirias.kuleuven.be/bitstream/123456789/496046/3/MICCAI2015_CNSF_camready.pdf>
(WM, GM, CSF + edema) might allow edema to be distinguished from free water
(see Figure 6 in particular). But this needs a 4 b-value (b=0 + 3 shells)
sequence, which you don't have... But the point here is that the signal
from this particular pathology is clearly distinct from that of CSF.

So the question then becomes whether normalising to the b=0 is the right
thing to do in principle, even if the CSF signal weren't a problem. I think
if you do the maths, you'll find it's not as simple as it might seem, for
exactly the same reasons I've given previously regarding CSF contamination.
Consider a voxel containing 50% WM, 50% 'pathology', i.e. whatever is
causing the shine-through, be it edema or something more complex (maybe
demyelination is also contributing in MS...?). The raw DW signal (and hence
the FOD, given the linearity of the operation) would essentially be the
linear sum of the DWI signal from the two compartments, which means there
would be a large amount of contamination from the shine-through, presumably
leading to a drastic increase in noise for the FOD, along with a increased
isotropic component (assuming the shine-through signal is not directional -
not sure how true that is). The main point though is that the WM part of
that FOD should be appropriately scaled. If you normalised to the b=0
signal, you'd be computing: S = ( ½DW_WM + ½DW_ST ) / (½b0_WM + ½b0_ST) -
so just like in the case of CSF, assuming b0_ST >> b0_WM, what this
essentially accomplishes is a large overall reduction in the FOD amplitude,
WM and edema alike. The amount of contamination would be the same relative
to the FOD amplitude, but the FOD would be strongly attenuated if
normalising to the b=0. This might look cleaner, but it would definitely
cause the tractography to struggle or fail outright if trying to track
through such regions, even though there may be significant amount of
genuine WM present. Depending on your particular question, this may or may
not be a bias you can live with... Either way, the tracking would struggle,
due to large amounts of contamination when *not *normalised, or due to very
low FOD amplitudes when normalised.

So I don't how you could somehow remove the edema signal from the DWI
signal using 2 b-values only. It *might* be possible to perform some kind
of partial volume correction *if* you had an independent estimate of the
partial volumes, but this sounds far from trivial to do robustly. Maybe
Thijs can do some magic here - I'd definitely recommend you take him up on
his offer and have that coffee...

Cheers,
Donald.

On 20 January 2016 at 06:59, Thijs Dhollander <
thijs.dhollander at florey.edu.au> wrote:

> Hi Scott,
>
> Interesting issue this is indeed! So to summarise what you are reporting
> (just to make sure I'm getting it right): you're seeing hyperintense voxels
> at b=3000 in the MS lesion, but the ADC map shows no clear decrease in
> values compared to surrounding (supposedly healthy) WM tissue, is that
> correct?  Are you sure there's not even a slight decrease in ADC?  Due to
> the log-transform, it might be hard to spot; definitely try to increase the
> contrast in your viewer to investigate the area.  I'm not an expert on MS,
> but doing a few searches, I seem to run into articles describing actual
> increased restriction in certain (stages of) MS lesions (presenting ADC
> maps with hypointense lesions.
>
> Let's say you would indeed be observing no decrease in ADC at all, yet a
> clear increase in b=3000 DWI values, that would mean an obvious increase in
> value of the b=0 image(s): are you seeing that as well in the data at hand?
>
> If the latter is true, I think I can potentially help you out with these
> partial volume issues (over time).  I notice you're at Unimelb.  I'm
> working at the Florey (in Heidelberg, but mobile and in the Parkville area
> at times); we should have a coffee and a chat some time, and take a good
> hands-on look at those data you're sitting on. ;-)
>
> Cheers,
> Thijs
>
>
>
> On 20 January 2016 at 14:54, Scott Cameron Kolbe <kolbes at unimelb.edu.au>
> wrote:
>
>> Dear Donald, Rob and Dave
>>
>> Just found this discussion whilst going through the archives. Sorry to
>> bring it up again but I’ve been having similar concerns to Jan recently
>> whilst working on multiple sclerosis data. The T2 “shine-through” effect
>> that Jan is referring to in MS results from T2 signal enhancement that is
>> not related to changes in CSF volume fraction.  Therefore hyperintense
>> voxels in DWI, considered by SD to be highly restricted diffusion, are in
>> fact not that different from surround tissue on the ADC map.
>>
>> This problem might be solved by acquiring low-B images for normalisation
>> where freely diffusing fluid is essentially annulled but even lightly
>> restricted water is not. However, we are sitting on quite a bit of data
>> with only b=0 or b=3000. Do you think partial volume correction might be
>> possible?
>>
>> cheers
>> Scott
>>
>>
>> Scott Kolbe
>> kolbes at unimelb.edu.au
>>
>>
>>
>> _______________________________________________
>> Mrtrix-discussion mailing list
>> Mrtrix-discussion at www.nitrc.org
>> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>>
>
>
> _______________________________________________
> Mrtrix-discussion mailing list
> Mrtrix-discussion at www.nitrc.org
> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>
>

-- 
*Dr J-Donald Tournier (PhD)*

*Senior Lecturer, **Biomedical Engineering*

*Division of Imaging Sciences & Biomedical EngineeringKing's College London*

*A: Department of Perinatal Imaging & Health, 1st Floor South Wing, St
Thomas' Hospital, London. SE1 7EH*
*T: +44 (0)20 7188 7118 ext 53613*
*W: http://www.kcl.ac.uk/medicine/research/divisions/imaging/departments/biomedengineering
<http://www.kcl.ac.uk/medicine/research/divisions/imaging/departments/biomedengineering>*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20160120/6374bb58/attachment-0001.html>