[Neuroimaging] [Dipy] MAPMRI between subject comparisons
Rutger Fick
fick.rutger at gmail.com
Fri Jan 12 18:40:33 EST 2018
Hi Ping,
All I can say is that dipy's RTOP is in 1/mm^3, I don't know about TORTOISE.
Looking at the images it looks like TORTOISE's RTOP estimation has some
areas that produced negative values (black patches in the middle of the
bright white areas in the middle of the left image). This indicates their
regularization is failing.
On the other hand, I think maybe you can tweak dipy's mapmri settings to
make the RTOP estimation look even better. Try the following setting:
mapmod = mapmri.MapmriModel(gtab,
laplacian_regularization=True, # this
regularization enhances reproducibility of estimated q-space indices by
imposing smoothness
laplacian_weighting="GCV", # this makes it
use generalized cross-validation to find the best regularization weight
positivity_constraint=True) # this ensures
the estimated PDF is positive
MAPL is just a name for imposing signal smoothness during MAP-MRI's signal
fitting, which is in fact the default setting when you fit dipy's mapmri (
so you've already been using it).
The general idea is that imposing a bit of smoothness in MAP-MRI's signal
reconstruction will make estimation of q-space indices (RTOP RTAP NG etc.)
more robust. The "GCV" option makes the weight selection of the
laplacian-regularization data dependent using generalized cross-validation.
Setting positivity_constraint=True then also forces the solution to have a
positive PDF, which was the original approach by Ozarslan et al (2013).
Using both MAPL and positivity should then give you the best possible
reconstruction of RTOP and others.
The relevant citation for the positivity constraint is [1], while for MAPL
is [2]. Just cite both those papers when you use dipy's implementation :-)
Best,
Rutger
[1] Özarslan, Evren, et al. "Mean apparent propagator (MAP) MRI: a novel
diffusion imaging method for mapping tissue microstructure." *NeuroImage*
78 (2013): 16-32.
[2] Fick, Rutger HJ, et al. "MAPL: Tissue microstructure estimation using
Laplacian-regularized MAP-MRI and its application to HCP data." *NeuroImage*
134 (2016): 365-385.
On 8 January 2018 at 17:15, Ping-Hong Yeh <pinghongyeh at gmail.com> wrote:
> Hi Mauro, Rutger,
>
> Thank you both for the instructive reply.
>
> We used SMS sequence (GE 3T) to acquire multi-shell DWI (three shells b=
> 1000, 2000, and 3000 with 90 directions for each shell and an additional 18
> volumes of b0) with an inter-slice acceleration factor of 3 and an in-plain
> acceleration of 2 (ASSET). The original data has N=130* 130 with an
> in-plain resolution of 1.7 *1.7 mm, and DWI data was registered to T2W with
> a final resolution of 1*1*1 mm using ANTS nonlinear warping implemented
> in TORTOISE toolkit for distortion correction, after running some
> preprocessing steps such as noise reduction, bias correction.
>
> I have not calculated the mean SNR for each shell, i think there is below
> 5 in b=3000.
>
> I have attached two examples of RTOP snapshots, one was created by
> TORTOISE, and one by DIPY using "map_model.fit" of the same data. There
> is a difference of a difference of scaling around 10^6 between the two, I
> think this is due to the difference of unit used for the DTI fitting
> between two tools?
>
> I am concerned about high-intensity values over the subcortical regions,
> such as brainstem, and inhomogeneity over the main white matter tracts such
> as corpus callosum. Are those seemed normal looking to you? If not, what
> can cause those artifacts? Are suggestions in correcting those?
>
> What even puzzles me is that i found there is a significant reduction of
> Principal Anisotropy in the disease group than the control group (which
> is not what i expected) over the frontal white matter and gray matter of
> insular region using voxel-wise analysis after correcting multiple
> comparisons. This is done by using the outputs from the TORTOISE toolkit. I
> am still working on the DIPY ones.
>
>
> Rutger,
>
> I have not tried the MAPL method yet, but i will definitely give it a shot
> later. What additional information or aspects does MAPL provide us,
> comparing to the conventional MAP-MRI method? BTW, can DIPY output Principal
> Anisotropy map and Non-Gaussianity map as well? If so, what is the syntax
> for making those?
>
> Thank you,
>
> Ping
>
>
> On Fri, Jan 5, 2018 at 5:53 AM, Rutger Fick <fick.rutger at gmail.com> wrote:
>
>> Hello Ping,
>>
>> It is hard to debug what you're doing without any other information about
>> your model settings or what data you're fitting.
>> I will just describe some possible issues that you could be running into:
>>
>> - MAP-MRI metrics will not be reproducible across subjects if the
>> gradient tables of these subjects are different. Reason is that MAP-MRI is
>> basically a non-parametric inter/extra-polator and will just smoothly
>> attenuate to zero after the last b-value point in acquisition scheme. If
>> this point changes between schemes, then the extrapolation begins at
>> different points, which results in different q-space index values (because
>> they are based on the extrapolation to infinity). Relevant references would
>> be [1, 2], where is shown that MAPMRI qspace indices *are *reproducible
>> between subjects (with the same scheme).
>> - The scaling factor is calculated internally using DTI, so that
>> should not be a user issue. It is worth looking at the
>> fitted_mapmri_model.mu property, which contain the estimated scaling
>> factors (ux, uy, uz). If all these factors are the same (or have been
>> truncated to their minimum allowed value), this means there is a problem
>> with the data itself (DTI failed to fit properly).
>> - With respect to artifacts, I'm not sure what kind of artifacts
>> you're seeing, but depending on the problem several things could be
>> happening.
>> - If you're using laplacian regularization and the data is very
>> anisotropic, then the automatic regularization estimation using GCV could
>> be giving a too low value. Try setting it to a fixed value (0.2 for
>> example), or if it was already set to a fixed value, try increasing it to
>> see what happens (but not too much because you'll just be making everything
>> flat).
>> - If you're not using positivity constraint as well -> use it as
>> well. The best results are typically found when both GCV and positivity
>> constraint are used (but it also takes the longest to fit).
>> - Of course,if the data itself is bad (very noisy or some crazy
>> distortion) then MAPMRI cannot do much about it. As I said above, it will
>> just smoothly fit the data it is given. It is always important to look at
>> the data itself you are fitting, and if you see a very badly distorted DWI,
>> then remove it from from the data set.
>>
>> Let us know what kind of acquisition schemes you're using, and explain
>> what kind of artifacts you're seeing. Hard to make a concrete judgement
>> otherwise.
>>
>> Best,
>> Rutger
>>
>> [1] Avram, Alexandru V., et al. "Clinical feasibility of using mean
>> apparent propagator (MAP) MRI to characterize brain tissue microstructure."
>> *NeuroImage* 127 (2016): 422-434.
>> [2] Fick, Rutger HJ, et al. "MAPL: Tissue microstructure estimation using
>> Laplacian-regularized MAP-MRI and its application to HCP data."
>> *NeuroImage* 134 (2016): 365-385.
>>
>>
>>
>> On 5 January 2018 at 11:46, Mauro Zucchelli <mauro.zucchelli88 at gmail.com>
>> wrote:
>>
>>> Hi! Low SNR in multi-shell data with high b-values are a problem for all
>>> the higher order models, including MAP-MRI and many compartmental models.
>>> Moreover, MAPMRI presents numerous parameters tha you can adjust in
>>> order to maximize its performances. Can you give us more information on
>>> your dataset? E.g. SNR, number of samples, number of b-values, etc.
>>>
>>> Kind regards,
>>>
>>> Mauro
>>>
>>> On Wed, Jan 3, 2018 at 8:43 PM, Ping-Hong Yeh <pinghongyeh at gmail.com>
>>> wrote:
>>>
>>>> Hi Dipy users,
>>>>
>>>> I am wondering if the MAP-MRI measures such as RTOP, RTAP, RTPP, NG
>>>> etc are ready to use for between-subject comparisons. Are there any scaling
>>>> factor that needs to be applied beforehand.
>>>>
>>>> I've noticed that MAP-MRI measures are very susceptible to artifacts.
>>>>
>>>>
>>>> Thank you.
>>>>
>>>> Ping
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Neuroimaging mailing list
>>>> Neuroimaging at python.org
>>>> https://mail.python.org/mailman/listinfo/neuroimaging
>>>>
>>>>
>>>
>>> _______________________________________________
>>> Neuroimaging mailing list
>>> Neuroimaging at python.org
>>> https://mail.python.org/mailman/listinfo/neuroimaging
>>>
>>>
>>
>> _______________________________________________
>> Neuroimaging mailing list
>> Neuroimaging at python.org
>> https://mail.python.org/mailman/listinfo/neuroimaging
>>
>>
>
> _______________________________________________
> Neuroimaging mailing list
> Neuroimaging at python.org
> https://mail.python.org/mailman/listinfo/neuroimaging
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/neuroimaging/attachments/20180113/5e8f8bf1/attachment.html>
More information about the Neuroimaging
mailing list