[Neuroimaging] [dipy] Import csd model precomputed by mrtrix
Paolo Avesani
avesani at fbk.eu
Fri May 27 05:07:54 EDT 2016
Thanks for your replies.
A brief update.
Samuel is right. I have in mind to manage the multi-shell model from
mrtrix3.
It is clear that up to now it is not viable until the work in progress from
Bago becomes available.
I tried to follow the example suggested by Ariel, focusing on a single
shell csd model (coming from dwi2fod command of mrtrix3),
and setting the sphere as the one used by mrtrix3 (a 300 points). The
"predict" method raised an issue of disalignment of matrix size.
The issue is at line 217 of csdecon.py when a product between
"predict_matrix" and "sh_coeff" takes place.
The "predict matrix" has size (90,45): 90 volumes, 45 sh
The "sh_coeff" has size (145, 174, 145, 45): the xyz dimensions of the
volume, 45 sh
It looks the size of predict matrix is wrong. It should be (145,174,145,90).
On Thu, May 26, 2016 at 4:07 AM, Samuel St-Jean <stjeansam at gmail.com> wrote:
> Indeed, I thought I read the multitissue version was used here, my mistake
> as the question does not directly imply that.
>
> The regular single shell hopefully does the same (at least I can
> personally attest the mrtrix2 version and dipy version give similar fodf up
> to a small rounding factor, but that was before the cholesky decomposition
> step, so now they should behave the same).
> On May 26, 2016 09:37, "Bago" <mrbago at gmail.com> wrote:
>
>> Samuel mrtrix has at least two models implemented. The multi-shell
>> (multi-tissue) model cannot be used with single shell data and the original
>> CSD model cannot be used with multi-shell data.
>>
>> Bago
>>
>> On Wed, May 25, 2016 at 5:13 PM Samuel St-Jean <stjeansam at gmail.com>
>> wrote:
>>
>>> Their wiki explains it, a sqrt(2) to normalize is used in mrtrix3, so
>>> multiplying your coefficients with that and using the mrtrix2 functions
>>> should do it.
>>>
>>> Although dipy only does single shell, so conclude with consideration
>>> that the algorithm is different from mrtrix3. Also, csd is a bad signal
>>> predictor (but good for angle estimation), see the sparc dmri challenge
>>> paper for example.
>>> On May 26, 2016 07:55, "Ariel Rokem" <arokem at gmail.com> wrote:
>>>
>>>>
>>>>
>>>> On Wed, May 25, 2016 at 4:41 PM, Bago <mrbago at gmail.com> wrote:
>>>>
>>>>> I believe they did change their basis (please correct me if I'm wrong
>>>>> but I believe they went from a non-normalized SH basis to a normalized SH
>>>>> basis).
>>>>>
>>>>>
>>>> So they have the same basis as dipy now, but the coefficients appear in
>>>> a different order? That should make life even easier!
>>>>
>>>>
>>>>> Also projecting onto a sphere is one way to _estimate_ the
>>>>> coefficients in a different basis. The cleaner way is to just re-order the
>>>>> coefficients and apply the appropriate scaling. If both basis are
>>>>> normalized (which dipy is) the scaling should be 1 or -1.
>>>>>
>>>>>
>>>> Fair point, but to be just a little bit facetious: given enough points
>>>> on the sphere and knowledge of the target maximal order of the
>>>> coefficients, wouldn't estimating be the same as transforming? Works for
>>>> the FFT, I believe :-)
>>>>
>>>> Bago
>>>>>
>>>>> On Wed, May 25, 2016 at 3:31 PM Ariel Rokem <arokem at gmail.com> wrote:
>>>>>
>>>>>> On Wed, May 25, 2016 at 1:09 PM, Bago <mrbago at gmail.com> wrote:
>>>>>>
>>>>>>> Hi Paolo,
>>>>>>> mrtrix and dipy define the SH basis slightly differently, so the
>>>>>>> precomputed FOD values need to be adjusted if you want to skip the fit step
>>>>>>> and initialize the Fit object directly. IRC we don't currently have the
>>>>>>> code to do that, but it would be something we'd like to incorporate.
>>>>>>>
>>>>>>> Did they change their basis set when they transitioned to mrtrix3?
>>>>>> We do have these functions:
>>>>>>
>>>>>> https://github.com/nipy/dipy/blob/master/dipy/reconst/shm.py#L852-L923
>>>>>>
>>>>>> That should work with the previous version of mrtrix (mrtrix2?). You
>>>>>> can use these to transform between coefficient sets:
>>>>>>
>>>>>> sf = sh_to_sf(mrtrix_coeffs, sphere, sh_order,
>>>>>> basis_type='mrtrix')
>>>>>> dipy_coeffs = sf_to_sh(sf, sphere, sh_order, basis_type=None) #
>>>>>> This defaults to the dipy basis set
>>>>>>
>>>>>> and then use the CSD model object to predict:
>>>>>>
>>>>>> from dipy.reconst.csdeconv import ConstrainedSphericalDeconvModel
>>>>>> csd_model = ConstrainedSphericalDeconvModel(gtab, response,
>>>>>> sh_order=sh_order) # Note: you still need to calculate the response
>>>>>> function!
>>>>>> pred_signal = csd_model.predict(dipy_coeffs, gtab, S0)
>>>>>>
>>>>>> I think that something like this should work (but I haven't tried it
>>>>>> myself).
>>>>>>
>>>>>>
>>>>>>> I have a WIP version of the multi-shell CSD model on a separate
>>>>>>> branch, I plan on merging it but wasn't intending to get to that for a few
>>>>>>> months. If you'd like to look at before then I can push the branch up to
>>>>>>> github.
>>>>>>>
>>>>>>> Sounds interesting! I'd love to see what you have so far!
>>>>>>
>>>>>> Cheers,
>>>>>>
>>>>>> Ariel
>>>>>>
>>>>>>
>>>>>>> Bago
>>>>>>>
>>>>>>> On Wed, May 25, 2016 at 2:31 AM Paolo Avesani <avesani at fbk.eu>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> I would like to take advantage of the "predict" method of
>>>>>>>> reconstruction models in dipy. The goal is to assess the quality of results.
>>>>>>>>
>>>>>>>> I have already computed the reconstruction models using mrtrix3 and
>>>>>>>> stored the ODF files. For this reason I would need to initialize the csd
>>>>>>>> model by importing the data from ODF stored by mrtrix3.
>>>>>>>>
>>>>>>>> The questions are manifold:
>>>>>>>> - may I initialize the csd model by providing the precomputed
>>>>>>>> values and skipping the "fit" step?
>>>>>>>> - may I import the value of precomputed model from a file stored by
>>>>>>>> mrtrix3?
>>>>>>>> - is the csd model in dipy compliant with the output of multi-shell
>>>>>>>> csd model computed by mrtrix3?
>>>>>>>>
>>>>>>>> I hope my questions and my goal is formulated clearly.
>>>>>>>> Thanks for your support.
>>>>>>>> Paolo
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> 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
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> 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
>
>
--
-------------------------------------------------------
Paolo Avesani
Fondazione Bruno Kessler
via Sommarive 18,
38050 Povo (TN) - I
phone: +39 0461 314336
fax: +39 0461 302040
email: avesani at fbk.eu
web: avesani.fbk.eu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/neuroimaging/attachments/20160527/f1e46da5/attachment-0001.html>
More information about the Neuroimaging
mailing list