[Rtk-users] Reconstruction artefact with SART

[Simon Rit]

Thanks for sharing the result. We should definitely try to work on a more
efficient and matched projector / backprojection pair. In the meantime, my
feeling is that you could also use some spatial regularization to reduce
the effect of unmatched projectors.
Best regards,
Simon

On Tue, Oct 29, 2019 at 6:19 PM vincent <vl at xris.eu> wrote:

> Hi Simon,
>
> well the step size marginally improved the result.  But using
> CudaVoxelBased as the BP rather than CudaRayCast completely solved it !
>
> Thank you for your time and advice,
>
> best regards,
>
> Vincent
> On 27.10.19 21:42, Simon Rit wrote:
>
> Hi Vincent,
> I'm not sure why the step size would make such a large difference but
> that's an interesting lead indeed. Feel free to add the step size as an
> option, I hope this will indeed solve your problem!
> Best regards,
> Simon
>
> On Fri, Oct 25, 2019 at 11:55 AM vincent <vl at xris.eu> wrote:
>
>> Hi Simon,
>>
>> thank you very much for the suggestions.  I think I have a lead and I was
>> wondering if you could give me your opinion before I roll up my sleeves and
>> dive into the code.
>>
>> Since the issue is number-of-voxels-dependent, and because I use
>> CudaRayCast both for forward and backward projections, my guess was that it
>> has to do with  the forward projection operator, and more specifically with
>> the step size along the ray;  while it is an option in
>> rtkforwardprojections, it is not customizable in rtksart.  To investigate
>> further, I reprojected the reconstructed FDK volume (700x700x700 voxels,
>> voxel size = 0.23 mm) 3 times, once with Joseph, once with CudaRayCast and
>> the default step (1) and one with a custom step (0.1).  The vertical
>> streaks appear as I expected in the second case but not in the others:
>>
>> https://ibb.co/ZX57FPM
>> https://ibb.co/MnMk5Ny
>> https://ibb.co/gzfPQG1
>>
>> So I think that adding the step size option in rtksart should get rid of
>> my problem, but I would like your feedback before, as I trust your
>> experience more than my intuition :) !
>>
>> Thanks again,
>>
>> Vincent
>> On 23.10.19 22:30, Simon Rit wrote:
>>
>> Hi Vincent,
>> There is a problem indeed! It's hard to guess what's the problem from the
>> images you sent. I would check the projections to see what could be wrong.
>> The first thing I would check for example is whether air is really at 0 (as
>> it should be). You can also look at the difference between the forward
>> projections of this first iterate and the measured projections. It's known
>> that SART should be stopped at a few iterations but 1 should not give such
>> a result. You can always check if the conjugate gradient provides a better
>> result, you can add some regularization in it.
>> Good luck solving this!
>> Simon
>>
>>
>> On Wed, Oct 23, 2019 at 3:03 PM vincent <vl at xris.eu> wrote:
>>
>>> Hi everyone,
>>>
>>> I am trying to reconstruct a volume representing an aluminum casted part
>>> and investigate the differences between FDK and SART.
>>>
>>> Here are the parameters involved: the images are 700*820 pixels and I
>>> have 900 projections.  The physical pixel size is 0.28 mm and the sdd
>>> and sid are 558.3 mm and 461.3 mm respectively, which gives me a
>>> magnification factor of 1.2 roughly.  I performed one iteration of SART
>>> with CudaRayCast as forward and backward projectors and I wanted to
>>> reconstruct a 700*700*700 volume with a 0.23 mm (= 0.28/1.2) voxel
>>> size.  The result is shown on the two pictures attached:
>>>
>>> https://ibb.co/8sLdJxw
>>> https://ibb.co/N3NLg2D
>>>
>>> The result is getting worse as I increase the number of iterations
>>>
>>> I have a hard time understanding where the stripes come from.  A
>>> reconstruction with the standard 256*256*256 volume and 1mm spacing
>>> gives a "smoother" result, closer to what FDK gives.  I thought it might
>>> be related to the linear system resolution, but unless I am mistaken, I
>>> am trying to solve a system with 700*700*700 unknown with 700*820*900
>>> equations, which should be ok.
>>>
>>> Furthermore, the reconstruction with FDK at the same "full" resolution
>>> gives satisfactory results.
>>>
>>> Any help/explanation will be greatly appreciated !
>>>
>>> I thank you very much in advance,
>>>
>>> Vincent
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> Rtk-users mailing list
>>> Rtk-users at public.kitware.com
>>> https://public.kitware.com/mailman/listinfo/rtk-users
>>>
>>> _______________________________________________
>> Rtk-users mailing list
>> Rtk-users at public.kitware.com
>> https://public.kitware.com/mailman/listinfo/rtk-users
>>
> _______________________________________________
> Rtk-users mailing list
> Rtk-users at public.kitware.com
> https://public.kitware.com/mailman/listinfo/rtk-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://public.kitware.com/pipermail/rtk-users/attachments/20191029/dbae4b4a/attachment.html>