[Rtk-users] SIRT Reconstruction

[Cyril Mory]

Hi Lotte,

I had a very quick look at what you are doing. It looks fine to me, you 
just need to perform way more iterations in the SIRT case. 100 would be 
a good start. SIRT is more stable than SART when there are 
inconsistencies in the projection data, but converges slowly.

An alternative to SIRT, which minimizes the same cost function with a 
faster algorithm, is the conjugate gradient algorithm. You should obtain 
nice results with something like 30-40 iterations (look for 
rtk::ConjugateGradientConeBeamReconstructionFilter if you want to give 
it a try).

Best regards,
Cyril


On 18/07/2017 11:17, Lotte Schyns wrote:
> Hello,
>
> We are having problems with the SIRT reconstructions. They seem very
> strange and blurry. However, other reconstructions (SART, FDK,
> iterativeFDK) look perfect. I uploaded an example of a SART and a SIRT
> reconstruction (same parameters) to
> https://www.dropbox.com/sh/7a4bkyjg43zjmy7/AAAp4tJrk9HedMEEuIKsGYDSa?dl=0.
> I also uploaded a minimalistic version of the code that I used (the
> paths in CMakeLists.txt probably need to be adapted to your system). You
> can alternate between SART and SIRT by changing line 143. The raw data
> is also available on dropbox. I didn't use
> rtkXRadRawToAttenuationImageFilter, because our projections are already
> corrected for the dark field and flood field, so
> (signal-dark)/(flood-dark). I just take the natural logarithm and
> multiply by -1. Do you know what could be the problem with the SIRT
> reconstructions? Are we using wrong parameters? Thanks for your time.
>
> Lotte
> _______________________________________________
> Rtk-users mailing list
> Rtk-users at public.kitware.com
> http://public.kitware.com/mailman/listinfo/rtk-users