Hello Simon, I finally managed to build RTK-2.5.0 against ITK-5.3.0 with CUDA support on Windows with VS2022. Since I have met some problems along the way, maybe this will help someone:
* Current CUDA 12.5 doesn't work, 11.8 does. Maybe the documentation should be updated? See https://discourse.itk.org/t/about-gpu-version-compilation-of-rtk/5905 * Before building ITK, I had to add "find_package(CUDAToolkit REQUIRED)" in the CMakeList.txt of the CudaCommon module to fix a CMake error * There still was a linker error in some ITK test project, which I ignored * There is a missing const in ITK-5.3.0\Modules\IO\XML\include\itkDOMNode.h which caused a compiler error when building RTK. This is fixed in ITK/master It also turns out that I apparently don't have enough GPU memory (8GB) to even use CUDA on full-resolution data: Description: ITK ERROR: CUDA ERROR: out of memory Best regards, Christian ________________________________ Von: Simon Rit <simon....@creatis.insa-lyon.fr> Gesendet: Dienstag, 4. Juni 2024 14:15 An: Christian Riesch <christian.rie...@live.com> Cc: rtk-us...@openrtk.org <rtk-us...@openrtk.org> Betreff: Re: [Rtk-users] Reconstruction geometry Hi, The way I calculate the --neworigin depends on the spacing as it passes the coordinates of the center of the first pixel which would change if you bin the data. Maybe you can use the first parameter I gave you originally. Note that you can keep the full resolution and bin the data with --binning 6,6 if you'd like, the origin is then automatically adjusted. For iterative reconstruction, I use rtkconjugategradient by default with a quadratic regularization of the gradient (parameter --gamma). But be aware that such truncated data are difficult to reconstruct. BTW, you might want to try the option --pad 1 to manage truncation in rtkfdk. Good luck! Simon On Tue, Jun 4, 2024 at 12:41 PM Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> wrote: Hello Simon, First of all, thank you for your efforts. The command line below indeed produces the best result I have seen using RTK so far. When going back to the full resolution images, is it correct to just set --newspacing=0.09, or do I need to modify something else too? It seems that this results in a much noisier reconstruction compared to the downscaled images. I also remember you mentioning iterative reconstruction. Do you have any pointers for that, besides using rtksart instead of rtkfdk? Best regards, Christian ________________________________ Von: Simon Rit <simon....@creatis.insa-lyon.fr<mailto:simon....@creatis.insa-lyon.fr>> Gesendet: Montag, 3. Juni 2024 16:14 An: Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> Cc: rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org> <rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org>> Betreff: Re: [Rtk-users] Reconstruction geometry Hi, FDK's formula is an integral over 360°. If there is an angular gap above --short (20° by default), rtkfdk will try to apply Parker short scan. If you disable Parker's weighting (with --short 360), then you'll get huge artifact because the two projections adjacents to the angular gap will get ~135° weights whereas the other projections get ~1° weight. I don't know what Astra does, it might simply apply a constant angular weights but rtkfdk is able to calculate non constant weights (which is required for, e.g. Elekta CBCT acquisitions). If you want to circumvent that, you can add 0-filled projections at the beginning and at the end of the acquisition. But rtk also guesses the flat field constant of your projections so you need to account for that too.... So I got some results by adding a 7000 filled 000.tif and 092.tif projections and running rtksimulatedgeometry -o"geo.xml" -n93 -f-46 --sdd=482.53 --sid=75.46 --arc 92 rtkfdk -p . -r '.*.tif' -o fdk.mha --newspacing=0.594 --neworigin=-151.767,-151.47,0 -g geo.xml --spacing 0.09 --dimension 512 --hardware cuda --short 360 --i0 7000 I can't say if the result is good, I don't know what I'm looking at and 90° is a really short scan. But you can maybe try if it's close to ASTRA's result. Cheers, Simon On Mon, Jun 3, 2024 at 11:16 AM Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> wrote: Hello Simon, I realize that a 90° scan only allows for a partial reconstruction of the volume. This approach seems to be quite common in industrial applications due to space and time constraints, but I have not found much information about it in the literature. The Astra toolbox's FDK implementation seems to cope with this quite well, at least in the center of the volume. The rescaled pixel size should be 6 * 0.099 mm = 0.594 mm. Best regards, Christian ________________________________ Von: Simon Rit <simon....@creatis.insa-lyon.fr<mailto:simon....@creatis.insa-lyon.fr>> Gesendet: Sonntag, 2. Juni 2024 08:00 An: Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> Cc: rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org> <rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org>> Betreff: Re: [Rtk-users] Reconstruction geometry Hi Christian, Thanks for the share. It was not clear to me that you had a 90° scan sorry. FDK is not appropriate for such a 90° scan, it's meant for a full 360° scan. rtkfdk tries to apply a short scan but Parker weighting is for 180°+fan-angle. So I'm not sure what to expect from FDK with such a scan. The projections that you shared are 512x511, can you please clarify the pixel size? I would try iterative reconstruction with such a scan. Cheers, Simon On Fri, May 31, 2024 at 12:25 PM Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> wrote: Hello Simon, Thanks for the quick response. So I have tried rtksimulatedgeometry with --arc 273 together with adding --neworigin=-153.5985,-153.2025,0 to rtkfdk. The result is different than before, but still looks very wrong (two ribbons and some vertical lines). I may be able to share some data later. Best regards, Christian ________________________________ Von: Simon Rit <simon....@creatis.insa-lyon.fr<mailto:simon....@creatis.insa-lyon.fr>> Gesendet: Donnerstag, 30. Mai 2024 08:48 An: Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> Cc: rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org> <rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org>> Betreff: Re: [Rtk-users] Reconstruction geometry Hi, I can see two problems: - you need to indicate the short scan to rtksimulatedgeometry, e.g. --arc 273 or, if the angle convention is opposite in your system, --arc -273. You may have to tune the --first_angle option to position the volume as you wish but that should not prevent accurate reconstruction. - you need to center your projections, i.e., add the option --neworigin=-153.5985,-153.2025,0 to rtkfdk, computed with -0.5*(3104-1)*0.099 and -0.5*(3096-1)*0.099. Don't hesitate to share a dataset if that does not help. Good luck! Simon On Thu, May 30, 2024 at 6:58 AM Christian Riesch <christian.rie...@live.com<mailto:christian.rie...@live.com>> wrote: Hello, I'm currently trying to use RTK for a reconstruction from a stack of images that resulted from a 90° scan, but struggling to get the parameters right. On the other hand, I was able to get what looks like a faithful reconstruction using the Astra toolbox and the script from https://tomroelandts.com/articles/astra-toolbox-tutorial-reconstruction-from-projection-images-part-1 I have these parameters from the X-ray device that captured the images (size 3104x3096): "fod_mm": 75.46, (focus to object distance) "odd_mm": 407.07, (object to detector distance) "pixel_size_mm": 0.099, "start_arc_deg": 45.0, "stop_arc_deg": -45.0 (There is also "detector_offset_mm": 9.8, which I have ignored so far with both Astra and RTK.) So I have tried to create a geometry using .\rtksimulatedgeometry.exe -o"geo.xml" -n91 -f45 -a90 --sdd=482.53 --sid=75.46 where sdd = fod_mm + odd_mm and sid = fod_mm, and a reconstruction with .\rtkfdk.exe -p"images" -r".*tif" -g"geo.xml" -o"out.tif" --dimension=512,512,200 --newspacing=0.099 However, out.tif contains mostly zeros and a weird ribbon-like shape in the center. So I think my geometry is wrong, but I don't know how to fix it. For comparison, here are the input parameters for the Astra script: distance_source_origin = 75.46 # [mm] distance_origin_detector = 407.07 # [mm] detector_pixel_size = 0.099 # [mm] num_projections = 91 start_angle_deg = 45 stop_angle_deg = 135 Any pointers would be appreciated. Best regards, Christian Riesch _______________________________________________ Rtk-users mailing list rtk-us...@openrtk.org<mailto:rtk-us...@openrtk.org> https://www.creatis.insa-lyon.fr/mailman/listinfo/rtk-users
_______________________________________________ Rtk-users mailing list rtk-us...@openrtk.org https://www.creatis.insa-lyon.fr/mailman/listinfo/rtk-users
