CERN/TIGRE

Cone FDK and backprojection issues with 3kx3k panel

paskino opened this issue · 2 comments

paskino commented

Expected Behavior

Obtaining a reconstruction or backprojection without stripes.

Actual Behavior

We observe something that looks like the an int32 counter going beyond the max value when backprojecting (or FDK-ing) a cone beam dataset acquired with a 3054x3054 panel. This happens with fdk weights:

BackProjection
image
FDK
image

This for backprojection with matched weights
image

Code to reproduce the problem (If applicable)

This is a user report in CIL. I currently don't have a machine with TIGRE 2.4 to try this onto.

Specifications

  • MATLAB/python version:
  • OS: Windows
  • CUDA version:
  • TIGRE version 2.4 via CIL
AnderBiguri commented

Hmmmm, this is bad.

Let me investigate. I thought this was solved here (0778480), but I must have missed something.

AnderBiguri commented

@paskino I know I am super late on this, life got in the way. I think I need the user for this. It will likely be related not just to the image sizes, but the number of GPUs they have available, and which ones. I need that info to reproduce it.