Real-time 3D cone beam reconstruction

The paper presents a comparison of filtered back-projection and iterative approaches (modified SIRT and OSEM) where backprojection is performed on FPGA hardware and forward projection is realized by a graphics card. The FPGA-architecture allows scaling parallel processing of the backprojection with a constant memory bandwidth and it performs its task about a factor of 6 faster than conventional graphics card architectures (using a Xilinx Virtex II 6000 compared to a floating-point implementation on a graphics hardware). For the forward projection a graphics card (with NVIDIA GeForce FX 5900) is used. The system reconstructs a volume of 256³ voxels and 60 projections (5122) within 6 seconds for each backprojection step and 12.5 s for the forward projection. As assumed filtered backprojection is the fastest method but shows most artifacts. SIRT and OSEM generate results with suppressed artifacts and require between 4 (SIRT) and 3 (OSEM) minutes for reconstruction. The limited precision of on the FPGA does not lead to any degradation in the reconstructed image quality.