SVMBIR: Fast parallel-beam MBIR reconstruction#
svmbir is a Python package for Model Based Iterative Reconstruction (MBIR) of parallel-beam and fan-beam tomography data. It wraps the super-voxel C code sv-mbirct [A3] [A4], which runs on multi-core CPUs.
Key features:
Fast reconstruction: the super-voxel algorithm is 100 to 1000 times faster than conventional MBIR code on a CPU.
Parallel-beam and fan-beam geometries (see Overview).
Bayesian reconstruction with a qGGMRF prior, well suited to sparse-view and noisy data.
A proximal map interface for Plug-and-Play priors [A2] [A1].
Automatic parameter selection, with a small set of parameters for fine-tuning.
A function interface of a few calls:
project,backproject, andrecon.
For GPU reconstruction, cone-beam and other geometries, and new development, see MBIRTorch, the current package in the OpenMBIR family.
A reconstruction is one function call on a sinogram and its view angles.
Super-voxel coordinate descent uses every core and the cache well.
The proximal map interface accepts any denoiser as the prior.