mbirtorch

MBIRTorch: Model-Based Iterative Reconstruction (MBIR) for tomographic reconstruction using PyTorch.

Features include:

• Multiple geometries: parallel beam, cone beam (including curved detector and helical), translation mode, and multi-axis parallel.
• 4D reconstruction: a time sequence of volumes from a single continuous scan, using multi-agent consensus equilibrium (MACE).
• Preprocessing routines for NSI and Zeiss scanners, plus geometry calibration.
• Utilities for metal artifact reduction and stripe removal.
• Hyperspectral neutron data: maximum-likelihood dehydration into component maps and spectra, denoising, and the mbirtorch-hsnt command line (documentation).
• Interactive slice and geometry viewers.
• Informative demos and extensive documentation.
• Seamless operation on 1 or more GPUs, Mac MPS, or CPU.
• Compiled torch and Triton kernels for efficiency.

Available on PyPI via

pip install mbirtorch

Reconstruct in one line:

import mbirtorch
recon, recon_dict = mbirtorch.recon_simple_parallel(sinogram, angles)

Full documentation at https://mbirtorch.readthedocs.io/

Citation

Please use the following BibTeX citation when referencing this software.

@misc{mbirtorch,
  title = {{MBIRTorch}: {H}igh-performance tomographic reconstruction using {PyTorch}},
  author = {Gregery T. Buzzard and Charles A. Bouman and Jingsong Lin and Ziyun Li},
  howpublished = {Software library available from \url{https://github.com/cabouman/mbirtorch}},
  note = {Version 0.1.2},
  year = 2026
}

GitHub's "Cite this repository" button on the repository page generates this citation from CITATION.cff.

mbirtorch is a PyTorch port of MBIRJAX; please also cite it when referencing the underlying methods.