Overview

MBIRJAX is a Python package for Model Based Iterative Reconstruction (MBIR) of images from tomographic data.

Here are the reasons to use MBIRJAX:

• Image quality: MBIR offers the best image quality because it uses a forward (sensor) and prior (image) model.

• Ease of use: MBIRJAX has built-in automatic parameter selection, so it will produce a good reconstruction the first time.

• Speed: MBIRJAX is fast (for MBIR) because:

  • GPU power: Uses JAX to harness the power of GPUs, CPUs, and clusters.

  • VCD algorithm: The vectorized coordinate descent algorithm has fast convergence to both reduce reconstruction time and improve image quality.

• Flexibility:

  • OO Python interface: Based on object-oriented python, so it is modular and easy to use.

  • Plug-and-Play prior models: Supports proximal map interfaces, so it can be used with PnP deep neural net priors.

We provide simple bash demo scripts located in [mbirjax/demo] that make it easy to get started. We also have a bash install script at [mbirjax/demo]. Also, installing JAX is not too difficult on most platforms and is getting easier.

Geometry

Right now MBIRJAX supports parallel-beam imaging geometry as shown below, but more geometries are on the way.

Parallel-beam geometry