PyTomography#
PyTomography is a python library for medical image reconstruction. It uses the functionality of PyTorch to (i) enable fast GPU-accelerated reconstruction and (ii) permit easy integration of deep-learning models in traditional reconstruction algorithms. If you use PyTomography in your own research, please cite the following: https://arxiv.org/abs/2309.01977
Features#
Supported Modalities
Single Photon Computed Emission Tomography (SPECT)
System matrix modeling includes attenuation correction, PSF modeling, scatter correction
2D Positron Emission Tomography (PET)
System matrix modeling includes attenuation correction and radially dependent PSF modeling.
Reconstruction Algorithms
Filtered Back Projection (FBP)
Statistical Iterative Algorithms
Options exist to include anatomical information (such as MRI/CT) when using priors/regularization.
Supported Datatypes
DICOM
Ability to open and align SPECT/CT data and create attenuation maps
Repository of collimator parameters for different scanners for obtaining PSF information
SIMIND output files (interfile)
Functionality to combine multiple sets of projections (representing different organs/regions) into a single set of projection data
Installation#
Firstly, it is important that all your graphics card drivers are up to date. To ensure this, go to this link, enter your graphics card information, and download/install the newest possible driver. It is also recommended that you download anaconda using this link to manage your python environments.
Firstly, create a pytomography virtual environment using anaconda:
conda create --name pytomography_env
Then you can activate your environment, and install PyTomography:
conda activate pytomography_env
pip install pytomography
Be sure to check out Tutorials for some simple examples. If you wish to make a contribution, please read the Developer’s Guide.
Examples#
Example 1: SPECT/CT Images of patient receiving targeted radionuclide therapy with Lu177 DOTATATE for neuroendocrine tumours (4 cycles of 7.4 GBq/cycle administered every 8 weeks). Row 1: OSEM (5 iterations, 8 subsets); row 2: BSREM (30 iterations, 8 subsets) using the RDP prior; row 3: BSREM (30 iterations, 8 subsets) using the RDP prior with anatomical weighting.
Example 2: Reconstructed SIMIND maximum intensity projections corresponding to activity distribution of typical patient treated with with Lu177 DOTATATE. Ground truth was generated using the XCAT phantom software. Shown are reconstructions with OSEM (8 iterations and 120 iteratons), BSREM with the RDP prior (120 iterations), and BSREM with the RDP prior and anatomical weighting (120 iterations). SPECT projections are overlaid with MIPs of the attenuation coefficient (hence the visible skeleton, even though it had no sigificant uptake).