pyronn-torch

This repository provides PyTorch bindings for PYRO-NN, a collection of back-propagatable projectors for CT reconstruction.

Feel free to cite our publication:

@article{PYRONN2019,
author = {Syben, Christopher and Michen, Markus and Stimpel, Bernhard and Seitz, Stephan and Ploner, Stefan and Maier, Andreas K.},
title = {Technical Note: PYRO-NN: Python reconstruction operators in neural networks},
year = {2019},
journal = {Medical Physics},
}

Installation

From PyPI:

pip install pyronn-torch

From this repository:

git clone --recurse-submodules --recursive https://github.com/theHamsta/pyronn-torch.git
cd pyronn-torch
pip install torch
pip install -e .

You can build a binary wheel using

python setup.py bdist_wheel

Important

If you're using an older CUDA version you might get an error about '-allow-unsupported-compiler' not being a valid compiler option. In that case remove that compiler option from this project's setup.py.

Usage

import pyronn_torch
import numpy as np

projector = pyronn_torch.ConeBeamProjector(
    (128, 128, 128),  # volume shape
    (2.0, 2.0, 2.0),  # volume spacing in mm
    (-127.5, -127.5, -127.5),  # volume origin in mm
    (2, 480, 620),  # projection_shape (number of projections, height, width)
    [1.0, 1.0],  # projection_spacing in mm
    (0, 0),  # projection_origin
    np.array([[[-3.10e+2, -1.20e+03,  0.00e+00,  1.86e+5],
               [-2.40e+2,  0.00e+00,  1.20e+03,  1.44e+5],
               [-1.00e+00,  0.00e+00,  0.00e+00,  6.00e+2]],
              [[-2.89009888e+2, -1.20522754e+3, -1.02473585e-13,
                1.86000000e+5],
               [-2.39963440e+2, -4.18857765e+0,  1.20000000e+3,
                1.44000000e+5],
               [-9.99847710e-01, -1.74524058e-2,  0.00000000e+0,
                6.00000000e+2]]])  # two projection matrices in shape (n, 3, 4)
                # optionally: source_isocenter_distance=1, source_detector_distance=1 for a scalar weighting the projections
)
projection = projector.new_projection_tensor(requires_grad=True)

projection = projection + 1.
result = projector.project_backward(projection, use_texture=True)

assert projection.requires_grad
assert result.requires_grad

loss = result.mean()
loss.backward()

Or easier with PyCONRAD (pip install pyconrad)

projector = pyronn_torch.ConeBeamProjector.from_conrad_config()

The configuration can then be done using CONRAD (startable using conrad from command line)