## FRAMEWORK

The python framework for the PYRO-NN layers implemented in (https://github.com/csyben/PYRO-NN-Layers)

## PYRO-NN

PYRO-NN brings state-of-the-art reconstruction algorithm to neural networks integrated into Tensorflow.

To use pyronn you need to build the operators from sources or install the provided binaries from https://github.com/csyben/PYRO-NN-Layers

The publication can be found under (pre-print coming soon)

## Installation

Install via pip :

`pip install pyronn`

or if you downloaded this repository (https://github.com/csyben/PYRO-NN) using:

`pip install -e .`

If you encounter a problem during the installation have a look at our wiki: https://github.com/csyben/PYRO-NN/wiki

## Changelog

Can be found CHANGELOG.md.

## Usage

PYRO-NN comes with all relevant helper classes to easily run the projection and back-projection operators within the Tensorflow context.

To use the Layers a geometry object is needed:

```
from pyronn.ct_reconstruction.geometry.geometry_parallel_2d import GeometryParallel2D
volume_size = 256
volume_shape = [volume_size, volume_size]
volume_spacing = [1, 1]
# Detector Parameters:
detector_shape = 512
detector_spacing = 1
# Trajectory Parameters:
number_of_par_projections = 360
angular_range = 2 * np.pi
# create Geometry class
par_geometry = GeometryParallel(volume_shape, volume_spacing, detector_shape, detector_spacing, number_of_fan_projections, angular_range)
```

After defining the basic geometry parameters, a trajectory need to be set. The circular_trajectory class computes an idealiyed circular trajectory for a given geometry. For 2D parallel- and fan-beam geometry a trajectory is described using the central ray vectors. For 3D cone-beam geometry the trajectory is described with projection matrices.

The trajectory can be calculated and set as follows:

```
from pyronn.ct_reconstruction.helpers.trajectories import circular_trajectory
par_geometry.set_central_ray_vectors(circular_trajectory.circular_trajectory_2d(par_geometry))
```

At this point the geometry is fully setup and can be used to create projections and reconstructions. The Layers just takes the respective input tensor and the geometry object to conduct the projection, reconstruction respectively. PYRO-NN also provides convinient general way to create sinograms and reconstructions. The generate methods are generalized and take the input data, the layer to be used and the geometry. The only restriction is that the generation methods are within the Tensorflow session scope:

```
from pyronn.ct_reconstruction.layers.projection_2d import parallel_projection2d
from pyronn.ct_reconstruction.layers.backprojection_2d import parallel_backprojection2d
from pyronn.ct_reconstruction.helpers.misc import generate_sinogram as sino_helper
from pyronn.ct_reconstruction.helpers.misc import generate_reco as reco_helper
from pyronn.ct_reconstruction.helpers.phantoms import shepp_logan
phantom = shepp_logan.shepp_logan_enhanced(par_geometry.volume_shape)
with tf.Session as sess:
sinogram = sino_helper.generate_sinogram(phantom, parallel_projection2d, par_geometry)
reconstruction = reco_helper.generate_reco(sinogram, parallel_backprojection2d, par_geometry)
```

In the following the example using the Layers directly is shown (Note that the Layers are within the Tensorflow graph context and therefore need to be evaluated before the result can be accessed):

```
from pyronn.ct_reconstruction.layers.projection_2d import parallel_projection2d
from pyronn.ct_reconstruction.helpers.phantoms import shepp_logan
phantom = shepp_logan.shepp_logan_enhanced(par_geometry.volume_shape)
with tf.Session as sess:
result = parallel_projection2d(phantom, par_geometry)
sinogram = result.eval()
```

Using the PYRO-NN Layers directly registers the respective gradient, thus they can be used as normal Tensorflow Layers within the graph. For more details checkout the examples which are covering the different geometry and application cases.

## Potential Challenges

Memory consumption on the graphics card can be a problem with CT datasets. For the reconstruction operators the input data is passed via a Tensorflow tensor, which is already allocated on the graphicscard by Tensorflow itself. In fact without any manual configuration Tensorflow will allocate most of the graphics card memory and handle the memory management internally. This leads to the problem that CUDA malloc calls in the operators itself will allocate memory outside of the Tensorflow context, which can easily lead to out of memory errors, although the memory is not full.

There exist two ways of dealing with this problem:

1. A convenient way is to reduce the initially allocated memory by Tensorflow itself and allow a memory growth. We suggest to always use this mechanism to minimize the occurrence of out of memory errors:

```
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.gpu_options.allow_growth = True
# ------------------ Call Layers ------------------
with tf.Session(config=config) as sess:
...
```

2. The memory consuming operators like 3D cone-beam projection and back-projection have a so called hardware_interp flag. This means that the interpolation for both operators are either done by the CUDA texture or based on software interpolation. To use the CUDA texture, and thus have a fast hardware_interpolation, the input data need to be copied into a new CUDA array, thus consuming the double amount of memory. In the case of large data or deeper networks it could be favorable to switch to the software interpolation mode. In this case the actual Tensorflow pointer can directly be used in the kernel without any duplication of the data. The downside is that the interpolation takes nearly 10 times longer.

Note that the hardware interpolation is the default setup for all operators.