This package contains code accompanying the manuscript of "On-the-fly machine learning for improving image resolution in tomography".

• Free software: GNU General Public License v3
• Documentation: [https://ahendriksen.github.io/on_the_fly]

It takes a few steps to setup On the fly machine learning for improving image resolution in tomography on your machine. We recommend installing Anaconda package manager for Python 3.

To install and execute the code in this package, conda on 64-bit Linux is required. Moreover, a CUDA 9.0-compatible graphics card and runtime is required.

To install this package, use conda and clone this GitHub project. To install the package into a new conda environment named otf, execute the following in the terminal:

conda create -n otf python=3.6
source activate otf
conda install -c astra-toolbox/label/dev  -c aahendriksen -c pytorch -c conda-forge -c owlas \
        msd_pytorch=0.5.1 \
        cudatoolkit=9.0 \
        flexdata \
        tomopy \
        dxchange \
        astra-toolbox=1.9.0.dev10 \
        cone_balls=0.2.2
conda install h5py ipython matplotlib numexpr pyopengl pyqtgraph scikit-image tqdm

# Now install the package using pip
git clone https://github.com/ahendriksen/on_the_fly.git
cd on_the_fly
pip install .

To learn more about the functionality of the package check out our examples folder.

The examples folder has the following hierarchy:

examples
├── cone_foam_fast
│   ├── cone_balls_spec.txt
│   ├── geometries
│   └── Makefile
├── cone_foam_full
│   ├── cone_balls_spec.txt
│   ├── geometries
│   └── Makefile
└── cone_foam_just_roi
    ├── cone_balls_spec.txt
    ├── geometries
    └── Makefile

The examples directory contains three directories with a Makefile. The cone_foam_full directory contains the specification of the data as it is used in the paper. Because generating each projection dataset can take 2 hours with a recent GPU, I have created cone_foam_just_roi where all voids have been removed that do not intersect the upper or central region of interest, and cone_foam_fast in which 90% of the voids have been removed. Generating the projections of cone_foam_fast should take roughly 10 minutes per projection dataset on a modern GPU.

Each directory contains a Makefile that can be used to generate

• Projection data
• Reconstructions
• Training and test sets

Moreover, the makefile contains instructions to train the networks and to process the input of the test set with the trained neural networks.

To generate projection data of the entire foam ball, a zoomed-in central region of interest, and a zoomed-in upper region of interest, run:

make data/zoom1/.dirstamp
make data/zoom4_centre/.dirstamp
make data/zoom4_top/.dirstamp

To reconstruct the entire volume and the regions of interest, run the following. All intermediate projection data will be saved in the processing directory.

make reconstruction/zoom/.dirstamp
make reconstruction/zoom4_centre/.dirstamp
make reconstruction/zoom4_top/.dirstamp

The following commands will generate the training and test set.

make training_set
make test_set

To train a neural network, execute either one of

make weights/unet-B1.torch
make weights/unet-A1.torch
make weights/unet-A9.torch
make weights/msd-B1.torch
make weights/msd-A1.torch
make weights/msd-A9.torch

To train for more than one epoch, make sure to edit the variable named EPOCH at the top of the Makefile.

To see how a neural network performs on the test set, execute either one of

make test/output-unet-B/.dirstamp
make test/output-unet-A1/.dirstamp
make test/output-unet-A9/.dirstamp
make test/output-msd-B/.dirstamp
make test/output-msd-A1/.dirstamp
make test/output-msd-A9/.dirstamp
make test/output-bicubic/.dirstamp

Another option is to bicubically upsample the whole volume. This is performed on the last line.

• Allard Hendriksen - Initial work

See also the list of contributors who participated in this project.

• We thank milesial for making available a high-quality pytorch version of the U-net network architecture.

If you have any issues, questions, or remarks, then please open an issue on GitHub.

This project is licensed under the GNU General Public License v3 - see the LICENSE.md file for details.