Helsinki Tomography Challenge 2022 - Technical University Dortmund & Heinrich Heine University Düsseldorf
Thomas Germer1, Jan Robine2, Sebastian Konietzny2, Stefan Harmeling2, Tobias Uelwer2
1 Department of Computer Science, Heinrich Heine University Düsseldorf, Germany
2 Department of Computer Science, Technical University Dortmund, Germany
This is a submission to the Helsinki Tomography Challenge 2022.
The goal of the challenge is to compute binary segmentations from limited angle sinograms, consisting of two-dimensional X-ray projections of an acrylic disk.
A dataset is available online:
Since the dataset is quite small, we generate synthetic data to train a neural network. First, we generate a large number of synthetic images by placing various shapes on top of a disk. We then compute sinograms from those images using the forward operator from the Helsinki Tomography Toolbox and Astra Toolbox. Lastly, we train a neural network to directly predict the images from the subsampled sinograms.
To install:
pip3 install -r requirements.txt
Use the following command to generate images from subsampled sinograms in the directory data/limited.
The images will be written to the directory output.
python3 main.py data/limited output
The model has been stored with git lfs and should be cloned automatically. Alternatively, if the bandwidth limit of git lfs has been exceeded, you can download the model with:
wget https://asdf10.com/model3.pth -O model.pth
It's SHA256 checksum is b41b5503deef2dc513923eb6b5ced4375e370d5de70790e7af71be36d9b02a0d.
You can then run the following command to compute the mean score on the subsampled dataset.
python3 main_compute_score.py output
| Angles | Limited sinogram | Prediction | Ground truth |
|---|---|---|---|
| 30° |  |
 |
 |
| 30° |  |
 |
 |
| 30° |  |
 |
 |
| 30° |  |
 |
 |
| 90° |  |
 |
|
| 90°, start 240° |  |
 |
|