- This repository regularly exceeds GitHub's (very small) bandwidth quota.
- If possible, please download from one of the links below instead of cloning the repository.
- UPDATE: We are still regularly over quota. Therefore, the file "model.pth" and the files "data/limited/*.mat" have been removed from this repository. To download the full repository, use one of the links below instead.

• Download ➔ https://uni-duesseldorf.sciebo.de/s/examkbpuy4Jj8lE
• Mirror ➔ https://dbs.cs.uni-duesseldorf.de/research/HTC/HTC2022-TUD-HHU-version-2024-01-08.zip
• Mirror ➔ https://drive.google.com/file/d/1mfjbBdKHEyK7FvjpZnBGyI7vOw-MMsFJ

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:

• https://www.fips.fi/HTCdata.php
• https://zenodo.org/record/6984868

Method Description

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.

Installation instructions

To install:

pip3 install -r requirements.txt

We were able to install all dependencies with both Python 3.8.10, 3.9.2 and 3.10.6 on Ubuntu 20.04 and 22.04, but other versions and similar operating systems might work as well.

Usage instructions

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/model1.pth -O model.pth

It's SHA256 checksum is ac43bf074b2756f07b995284c50ccda9a2f1d17d8be38ea0e84609501bc5230a.

You can then run the following command to compute the mean score on the subsampled dataset.

python3 main_compute_score.py output

Examples

Angles	Limited sinogram	Prediction	Ground truth
30°
30°
30°
30°
90°
90°, start 240°