Code for Helsinki Tomography Challenge 2022

Authors, institution, location

Below is the information for the authors.

• Author Chao Wang and Ji Li
• Institution Department of Statistics & Data Science and Department of Mathematics, National University of Singapore
• Location 21 Lower Kent Ridge Rd, Singapore 119077

---

Brief description of your algorithm and a mention of the competition

Our reconstruction algorithm is a deep learning approach, as there have been provided the training datasets. Our deblurring network backbone is a U-net. We use the supervised learning.

The given dataset is very limited. To address such issue, we proposed the following data augmentations:

• To increase the dataset scale, we propose using the subsets of the full sinogram to construct several dataset.
• To mitigate the learning difficulty, we plug the FBP recovery (using astra-toolbox) as the input, as well we observe that the circle is generally destroyed.
• We test the unrolling network, it does not work for the limited dataset.

This code repository is uploaded for competition of Helsinki Tomography Challenge 2022.

Installation instructions, including any requirements

See the requirement.txt to install the dependent packages and libraries.

• Clone the github repository

  git clone https://github.com/chaow-mat/HTC2022.git
  cd HTC2022

• Use virtualenv to construct the virtual environment

  pip3 install virtualenv
  virtualenv --no-site-packages --python=python3 htc2022
  source htc2022/bin/activate # enter the environment
  # deactivate

• Install the Cython module

  pip3 install Cython
  

• Install Astra-toolbox

  git clone https://github.com/astra-toolbox/astra-toolbox.git
  cd astra-toolbox/build/linux
  bash ./autogen.sh   # when building a git version
  bash ./configure --with-cuda=/usr/local/cuda \
              --with-python \
              --with-install-type=module
  make
  make install
  cd ~current working directory

• Install requirements

  pip3 install -r requirements.txt # install the dependency 
  

• (Automatically download the trained models) In our code, we use the gdown to obtain the training checkpoints in google driver. The following links is used in our evaluations.

  • Files for seven models are readable by anyone at Here
  • If downloading fails, please download them from the above links and put them in the folder checkpoints

Usage instructions

• For recovering the binary segmented image from sinogram data, using the following command line in terminal

  CUDA_VISIBLE_DEVICES=0 python3 main.py /path/to/input/folder/ /path/to/output/folder/ groupNbr

Here groupNbr is the difficulty level of the recovery task, its values are the integers from 1 to 7, 1 means that there are 90-degree projected sinogram data, 7 means that there are only 30-degree projected sinogram data