mathimag / data_driven_convex_regularization

A PyTorch implementation of the data-driven convex regularization approach for inverse problems

data_driven_convex_regularization

This repo contains python scripts for implementating data-driven convex regularization for inverse problems (sparse-view CT reconstruction, in particular). For a detailed description of the algorithm and theoretical results, see: https://arxiv.org/abs/2008.02839.

If you use these scripts in your research, consider citing the paper:

@misc{mukherjee2021learned,
      title={Learned convex regularizers for inverse problems}, 
      author={Subhadip Mukherjee and Sören Dittmer and Zakhar Shumaylov and Sebastian Lunz and Ozan Öktem and Carola-Bibiane Schönlieb},
      year={2021},
      eprint={2008.02839},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Steps to run the scripts:

• The phantoms used in our CT experiments are available (as .npy files) here: https://drive.google.com/drive/folders/1SHN-yti3MgLmmW_l0agZRzMVtp0kx6dD?usp=sharing. Download the .zip file containing the phantoms, unzip, and put inside the cloned directory.
• Create a conda environment with the required dependencies by conda env create -f environment.yml, and then activate it by conda activate env_deep_learning.
• Run python simulate_projections_for_train_and_test.py to simulate the projection data and the FBP solutions.
• Train a convex regularizer by python train_convex_reg.py.
• Evaluate the model on test slices by running python eval_convex_reg.py.
• If you want to test the model for a different acquisition geometry, appropriately modify the acquisition parameters in simulate_projections_for_train_and_test.py.