Convergent Data-driven Regularizations for CT Reconstruction

Samira Kabri¹, Alexander Auras², Danilo Riccio³, Martin Benning^3,4, Michael Moeller² and Martin Burger¹

¹Department of Mathematics, University of Erlangen-Nuremberg
²Institute for Vision and Graphics, University of Siegen
³School of Mathematical Sciences, Queen Mary University of London
⁴The Alan Turing Institute, British Library

This repository contains the official pytorch implementation of Convergent Data-driven Regularizations for CT Reconstruction.

Features

• Calculation of $\Pi$, $\Delta$ and $\Gamma$
• Calculation of the analytic coefficients $g_n$
• Learning of the coefficients $\overline{g_n}$
• Calculation of the analytic filter $\rho$
• Learning of the filter $\overline{\rho}$
• Logging of results via tensorboard
• Saving of results in .pt files

Dependencies

Dependencies to execute the calculations and the used versions:

• torch (1.12.1)
• hydra-core (1.2.0)
• pytorch-lightning (1.7.1)
• torchmetrics (0.9.3)
• radon (1.0.0) (either as submodule or from the repository at https://github.com/AlexanderAuras/radon)

Optional dependencies for the visualizations:

• notebook
• matplotlib
• tensorflow

Dependencies to view tensorboard results:

• tensorboard

Installation

• Install conda and create environment (optional)
• Install dependencies
• Clone repository: git clone https://github.com/AlexanderAuras/LearnedRadonFilters
• Adjust the output directory in configs/default.yaml and the paths in mnist_datamodule.py

Usage

• Adjust the configurations in the configs folder
  • Adjust configurations for datasets in configs/dataset/<dataset>.yaml
  • Adjust configurations for the models in configs/model/<model>.yaml
  • Adjust general configurations in configs/default.yaml
• Execute the training/calculations with python main.py hydra.job.name=<name>
• View tensorboard results via tensorboard --logdir=<output_directory>
• View saved .pt files in the jupyter notebooks plots.ipynb, test.ipynb or visualization.ipynb
  • For every run $\Pi$, $\Delta$ and $\Gamma$ are saved in pi.pt, delta.pt and gamma.pt
  • For FFT-runs, coefficients.pt contains $\rho$ or $\overline{\rho}$
  • For SVD-runs, coefficients.pt contains $g_n$ or $\overline{g_n}$.
  • For SVD-runs the singular values are saved in singular_values.pt