No Stripes Net

A neural network to remove stripe artifacts in sinograms.
The network takes a sinogram with stripes as input, and learns to remove those stripes from it.
The network can train on both synthetic and real-life data.

Requirements

• A Linux machine with a GPU and CUDA
• Conda
• Python 3.9+
• PyTorch
• TomoPy

For a full list of requirements, see environment.yml

Installation

First, set up the project environment:

• Clone the repository: git clone https://github.com/dkazanc/NoStripesNet.git
• Create the conda environment: conda env create -f environment.yml
• Activate the conda environment: conda activate nostripesnet

The Repository

• network/ contains Python code to train and test a model, as well as the dataset and visualiser classes.
• run_scripts/ contains bash scripts to generate masks & datasets and train/test models.
• simulator contains Python code to generate masks & datasets.
• utils/ - contains utility functions used throughout the codebase.
• TUTORIAL.md is a walkthrough of how to generate a dataset, and train & apply a model.
• apply_model.py is a program that applies a model to a given tomographic scan.
• graphs.ipynb is a Jupyter Notebook used to create the graphs in the paper.
• residuals.ipynb is a Jupyter Notebook used to create the residual images in the paper.
• rmse.ipynb is a Jupyter Notebook used to calculate the RMSEs in the paper.
• submit.sh is a bash script to train a model on multiple nodes, using multiple GPUs on each.
• visualize_results.ipynb is a Jupyter Notebook used to visualize the results of a model.

Running the Code

A full walkthrough of how to generate a dataset and train a model can be found here.
To apply a trained model to a tomographic scan, see run_scripts/apply_model.sh.