Hands-on Session 4: Image Reconstruction using the PyTorch and SPyRiT Packages

This code was used during a hands-on session given at the Deep Learning for Medical Imaging School 2023.

The session was a practical introduction to image reconstruction, considering the limited-angle computed tomography problem. Participants were invited to run the cells, answer the questions, and fill in blanks in the code of main.ipynb. All answers and the solution code are given in main_with_answers.ipynb

The hands-on session followed a lecture on the topic.

• Slides
• Video

Authors: (version 2023) L Amador, E Chen, N Ducros, H-J Ling, K Mom, J Puig, T Grenier, E Saillard,
Authors: (version 2021): N Ducros, T Leuliet, A Lorente Mur, Louise Friot-Giroux

Contact: [email protected], CREATIS Laboratory, University of Lyon, France.

Install the dependencies

We recommend using a virtual (e.g., conda) environment.

conda create --name new-env
conda activate new-env

Our notebook primarily relies on the SPyRiT package that can be installed via pip. The notebook was tested with version 2.1.

conda install pip
pip install spyrit==2.1
pip install ipykernel
pip install scikit-image
pip install h5py

Get the scripts and data

1. Get source code from GitHub

    git clone https://github.com/openspyrit/spyrit-examples.git        
   

2. Go into spyrit-examples/2023_DLMIS/

   cd spyrit-examples/2023_DLMIS/    
   

3. Download the image database at this url and extract its content

   • Windows PowerShell

   wget https://www.creatis.insa-lyon.fr/~ducros/spyritexamples/2023_DLMIS/data.zip -outfile data.zip
   tar xvf data.zip 

   The directory structure should be

    |---spyrit-examples
    |   |---2023_DLMIS
    |   |   |---data
    |   |   |   |---
    |   |   |---main.ipynb
    |   |   |---main_with_answers.ipynb
    |   |   |---train.py
   

At this this point you ready to go!

Try to complete main.ipynb or run main_with_answers.ipynb

To go further

We provide train.py to train a network from a single command line

python train.py

By default, all networks are trained for 60 view angles during 20 epochs. For other values (e.g., 40 angles and 100 epochs), consider

python train.py --angle_nb 40 --num_epochs 100

To specify training parameters such as the batch size or learning rate, and for other options, type python train.py --help