Brain Tumour Segmentation (BraTS) on the Basis of a
U-Net architecture

How to run our code

Set up the project

0. Ensure required packages, listed in Code/requirements.txt, are installed on your system ( > pip install -r requirements.txt )
1. Download "Task01_BrainTumour.tar" under the Medical Segmentation Decathlon website and extract the tar file.
2. Provide the following folder structure:

BT_Segmentation_Project
    |_ README.md
    |_ Code
       |_ Data_Extraction.ipynb     (this file)
       |_ ...
    |_ Task01_BrainTumour           (Dataset)
       |_ dataset.json
       |_ imagesTr                  (here are the compressed nii-files)
       |_ labelsTr
       |_ imagesTs
    |_ Weights                      (trained model files, see section "Useful links")

3. Change into "Code" as a working directory
4. Run "python3 setup_project.py" in your shell, to extract the nii-files to numpy file (Attention!: bear in mind that the dataset needs 70+ GB of storage)

Additional ressources for our code

High-level code overview (Click to expand)

Code Overview

Notebooks

• demo.ipynb:
  Here we present our model training along with evaluation metrics and animated visualizations. The majority of what we report in our paper is produced using this code.

Python-files

• custom_losses.py:
  Contains custom programmed losses. In addition you can find the get_loss function to automatically calculate the loss for a provided loss-function.

• data_loading.py:
  This is the dataloader. The function get_train_test_iters gives you the data iterators.

• dataset_utils.py:
  Here we coded all functions for manipulating and processing data.

• train.py:
  Here you can find the train_model function to train a model.

Folders

• Optional:
  Here we stored additional experimental code that was not ultimately used in our trained models or demo.

• Architectures:
  Here you can find all model architectures that we built.

File-structure figure (Click to expand)

-

Useful links

Our executive report
Github repo
Model files

Acknowledgements

We would like to thank Prof. Dr. Chistoph Lippert, Tahir Miriyev, Noel Danz, Eshant English and Wei-Cheng Lai for sharing their experience and expertise with us. Special thanks to Noel Danz who gave insightful tips and comments to help guide the project.

Used Github-repos

• Inspiration for our U-Net implementation:
  https://github.com/Hsankesara/DeepResearch/blob/master/UNet/Unet.py
  

• Basis for adapted fuctions/classes, such as dice loss and focaltversky loss:
  https://github.com/Mr-TalhaIlyas/Loss-Functions-Package-Tensorflow-Keras-PyTorch
  

• Reference was made to pytorch documentation throught the coding process:
  https://pytorch.org/docs/stable/index.html
  

• Basis for the 2D Unet:
  https://github.com/mateuszbuda/brain-segmentation-pytorch