Project content & structure • Getting started • Usage • Citation

This repository contains the implementation of k-Strip, a complex valued convolutional neural network (CVNN) based method for automated skull stripping directly in the k-space, the raw data domain of MRI scans.

This work is part of the k-Radiomics project. For more information, please visit our project-website https://k-radiomics.ikim.nrw.

Project content & structure

This repository provides the architecture of a complex valued neural network in the style of a residual UNet.

We also provide multiple complex valued blocks and layer, as well as loss and activation functions, which can all be found in the folder src/models/layer.

Project folder structure:

├── logs                  # Location of logs created during training / testing
├── output                # Location of output created during training / testing
│   ├── training  
│   └── testing  
└── src  
    ├── configs           # Config files for training / tesing  
    ├── models            # Location of the network architecture
    │   └── layer         # Location of all building blocks needed for the network
    └── utils             # Utility functions, including dataloader

Implemented layers & functions

Toggle the lists below to see an overview of implemented layer, blocks and activation functions. For more details, have a look at the corresponding python files in the src/models/layer folder.

Complex layer & blocks

• cBatchNorm
• cConvolution
• cDoubleConvolution
• cResidualBlock
• cTransposedConvolution
• Spectral Pooling
• cDropout
• cUpsampling

Complex activation functions

• cReLU
• clReLU
• cELU
• cPReLU
• cSELU
• cTanh
• cSigmoid
• PhaseAmplitudeReLU
• PhaseReLU
• Cardioid
• AmplitudeReLU
• cLogReLU

Getting started

1. Clone repository:

   git clone https://github.com/TIO-IKIM/k-Strip.git
   

   Note: The training & testing scripts expect the repository to be saved directly in the home directory (Path.home()). If this is not the case, you just need to insert the correct path in the scripts.

2. Create a conda environment with Python version 3.11.2 and install the necessary dependencies:

   conda create --name k-strip python=3.11.2 --file requirements.txt
   

   In case of installation issues with conda, use pip install -r requirements.txt to install the dependecies.

   Note: The PyTorch version used needs CuDA 11.7 installed. You can change the PyTorch version in the requirements file.

3. Activate your new environment:

   conda activate k-strip
   

4. Insert all necesarry paths to your training / testing data in the config files in src/configs and you should be all set to start training your own CVNN!

   Note: The training config is already filled with the settings used for our publication, but you can change all parameters as you wish ... until you break it.

Usage

The repository includes an exemple for preparing the CC-359 for training the network by transforming it into fourier transformed 2D .pt slices. You can find the script at src/utils/prepare_data.py. Just download the dataset and adjust the paths in the script accordingly.

Training CLI

usage: Training [-h] [--e E] [--log LOG] [--tqdm] [--gpu GPU] [--config CONFIG] [--load LOAD]
                [--load_checkpoint LOAD_CHECKPOINT]

Train a CVNN.

options:
  -h, --help            show this help message and exit
  --e E                 Number of epochs for training
  --log LOG             Define debug level. Defaults to INFO.
  --tqdm                If set, do not log training loss via tqdm.
  --gpu GPU             GPU used for training.
  --config CONFIG       Path to configuration file
  --load LOAD           Load model by giving path to saved model. Defaults to None.
  --load_checkpoint LOAD_CHECKPOINT
                        Load checkpoint by giving path to checkpoint. Defaults to None.

Checkpoints of the model and examples of the validation results are saved every 10 epochs. The frequency can be changed in the script in if self.epoch % 10 == 0.

Testing CLI

usage: Tester [-h] [--config CONFIG] [--e] [--perc PERC] [--num_zero NUM_ZERO] [--eval_num EVAL_NUM]
              [--id ID] [--c C] [--device DEVICE]

Test a trained network.

options:
  -h, --help           show this help message and exit
  --config CONFIG      Path to configuration file
  --e                  Start inference. If selected, start inference. Defaults to True.
  --perc PERC          If selected choose single percentage value for binary masking in evaluation. If
                       None, iterate through range of values in case of evaluation. Defaults to None
  --num_zero NUM_ZERO  Choose one number of minmal amount non-zero pixels in segmentation masks to be taken
                       into account in evaluation. Use None to iterate through range of values in case of
                       evaluation. Defaults to None
  --eval_num EVAL_NUM  Define if not all samples are supposed to be used for evaluation. Defaults to None.
  --c C                Checkpoint number to be loaded. For this the path to the checkpoints need to be
                       defined in the config file. Defaults to None.
  --device DEVICE      Cuda device to run test on.

Datastructure

The network expects a certain datastructure which is shown below.

├── .../train  
│       ├── orig
│       └── label  
├── .../val  
│       ├── orig
│       └── label  
└── .../test  
        ├── orig
        └── label  

Citation

If you use our code in your work, please cite us with the following BibTeX entry.

@article{rempe2023k,
  title={k-strip: A novel segmentation algorithm in k-space for the application of skull stripping},
  author={Rempe, Moritz and Mentzel, Florian and Pomykala, Kelsey L and Haubold, Johannes and Nensa, Felix and Kr{\"o}ninger, Kevin and Egger, Jan and Kleesiek, Jens},
  journal={Computer Methods and Programs in Biomedicine},
  pages={107912},
  year={2023},
  publisher={Elsevier}
}