- The presented code in this repository is the implementation of Graph Neural Networks for modelling breast biomechanical compression, available at: https://doi.org/10.48550/arXiv.2411.06596
- Please cite as:
-
@article{awwad2024, title={Graph Neural Networks for modelling breast biomechanical compression}, author={Hadeel Awwad and Eloy García and Robert Martí}, journal={arXiv preprint}, year={2024}, doi={arXiv:2411.06596}, url={https://arxiv.org/abs/2411.06596} }
- The breast geometry used in this study, 'UncompressedBreast3,' was obtained from a publicly available dataset on Zenodo, available at: https://zenodo.org/records/4529852
- The mesh generation tool used is Pygalmesh implemented by Schlömer et al, available at: https://github.com/meshpro/pygalmesh
- The FEA-baseline using NiftySim and Phantom Reconstruction was adopted from the work of García et al, available at: https://github.com/eloygarcia/niftysim & https://github.com/eloygarcia/RadboudCompression
- The PhysGNN implementation used in this work was originally developed by Salehi et al, available at: https://github.com/YasminSalehi/PhysGNN
The following is the required folder structure for this project. If any of the folders are missing, please create them:
GNNs-BreastCompression/
├── Data_Generator/
├── Environment/
├── FEA-simulations/
├── Hold-out/
├── LODO/
├── process-compressed/
├── uncompressed_nrrd/
├── mesh_data.ipynb
├── niftysim.ipynb
├── preprocessing.ipynb
└── qualitative.ipynb
To set up the environment, please refer to the instructions in the Environment folder.
To build NiftySim Docker, please follow the instructions:
git clone https://github.com/eloygarcia/niftysim.git
cd niftysim
docker build -t niftysim:2.5 .
To build Reconstruct Phantom Docker, please follow the instructions:
git clone https://github.com/eloygarcia/RadboudCompression.git
cd "RadboudCompression/Phantom Reconstruction/Reconstruct Image"
docker build -t reconstruct-image .
Run the following command with the right paths, in the public dataset there is a metadata CSV file that will be needed:
The isotropic spacing that was used is 0.273
python GNNs-BreastCompression/preprocessing.py <dicom_folder> <csv_file> <output_folder> <isotropic_spacing>
Make sure to update the path as prompted in the notebook niftysim.ipynb. This notebook is for mesh generation, runs NiftySim simulations, and reconstructs the phantom image. Incremental simulations will be achieved by adjusting the thickness and offset of the plates with each run and manually saving the output displacements from the Niftysim Docker output.
Run mesh_data.ipynb to extract the data from the uncompressed mesh, and generate the random force directions
For Hold-out experiment, please refer to the instructions in the Hold-out folder. For Leave-one-deforamtion-out experiment, please refer to the instructions in the LODO folder.
To produce qualitative results of LODO experiment, run qualitative.ipynb and make sure to update the path as prompted for reconstructing the compressed phantom.