PhysGNN

• The presented code in this repository is the implementation of PhysGNN, available at: https://arxiv.org/pdf/2109.04352.pdf

• Please cite as:

@article{salehi2021physgnn,
  title={PhysGNN: A Physics-Driven Graph Neural Network Based Model for Predicting Soft Tissue Deformation in Image-Guided Neurosurgery},\
  author={Salehi, Yasmin and Giannacopoulos, Dennis},\
  journal={arXiv preprint arXiv:2109.04352},\ 
  year={2021}
}

• The code used for generating the datasets used in the paper is also included.

• The tumour and brain model are taken from the paper: "A machine learning approach for real-time modelling of tissue deformation in image-guided neurosurgery" by Tonutti et al, available at: https://github.com/michetonu/MALTIDEM--Machine-Learning-for-Tissue-Deformation-Modelling

• The Graph_load_batch function in pg_dataset.py has been adapted from the Graph_load_batch function in dataset.py of Position-aware Graph Neural Networks code by You et al. available at: https://github.com/JiaxuanYou/P-GNN

Data Generation

1. Please download:

   1. FEBio 2.9.1 @ https://febio.org/febio/febio-downloads/

   2. export_fig-master @ https://www.mathworks.com/matlabcentral/fileexchange/23629-export_fig

   3. GIBBON @ https://www.gibboncode.org/Installation/

   and place all the downloaded files in the "code" folder.

2. Please make the following installations:

   1. pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.12.0+cu113.html
   2. pip install tensorboardX
   3. pip install networkx

3. Please run data_generator_1.m to generate Dataset 1, and data_generator_2.m to generate Dataset 2. Formatted data will be saved in dataset_1 and dataset_2 folders.

4. Run dataset_full.py to generate the preprocessed data.

5. Run pg_dataset.py to create 1/11 of dataset1/dataset2 and pickle them. Each run requires 22GB of RAM. Dataset 1 and Dataset 2 use up 16.79 GB of RAM each.

PhysGNN Training

1. Select the desired configuration from training in main.py and simply run the code.

Reproducing Results

1. Select the desired configuration from reproduce.py and simply run the code.

To run the code on Colab

(Requirement: A Google Colab Pro Account)

Training PhysGNN on Colab

1. Upload dataset_1 and dataset_2 folders to your Google Drive.
2. Upload Dataset_Generation.ipynb to your Google Drive.
3. Set up the paths on your Colab notebook.
4. Run all the cells consecutively. Ensure proper selection of the dataset to be configured as a pytorch geometric dataset and its corresponding pickle name.
5. After one pytorch geometric dataset is created, restart the runtime to clear the RAM.
6. Each section of Dataset 1 and Dataset 2 takes 1 hour and 15 minutes to be generated.
7. After successful generation of the datasets, upload PhysGNN.ipynb to your Google Drive.
8. Run all the cells in a consecutive order.
   1. In the cell "Final Run for Training" select the dataset you wish to use
   2. In the "Final Run" section select the configuration you want to train.

To reproduce the results on Colab

1. Run all the cells until "Final Run for Training". Run the "Reproducing the Results" cell instead.
2. Select the Dataset and the model you wish to reproduce their results in the "Final Run" under the Reproducibility cell.
3. Results generated from the pickled configurations can be saved by setting the save parameter to 1.
4. Setting mean_mag_results to 1 generates the mean Euclidean errors reported (Table 4).
5. Setting max_error_results to 1 generates the max error results reported (Table 5).