How to increase the model accuracy and reproducibility in terms of top discriminative features when learning from heterogeneous decentralized datasets? Please contact ([email protected]) for inquiries. Thanks.
This work is accepted at the PRIME workshop in MICCAI 2022.
Investigating the Predictive Reproducibility of Federated Graph Neural Networks using Medical Datasets
Mehmet Yiğit Balık1, Arwa Rekik2, Islem Rekik1
1BASIRA Lab, Faculty of Computer and Informatics, Istanbul Technical University, Istanbul, Turkey
2Faculty of Medicine of Sousse, Sousse, Tunisia
Abstract: Graph neural networks (GNNs) have achieved extraordinary enhancements in various areas including the fields medical imaging and network neuroscience where they displayed a high accuracy in diagnosing challenging neurological disorders such as autism. In the face of medical data scarcity and high-privacy, training such data-hungry models remains challenging. Federated learning brings an efficient solution to this issue by allowing to train models on multiple datasets, collected independently by different hospitals, in fully data-preserving manner. Although both state-of-the-art GNNs and federated learning techniques focus on boosting classification accuracy, they overlook a critical unsolved problem: investigating the reproducibility of the most discriminative biomarkers (i.e., features) selected by the GNN models within a federated learning paradigm. Quantifying the reproducibility of a predictive medical model against perturbations of training and testing data distributions presents one of the biggest hurdles to overcome in developing translational clinical applications. To the best of our knowledge, this presents the first work investigating the reproducibility of federated GNN models with application to classifying medical imaging and brain connectivity datasets. We evaluated our framework using various GNN models trained on medical imaging and connectomic datasets. More importantly, we showed that federated learning boosts both the accuracy and reproducibility of GNN models in such medical learning tasks.
The codewas implemented using Python 3.8 (Anaconda), PyTorch 1.7.1 on Windows and Linux. GPU is not required to run the code.
- Go to https://repo.anaconda.com/archive/
- Download version for your system (We used Python 3.8 on 64bit Windows 10)
- Install the platform
- Create a conda environment by typing:
conda create --name reproducibleFedGNN python=3.8
Copy and paste the following commands to install all packages (CPU version)
$ conda activate reproducibleFedGNN
$ pip install numpy
$ pip install matplotlib
$ pip install pandas
$ pip install scikit-learn
$ pip install seaborn
$ pip install torch==1.7.1+cpu torchvision==0.8.2+cpu torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
$ pip install torch-scatter==2.0.6 -f https://data.pyg.org/whl/torch-1.7.1+cpu.html
$ pip install torch-sparse==0.6.9 -f https://data.pyg.org/whl/torch-1.7.1+cpu.html
$ pip install torch_geometric==1.7.1
$ pip install medmnistThese instructions are for CPU installation. If you want GPU installation, please visit (optional) PyTorch-Geometric’s web page (https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html) for description on installing GPU version. Code will check the version of dependencies and availability of GPU. If everything is configured correctly, it will utilize GPU automatically.
In case you want to use our framework, the input dataset should be a list of numpy arrays. Each numpy array is of size (n_r, n_r, n_v), where n_r and n_v are the number of regions and views, respectively. Our framework works with only one view. If you have multi-view graph please indicate the view by command line argument --view=VIEW_INDEX. We provided a data simulation coda in the python file simulate_data.py. If you want to use your dataset, you should put your data into the data folder and set command line argument --dataset=Folder_Name_of_Your_Dataset.
Usage of image based datasets (MedMNIST)
The image should be square and grayscale to satisfy (n, n, 1) format where n is number of rows and columns.
784x784 connectivity matrix (8-connected) is built and given as inputs to models. The weight of the connectivity link is determined by the absolute differences of two adjacent pixels. Heatmap of pixel weights can be plotted if the input type is the connectivity matrix.
The input type can be changed via options (for image based datasets only):
--input_type=imageto input as image matrix--input_type=adjto input as weighted connectivity matrix
After creating a simulated dataset, you can run the GNN models by running demo.py file in your activated reproducibleFedGNN environment.
$ conda activate reproducibleFedGNN
$ python demo.pyThe GNN models included are:
| Model | Paper |
|---|---|
| DiffPool | https://arxiv.org/abs/1806.08804 |
| GCN | https://arxiv.org/abs/1609.02907 |
| Component | Content |
|---|---|
| options.py | Includes hyperparameter and other options. You may modify it according to your needs. |
| demo.py | The main driver code. |
| fed_localmodel.py | Contains the implementation of the local model and federated averaging. |
| learning_modes.py | Contains the implementation of both federated and non-federated (baseline) learning. |
| fed_accuracy.py | It creates the plots of losses and accuracy comparison concerning one specific model. |
| fed_reproducibility.py | It creates the heatmaps of reproducibility matrices. |
| data/ | Contains dataset folders. |
| results/ | Contains plots after execution. |
results/ directory automatically created by demo.py
| Argument | Explanation |
|---|---|
| dataset | Name of the dataset (default: 'Demo') |
| input_type | Input type option of image data (options: ['image', 'adj'], default: 'image') |
| view | Index of view for multi-view graphs (default: 0) |
| comms_round | Number of communication rounds between local models and the global model (default: 5) |
| hospital_num | Number of hospitals (default: 3) |
| num_epochs | Number of epochs (default: 100) |
The figure demonstrates an example accuracy comparisons and average reproducibility matrices of federated method to baseline method for the models (DiffPool and GCN) for a demo data with 200 subjects (equally partitioned between two classes) where each subject represented by a 35x35 matrix.
The figure shows the learned weights of the features by the most reproducible model.
To install and run federeated reproducibility, check the following YouTube video: https://www.youtube.com/watch?v=bG54z0v75U0&list=PLug43ldmRSo2p3vVV5K9I4yWZuyXQ0_jL&index=3&ab_channel=BASIRALab
To learn about how federated reproducibility works, check the following YouTube video: https://www.youtube.com/watch?v=pnattjt981k&list=PLug43ldmRSo0bX8cOSuMWinXs-xem9q4o&index=4&ab_channel=BASIRALab
https://arxiv.org/abs/2209.06032
@inproceedings{Balik2022,
title={Investigating the Predictive Reproducibility of Federated Graph Neural Networks using Medical Datasets},
author={Balik, Mehmet Yigit and Rekik, Arwa and Rekik, Islem},
booktitle={International Workshop on PRedictive Intelligence In MEdicine},
year={2022},
organization={Springer}
}- Federated learning technique is built on RG-Select framework.
Georges, N., Mhiri, I., Rekik, I., Initiative, A.D.N., et al.: Identifying the best data-driven feature selection method for boosting reproducibility in classification tasks. Pattern Recognition 101 (2020) 107183
Nebli, A., Gharsallaoui, M.A., Gurler, Z., Rekik, I., Initiative, A.D.N., et al.: Quantifying the reproducibility of graph neural networks using multigraph data representation. Neural Networks 148 (2022) 254–265
McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communicationefficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics, PMLR (2017) 1273–1282