/BrainGB

Officially Accepted to IEEE Transactions on Medical Imaging (TMI, IF: 11.037) - Special Issue on Geometric Deep Learning in Medical Imaging.

Primary LanguageMATLABMIT LicenseMIT

logo

Update: BrainGB is now officially accepted to IEEE Transactions on Medical Imaging (TMI, IF: 11.037) - Special Issue on Geometric Deep Learning in Medical Imaging.

BrainGB is a unified, modular, scalable, and reproducible framework established for brain network analysis with GNNs. It is designed to enable fair evaluation with accessible datasets, standard settings, and baselines to foster a collaborative environment within computational neuroscience and other related communities. This library is built upon PyTorch and PyTorch Geometric.

To foster research, we provide an out-of-box package that can be installed directly using pip, with detailed tutorials available on our hosted website. For more details, please check our paper here.


HennyJie - BrainGB stars - BrainGB forks - BrainGB language lines license visitor issue

Library Highlights

Our BrainGB implements four main modules of GNN models for brain network analysis:

  • Node feature construction: studies practical and effective methods to initialize either positional or structural node features for each brain region.
  • Message passing mechanisms: update the node representation of each brain region iteratively by aggregating neighbor features through local connections.
  • Attention-enhanced message passing: incorporates attention mechanism to enhance the message passing scheme of GNNs.
  • Pooling strategies: operate on the set of node vectors to get a graph-level representation.

BrainGB also implements utility functions for model training, performance evaluation, and experiment management.

Installation

To install BrainGB as a package, simply run

pip install BrainGB

Alternatively, you can also download the repository from Github. The main package is under the src folder. If you choose to go with this method, please check the Specification of Dependencies section for dependency requirements.

Specification of Dependencies

BrainGB depends on the following frameworks:

torch~=1.10.2
numpy~=1.22.2
nni~=2.4
PyYAML~=5.4.1
scikit-learn~=1.0.2
networkx~=2.6.2
scipy~=1.7.3
tensorly~=0.6.0
pandas~=1.4.1
libsvm~=3.23.0.4
matplotlib~=3.4.3
tqdm~=4.62.3
torch-geometric~=2.0.3
h5py~=3.6.0

To install the dependencies, run:

pip install -r requirements.txt

Notice that if you install the package through pip, the dependencies are automatically installed.

Getting Started

To import the models detailed in the paper:

from BrainGB.models import GAT, GCN, BrainNN, GCN

The BrainNN is required and will be served as the parent module of the GAT, GCN models. You may choose either GAT or GCN as the submodule.

To initialize a GCN model

sample: Data = Data()  # A torch geometric data

num_features = data.x.shape[1]
num_nodes = data.x.shape[0]
gcn_model = GCN(num_features, num_nodes)

model = BrainNN(args.pooling, gcn_model, MLP(2 * num_nodes))

To initialize a GAT model, simply replace the GCN with GAT. Both models are customizable. Please refer to the Customizing Your Own GNN Models section for more details.

Customizing Your Own GNN Models

Node Feature Construction

In src.dataset.tranforms, BrainGB provides the BaseTransform base class, which offers a universal interface for node feature initialization for each brain region. Specifically, BrainGB implements the following node feature construction functions:

Node Features Option Name
Identity identity
Eigen eigenvector
Degree degree
Degree Profile LDP
Connection Profile adj

To adjust the type of node features, simply set the chosen option name for the input parameter node_features.

Message Passing Mechanisms

In models.gcn, BrainGB provides the base class MPGCNConv and different message vector designs including:

Message Passing Mechanisms Option Name
Edge Weighted weighted_sum
Bin Concat bin_concate
Edge Weight Concat edge_weight_concate
Node Edge Concat edge_node_concate
Node Concat node_concate

To adjust the message passing schemes, simply set the input parameter model_name as gcn and chose an option name for the parameter gcn_mp_type.

Attention-Enhanced Message Passing

In models.gat, BrainGB provides the base class MPGATConv and different versions of attention-enhanced message passing designs including:

Message Passing Mechanisms Option Name
Attention Weighted attention_weighted
Edge Weighted w/ Attn attention_edge_weighted
Attention Edge Sum sum_attention_edge
Node Edge Concat w/ Attn edge_node_concate
Node Concat w/ Attn node_concate

Note that some of these options are corresponding attention enhanced version of the message passing mechanism designs. Please refer to our paper for more details.

To adjust the attention-enhanced message passing schemes, simply set the input parameter model_name as gat and chose an option name for the parameter gat_mp_type.

Pooling Strategies

The pooling strategy is controlled by setting the self.pooling in the chosen model. Specifically, BrainGB implements the following three basic pooling strategies:

Pooling Strategies Option Name
Mean Pooling mean
Sum Pooling sum
Concat Pooling concat

To adjust the pooling strategies, simply set the chosen option name for the input parameter pooling.

Running Example Scripts

The repository also comes with example scripts. To train our model on any of the datasets we tested, simply run:

python -m main.example_main --dataset_name=<dataset_name> [--model_name=<model_name> --gcn_mp_type=<mp_mechanism>  --gat_mp_type=<attention_mp_mechanism> --node_features=<feature_name> --pooling=<pooling_name> --n_GNN_layer=<GNN_num> --n_MLP_layers=<MLP_num> --hidden_dim=<hidden_layer_dimension> --epochs=<epoch_num> --k_fold_splits=<split_num> --test_interval=<evaluation_interval_num>]

The dataset_name is the name of the dataset to use (required parameter). We include the following four datasets in our paper:

  • HIV
  • PNC (Can be downloaded here)
  • PPMI (Can be downloaded here)
  • ABCD (Can be downloaded here)

You can also construct your own datasets by following the instructions on neuroimaging preprocessing and brain network construction on our website.

Please place the dataset files in the datasets folder under the package examples folder. Create the folder if it does not exist.

The model_name specifies the backbone model type. Choose gcn to test the message passing variants without attention and gat to test the attention-enhanced message passing mechanisms. Specifically, use gcn_mp_type to set a message vector design and use gat_mp_type to set an attention-enhancing mechanism.

The node_features specifies the artificial node feature initialization for each brain region.

The pooling specifies the pooling strategy to get a graph-level representation for each subject.

You can also change other hyper-parameters, such as --n_GNN_layer, --n_MLP_layers, --hidden_dim, --epochs, etc., to adjust the detailed model design or control the training process. All those hyper-parameters can be automatically searched and optimized using the AutoML tool NNI by passing --enable_nni.

Contribution

Feel free to open an issue should you find anything unexpected or create pull requests to add your own work! We welcome contributions to this benchmark work and the package.

Citation

Please cite our paper if you find this code useful for your work:

@article{cui2022braingb,
author = {Cui, Hejie and Dai, Wei and Zhu, Yanqiao and Kan, Xuan and Chen Gu, Antonio Aodong and Lukemire, Joshua and Zhan, Liang and He, Lifang and Guo, Ying and Yang, Carl},
title = {{BrainGB: A Benchmark for Brain Network Analysis with Graph Neural Networks}},
journal={TMI},
year = {2022},
}