BrainRGIN
Extending from the existing graph convolution networks, our approach incorporates a clustering-based embedding and graph isomorphism network method in the graph convolutional layer to reflect nature of the brain sub-network organization and efficient expression, in combination with TopK pooling and attention-based readout functions.
Install / Use
/learn @bishalth01/BrainRGINREADME
BrainRGIN (Brain ROI Aware Graph Isomorphism Networks)
Description
BrainRGIN extends existing Graph Convolution Networks (GCNs) by incorporating clustering-based embeddings and a Graph Isomorphism Network (GIN) in the graph convolutional layers. This method better captures the structure and organization of brain sub-networks, providing efficient and expressive representations. Our approach combines TopK pooling and attention-based readout functions.
This work has been accepted for publication in Medical Image Analysis. The DOI for citation will be updated soon.
Creating Graphs
We used DGL (Deep Graph Library) to create graphs from the PyTorch Data object. Sample details and graphs can be found in the save_dgl_graphs/* directory.
How to Run the Code
- Main File:
hcp_main/main_rgin - Hyperparameters: Configurations for hyperparameter tuning are available in
wandb_sweeps/example_config.
For questions or contributions, feel free to open an issue or create a pull request!
