This project contains the architecture, training, and testing frameworks for our cVAE for harmonizing connectomes. Outputs include a set of site-invariant features and harmonized network properties (set of scalars).
This code accompanies the publication "Harmonizing 10,000 connectomes: Site-invariant representation learning for multi-site analysis of network connectivity and cognitive impairment".
- The data set up in ExampleData is how the dataloader_train and dataloader_test expect the data to be organized inside of DATA/TRAIN
- If running yourself, make sure to change the paths to your desired input/output locations.
The architecture follows this schematic:
Here is an example figure showing the clustering based on site and diagnosis in uncorrected (left) and corrected (right) connectome features.
We trained this model with Quadro RTX 5000 and 16GB RAM with PyTorch version 0.4.1.