A Federated Multigraph Integration Approach for Connectional Brain Template Learning

The connectional brain template (CBT) is a compact representation (i.e., single connectivity matrix) multi-view networks of given population. CBTs are especially very powerful tools in dysconnectivity diagnosis as well holistic mapping if they learned properly – i.e., occupy the center Even though accessing large-scale datasets much easier nowadays, it still challenging to upload all these clinical server altogether due data privacy and sensitivity. Federated learning, on other hand, has opened new era for machine learning algorithms where different computers trained together via distributed system. Each computer client) connected server, trains model with its local dataset sends learnt weights back server. Then, aggregates thereby outputting global encapsulating information drawn from privacy-preserving manner. Such pipeline endows generalizability power implicitly benefits diversity datasets. In this work, we propose first federated (Fed-CBT) framework learn how integrate connectomic collected by hospitals into representative map. First, choose random fraction train our model. Next, send their aggregate them. We also introduce weighting method aggregating take full benefit hospitals. Our best knowledge only estimate templates using graph neural networks. Fed-CBT code available at https://github.com/basiralab/Fed-CBT.