Sounds like they use a method called federated learning. I.e. each site can train a local model on their own data and then share model parameters to a central entity that aggregates several sites model parameters. Superior performance can then be achieved (and also raw training data between sites are not shared).
Nonetheless, I don’t like this…