HarmoFL: Harmonizing Local and Global Drifts in Federated Learning on Heterogeneous Medical Images