Partial Participation with Bias Correction for Hospital-Scale Hierarchical Federated Multi-Task Segmentation

Federated learning (FL) enables clinical institutions to collaboratively train deep learning models without sharing patient data, but a single central server is often impractical across national or regional regulatory and network boundaries. Hierarchical FL addresses this through regional aggregation servers; yet combining it with partial, heterogeneous client participation and multi-task learning can introduce participation bias and amplify task- and client-level data heterogeneity. We compare averaging-based aggregation schemes under full and partial client participation on two benchmark neuroimaging segmentation tasks, ATLAS for ischemic stroke lesion segmentation and BraTS for glioma segmentation. Experiments span hierarchical vs. single-server topologies and single-task vs. joint multi-task training under heterogeneous participation patterns. To our knowledge, this is the first systematic study to combine hierarchical multi-server FL, multi-task joint training, and explicit participation-bias correction for neuroimaging segmentation across ATLAS and BraTS.