Partial Participation and Data Heterogeneity in Cross-Clinic Federated Medical Image Segmentation: An Empirical Study

Precise medical image segmentation is a core component of automated diagnosis and treatment planning. Since the amount of data available at any single clinic is often limited, federated learning (FL) provides a natural framework for jointly training models across institutions without exchanging raw patient data. One of the key practical challenges in FL is partial client participation: some clinics may be unavailable in a given communication round because of connectivity, resource constraints, or clinical workload. Importantly, the resulting degradation can be further amplified by cross-clinic data heterogeneity, as participating clients may represent systematically different data distributions due to differences in acquisition protocols, imaging equipment, and patient populations. We study this problem in two medical image segmentation setups, both based on encoder–decoder architectures with client-specific encoders and a shared decoder. The first considers brain tumor segmentation with SegResNet on multi-modal FeTS dataset with natural heterogeneity. The second considers abdominal multi-organ segmentation on BTCV using a pretrained SwinUNETR model. We then investigate whether the adverse effects of partial client participation can be mitigated. Our results show that the error-corrected aggregation scheme PPBC yields more stable convergence under partial participation than baseline aggregation. These findings suggest that error-corrected aggregation can mitigate the combined effects of intermittent client availability and heterogeneous local data, providing a more robust training mechanism for real-world cross-clinic federations in which consistent participation of all institutions cannot be assumed.