Glioma pretraining boosts cross-center generalization of federated learning for MRI meningioma segmentation

For multi-center MRI meningioma segmentation, glioma-pretrained federated learning achieved external Dice 0.7503, comparable to centralized training (0.7452) and superior to meningioma-pretrained FL (0.7122; p<0.001).

• Centralized training on BraTS2023-Men showed a large external generalization drop: internal DSC 0.8958 fell to 0.7452 on the external SPHS cohort.
• Glioma-pretrained FL yielded DSC 0.7503, significantly outperforming meningioma-pretrained FL (DSC 0.7122, IoU 0.6301; Holm-adjusted p<0.001) and matching centralized training.
• Model was UMamba 2D using T1, T1c, and T2 MRI; FL was simulated across three clients.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.