145 - Federated Deep Ensemble Learning for MRI-Based Brain Metastasis Segmentation: An International Multi-Center Development

Purpose/Objective(s):

Precise brain metastasis (BM) segmentation is critical for accurate stereotactic radiosurgery (SRS) planning. Manual contouring is labor-intensive and subject to inter-observer variability, and development of deep learning automation solutions are limited by the for aggregation of large, centralized datasets with restrictive data protection requirements. To overcome this, we developed and validated a dedicated Federated Learning (FL) framework with a novel learnable ensemble fusion strategy collaboration, thereby establishing a platform for development of enhanced BM segmentation solutions that leverages global experience without dependence on large-scale data transfers.

Materials/Methods:

Two institutional patient cohorts (n=90, Institution-S; n=459, Institution-L) consisting of post-contrast T1w MRI and expert BM contours were retrospectively analyzed, with 7:1:2 for training/validation/test at each site. To capture multi-scale anatomical features, we implemented a multi-center spherical re-sampling to generate 27 locoregional views uniformly across the field-of-view, creating varied locoregional perspectives. A customized UNet++ architecture served as the core segmentation backbone. During inference, 27 independent segmentations per case were inversely mapped to the Cartesian grid and integrated via a learnable weighted ensemble strategy. Subsequently, institutional model parameters (no patient data) were uploaded to a central server to initiate the FL process. Ten communication rounds were implemented, utilizing central server model averaging and local model updates guided by institutional validation sets. Performance was compared against institution-specific ensemble models.

Results:

Independent testing at the larger site (Institution-L) demonstrated that the deep ensemble strategy achieved satisfactory BM detection performance (Dice=0.768, F1=0.820, sensitivity=0.868). In contrast, the data-limited site (Institution-S) revealed that its local model struggled with detection (Dice=0.583, F1=0.714, sensitivity=0.627). The proposed FL approach significantly (p<0.01) bolstered BM detection at the smaller institution (Dice=0.641, F1=0.724, sensitivity=0.806). Meanwhile, the FL model maintained a high-performance level at the larger institution (Dice=0.778, F1=0.813, sensitivity=0.835), proving that the federated framework retains high accuracy and generalizability across diverse data scales.

Conclusion:

This international study demonstrates that an ensemble-based FL framework is generalizable across institutions and enables development of high-fidelity, automated BM segmentation without requiring raw data sharing. This strategy provides a scalable, privacy-compliant pathway for improving BM care and standardizing SRS quality across global radiation oncology networks.