Improving Cross-Center Generalization for Multi-modal MRI Meningioma Segmentation: Centralized Training vs Federated Learning vs Transfer-Initialized Federated Learning with External Validation on The Second People’s Hospital of Shenzhen Data

Abstract

Purpose:To evaluate whether federated learning (FL) and transfer learning (TL) improve generalization for multi-modal MRI (T1c/T1n/T2w) meningioma (MEN) segmentation on a real-world external clinical dataset from The Second People’s Hospital of Shenzhen (SPHS), under privacy constraints and limited cross-center data sharing.Methods:We used 450 cases from BraTS2023_Men as the primary training source and adopted UMamba 2D as the backbone. The task was binary tumor segmentation with three MRI modalities (T1c, T1n, T2w). To harmonize annotation systems, SPHS labels retained tumor mass only (label1), while BraTS labels were mapped by merging label1 (non-enhancing core) and label3 (enhancing tumor) as “tumor mass,” excluding label2. SPHS data were processed with DICOM-to-NIfTI conversion, resampling to 1×1×1 mm, mutual-information (MI) rigid inter-modality registration (T1 as reference), and union-based multi-modal mask fusion followed by mild morphological smoothing. We compared three strategies with external validation on SPHS: (1) centralized training; (2) in-domain initialized FL (first three FedAvg rounds); and (3) transfer-initialized FL (initialized from a BraTS2023_Gli pretrained model; first three FedAvg rounds). Evaluation followed a case-averaged protocol, reporting Dice/IoU and error components (TP/FP/FN/TN, predicted/reference voxel counts).Results:The centralized model achieved Dice=0.8958 and IoU=0.8356 on the BraTS2023_Men test set and dropped to Dice=0.7452 and IoU=0.6241 on SPHS. In-domain initialized FL produced Dice=0.7070 (Global model 0) and then Dice=0.6253/0.6890/0.7122 across FedAvg rounds 1–3 on SPHS, with Round 1 showing pronounced over-segmentation (FP=21,169.13). Transfer-initialized FL followed a low-to-high trajectory on SPHS (Dice=0.2441/0.4138/0.7503 across rounds 1–3), with Round 3 being the best among all strategies and slightly exceeding the centralized SPHS Dice. A Gli-only model directly evaluated on SPHS achieved Dice=0.6972, below transfer-initialized FL Round 3.Conclusion:External validation suggests that FL does not guarantee generalization gains under non-IID and label-shift conditions. Transfer initialization from a related pretrained model, followed by a small number of federated adaptation rounds, was more likely to yield improved and more stable external generalization. Overall, “strong initialization + federated adaptation” appears to be a favorable pathway for cross-center MEN segmentation under privacy constraints.Keywords: meningioma segmentation; multi-modal MRI; UMamba; federated learning; transfer learning; external validation; generalization