Related Experiment Videos
Improving Cross-Center Generalization for Multi-modal MRI Meningioma Segmentation via Glioma-Pretrained Federated
Chendong Ni1, Kun Qian2, Chulong Zhang2
1Medical Physics Graduate Program, Duke Kunshan University; Radiotherapy Business Unit, Shanghai United Imaging Healthcare Co., Ltd.
Journal of Visualized Experiments : Jove
|August 10, 2026
Summary
Federated learning for meningioma segmentation improves robustness when initialized with a glioma model, outperforming standard methods and enhancing external generalization without data sharing.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Healthcare
- Neuro-oncology Research
Background:
- Automated meningioma segmentation across institutions is challenging due to domain shift from varying scanner protocols and image characteristics.
- Federated learning (FL) offers privacy-preserving multi-center model development but standard aggregation may not fully address cross-center domain differences.
- Quantifying the external generalization gap in meningioma segmentation is crucial for robust clinical deployment.
Purpose of the Study:
- To evaluate the external generalization gap in multi-modal MRI meningioma segmentation.
- To assess if initializing federated learning with a glioma model improves robustness compared to meningioma-pretrained FL.
- To determine if this approach enhances generalization without centralized data pooling.
Main Methods:
- A UMamba 2D architecture was used for binary meningioma segmentation on T1, T1c, and T2 MRI sequences.
- Three strategies were evaluated: centralized training on BraTS2023-Men, meningioma-pretrained FL, and Glioma-pretrained FL.
- External validation was performed on 174 independent clinical cases, assessing performance with Dice Similarity Coefficient (DSC) and Intersection over Union (IoU).
Main Results:
- Centralized training on BraTS2023-Men showed a significant drop in external generalization (DSC from 0.8958 to 0.7452).
- Meningioma-pretrained FL resulted in lower external performance (DSC = 0.7122).
- Glioma-pretrained FL achieved comparable performance to centralized training (DSC = 0.7503) and significantly outperformed meningioma-pretrained FL (p < 0.001).
Conclusions:
- Glioma-pretrained initialization provides a more robust starting point for federated meningioma segmentation.
- Federated learning initialized with a glioma model can improve external generalization in meningioma segmentation.
- This approach enhances robustness while preserving institutional data privacy.