Improving Cross-Center Generalization for Multi-modal MRI Meningioma Segmentation via Glioma-Pretrained Federated Learning

Summary

This study investigates the impact of Glioma-pretrained federated learning on external generalization in multi-modal MRI meningioma segmentation

Abstract

Automated meningioma segmentation on multi-modal MRI remains challenging when models are transferred across institutions, because scanner protocols, image characteristics, and annotation styles may differ between centers. Federated learning (FL) provides a privacy-preserving strategy for multi-center model development, but standard aggregation may not fully overcome cross-center domain shift. This study aimed to quantify the external generalization gap in MRI meningioma segmentation and evaluate whether Glioma-pretrained FL could improve robustness without centralized data pooling. A UMamba 2D architecture was used for binary meningioma segmentation using T1, T1c, and T2 MRI as model inputs. The protocol included 450 BraTS2023-Men cases as the source-domain meningioma dataset and 174 independent clinical cases from our institute as the external validation cohort. Three final meningioma segmentation strategies were quantitatively evaluated under the same external validation setting: centralized training on BraTS2023-Men, meningioma-pretrained FL across three simulated clients, and Glioma-pretrained FL initialized from a BraTS2023-Gli source model before federated optimization. Centralized training on Glioma was used only to generate the glioma-pretrained initialization and was not reported as an independently evaluated final meningioma segmentation strategy. Model performance was assessed using the Dice similarity coefficient (DSC) and Intersection over Union (IoU). Centralized training on BraTS2023-Men showed a clear external generalization drop, with DSC decreasing from 0.8958 on the internal BraTS2023-Men test set to 0.7452 on the SPHS cohort. Meningioma-pretrained FL yielded lower external performance (DSC = 0.7122), whereas Glioma-pretrained FL achieved comparable performance to centralized training on BraTS2023-Men and improved over meningioma-pretrained FL (DSC = 0.7503; IoU = 0.6301; Holm-adjusted p < 0.001). These results suggest that glioma-pretrained initialization provides a more robust starting point for federated meningioma segmentation and may improve external generalization while preserving institutional data privacy.

Introduction

Meningiomas are the most prevalent primary intracranial neoplasm in adults, accounting for approximately 38% of all primary brain and central nervous system tumors in the United States¹. Originating from meningothelial (arachnoid cap) cells, they encompass World Health Organization (WHO) grades I–III, spanning a wide clinical spectrum: grade-I tumors (approximately 80%) are generally associated with favorable outcomes following gross total resection, whereas grade-II (atypical) and grade-III (anaplastic) lesions exhibit substantially higher recurrence rates of 20–40% and exceeding 70% at ten years, respectively^2,3. Clinical decision-making—including microsurgical resection, stereotactic radiosurgery (SRS), fractionated radiotherapy, and active surveillance—depends on accurate and reproducible volumetric delineation of tumor extent across longitudinal imaging studies. Robust segmentation of meningiomas from magnetic resonance imaging (MRI) is therefore a prerequisite for consistent treatment planning and outcome assessment.

Multi-sequence MRI forms the foundation of meningioma characterization in routine clinical practice, with each sequence encoding complementary aspects of tumor biology and surrounding tissue response. Contrast-enhanced T1-weighted imaging (T1c) highlights regions of blood–brain barrier disruption through gadolinium-induced T₁ relaxation shortening, enabling clear visualization of enhancing tumor parenchyma and definition of the primary lesion boundary. Non-contrast T1-weighted imaging (T1) provides structural context for evaluating the arachnoid interface, dural attachment, and adjacent osseous involvement. T2-weighted imaging (T2) captures intra-tumoral heterogeneity and peritumoral edema, both of which are essential for understanding mass effect and for delineating treatment volumes in radiosurgical planning^3,4. Manual contouring remains the standard approach for meningioma delineation. During this process, clinicians typically consider complementary contrast information across multiple MRI sequences to define tumor extent. However, this procedure is time-consuming and is associated with substantial inter-observer variability, particularly in anatomically ambiguous regions such as the dural tail, where reported agreement metrics fall below 0.80^5,6. These limitations motivate the development of automated segmentation approaches that deliver consistent, scalable performance.

Deep learning has substantially advanced medical image segmentation, supported by benchmark datasets such as BraTS⁷ and by the emergence of multiple architectural paradigms, including convolutional models such as U-Net⁸ and nnU-Net⁹, Transformer-based models such as UNETR and Swin-UNETR^10,11, and state space model (SSM)-based approaches such as UMamba¹². Recent work has further explored explainable attention-guided Swin Transformer networks and global-local Vision Transformer/axial state-space designs for 3D brain tumor segmentation^13,14, highlighting the continued development of transformer- and state-space-based segmentation pipelines. Despite strong benchmark performance, meningioma-specific studies have historically been limited by relatively small cohort sizes (typically 50–200 cases)^15,16. The release of the BraTS 2023 meningioma dataset¹⁷, with over 1,000 cases, partially alleviates this limitation; however, most models are still trained and evaluated within curated datasets, and their performance often degrades when applied to independent clinical data¹⁸. This highlights a core issue: achieving robust cross-center generalization requires exposure to diverse multi-center data.

In practice, aggregating large-scale multi-center datasets is constrained by data privacy regulations and institutional policies, making centralized training difficult or infeasible. Federated learning (FL) addresses this constraint by enabling collaborative model training without sharing raw data, where each institution trains locally and contributes model updates to a global model via aggregation, as in Federated Averaging (FedAvg)¹⁹. Related work has also extended privacy-preserving FL to human-intention modeling in pediatric cerebral palsy using extended reality²⁰, and lightweight deep learning with virtual-reality visualization has been proposed for offline tumor segmentation in rural environments²¹, reflecting the broader need for deployable and privacy-aware medical AI workflows. While FL provides a practical framework for multi-center learning, it introduces a new challenge: each participating site typically has a limited local dataset, and independent local training before aggregation can lead to suboptimal local updates and degraded global performance, particularly when local data distributions diverge-a problem known as client drift²². A natural complementary strategy is to initialize the model using knowledge learned from a related source dataset²³. In brain tumor imaging, glioma represents a closely related source task with partially shared imaging characteristics and tumor-context features, making it a practical source domain for pretraining. Although meningiomas and gliomas differ in biological characteristics and imaging appearance, they share common imaging modalities and anatomical context, allowing transferable low- and mid-level representations to be learned³. Prior work has shown that such cross-tumor pretraining can improve performance in data-limited settings²³. Despite these developments, it remains unclear how FL performs for meningioma segmentation under realistic multi-center conditions, and whether glioma pretraining can improve federated training when each site has limited data. The relative benefits of FL alone versus Glioma-pretrained FL have not been evaluated under external clinical validation.

To address this gap, this study conducts a controlled comparison using an independent external cohort from SPHS. Specifically, we investigate three questions: (i) the magnitude of the cross-center generalization gap between models trained on benchmark data and external clinical data; (ii) whether FL can reduce this gap under limited local data conditions; and (iii) whether Glioma-pretrained FL (glioma -> meningioma) provides additional performance gains over both centralized training and standard federated approaches.

Protocol

This study was a retrospective clinical study. The study protocol was reviewed and approved by the Ethics Committee of Second People’s Hospital of Shenzhen for the collection and use of the SPHS clinical dataset, with the approval number 20190910. Written informed consent was obtained from all patients included in the SPHS cohort prior to enrollment in this study.

Patient data

This study used three datasets: BraTS2023-Men, BraTS2023-Gli, and an external clinical cohort from SPHS (SPHS-Men). Their characteristics are summarized below.

BraTS2023-Men

BraTS2023-Men is the meningioma track of the BraTS 2023 benchmark^24,25. For this study, 450 BraTS2023-Men cases were used as the source cohort for model development and internal evaluation. All images were preprocessed uniformly, including skull stripping, co-registration across modalities, and isotropic resampling to 1×1×1 mm³. Each case contains four MRI sequences (T1, T1c, T2, FLAIR) with voxel-wise annotations in three categories: non-enhancing tumor core (label 1), peritumoral edema (label 2), and enhancing tumor (label 3). For consistency with the clinical tumor definition used in this study, labels 1 and 3 were merged into a single tumor class, and label 2 (edema) was excluded, resulting in a binary segmentation target.

BraTS2023-Gli

BraTS2023-Gli is the glioma track of the BraTS benchmark and inherits the established BraTS multi-modal MRI conventions and annotation framework^7,24,26,27. In this study, Centralized training on Glioma was performed using 450 BraTS2023-Gli cases to provide a source model for Glioma-pretrained FL, using the same binary tumor-versus-background formulation later used for the meningioma task. The goal of including this cohort was not to increase sample size relative to BraTS2023-Men, but to test whether a task-related brain-tumor source domain could provide a more transferable initialization for subsequent federated adaptation.

SPHS-Men

SPHS-Men is an external clinical dataset collected from the Second People’s Hospital of Shenzhen under institutional review board approval. A total of 174 cases were included for external evaluation. Unlike BraTS, the data were acquired using routine clinical protocols without standardized preprocessing, leading to variability in voxel spacing, slice thickness, field of view, and scanner-dependent reconstruction parameters. Annotations were originally provided with two labels: tumor mass and edema/peritumoral change. To align with the BraTS-derived tumor definition used in this study, only tumor mass was retained, and edema was excluded. All SPHS cases were subsequently processed using a harmonization-oriented pipeline before external evaluation, as described in MRI acquisition and preprocessing. After preprocessing, each voxel at spatial coordinate r is represented by a three-channel input consisting of T1c, T1n, and T2 intensities,

and the segmentation task is defined as binary classification Y(r) ∈ 0,1 indicating tumor versus background. Representative examples and qualitative differences across datasets are shown in Figure 1.

Figure 1. Representative cross-dataset comparison. Top row: SPHS-Men; middle: BraTS2023-Men; bottom: BraTS2023-Gli. Columns show tumor mask, T1, T1c, and T2. Qualitative differences in lesion size, enhancement pattern, and boundary definition reflect the compounded imaging and annotation shifts encountered in this study. Please click here to view a larger version of this figure.

MRI acquisition and preprocessing

All BraTS data were distributed with standardized preprocessing already applied, including skull stripping, co-registration, and isotropic resampling. For the SPHS cohort, raw DICOM images were first converted to NIfTI format²⁸. The multimodal images were then rigidly registered to the native T1 space using mutual information-based alignment²⁹. Because skull stripping is a necessary preprocessing step for brain MRI segmentation, brain extraction was explicitly performed before model training. Specifically, DeepBet was applied to the T1-weighted image to generate a brain mask³⁰, and the resulting mask was propagated to the co-registered T1c and T2 volumes to remove non-brain tissue consistently across all modalities. The skull-stripped images were subsequently resampled to an isotropic resolution of 1 x 1 x 1 mm³. Binary tumor masks were finally refined using morphological closing and opening operations to suppress interpolation artifacts introduced during resampling. These harmonization steps were used to improve cross-dataset input consistency for external testing, but they were not intended to eliminate the underlying scanner-, protocol-, or annotation-level differences intrinsic to the SPHS cohort.

Study design

Model architecture

As shown in Figure 2, the UMamba 2D architecture integrates residual convolutional neural network modules with Mamba selective state-space model (SSM) blocks^12,31. This hybrid design enables the simultaneous extraction of local texture features via convolutions and long-range global structural context via SSMs, with a linear computational cost with respect to the number of spatial positions.

Figure 2. The UMamba 2D architecture. The encoder integrates residual convolutional blocks for local texture and boundary feature extraction with Mamba selective SSM modules for long-range context modeling. These components are connected via skip connections to a transposed-convolution decoder. The network input consists of a three-channel multi-modal MRI slice comprising T1, T1c, and T2 modalities. Please click here to view a larger version of this figure.

Encoder

Multi-modal MRI slices with C = 3 input channels are processed by a hierarchical encoder. At each stage s ∈ 1,…, S, the input feature map is first processed by a residual convolutional block to extract localized spatial patterns. The feature map is then flattened into a one-dimensional sequence and processed by a selective SSM module to capture long-range contextual dependencies. Spatial downsampling is performed using stride-2 convolutions, progressively generating hierarchical feature representations F_s ∈ R^C_s × H/2^s × W/2^s with increasing channel dimensions C_s ∈ 32, 64, 128, 256, 512.

Mamba selective SSM

The selective SSM operates on a one-dimensional sequence x ∈ R^L × d. Because spatial adjacency is disrupted during standard one-dimensional flattening of two-dimensional feature maps, a multi-directional scanning strategy is used: feature maps are scanned along four directions (row forward, row backward, column forward, and column backward) to preserve spatial context across both row and column directions. The continuous-time SSM is defined as:
    

where A ∈ R^N × N denotes the learnable state transition matrix initialized via the HiPPO (Hidden Polynomial Projection Operator) framework³¹. The projection matrices B(t) and C(t) are input-dependent, establishing a selective mechanism that enables the model to dynamically filter irrelevant tokens and retain important contextual information. The continuous system is discretized using a time step Δ via a zero-order hold (ZOH) transformation:

At each discrete step k, the parameters B, C, and Δ are computed as linear projections of the input token x_k, enabling input-dependent feature propagation. The HiPPO initialization ensures that the eigenvalues satisfy |λ_j(barA)| < 1, which keeps the hidden state representations ‖h_k‖ bounded across arbitrarily long spatial sequences.

Decoder and skip connections

The decoder reconstructs the original spatial resolution from the encoded representations. At each stage, transposed convolutions upsample the lower-resolution feature maps. Skip connections concatenate the multi-scale encoder features F_s with the corresponding upsampled decoder features, preserving high-frequency boundary information that is essential for precise voxel-level tumor delineation.

Training loss

Model optimization is guided by a composite loss function:

where p_θ(r) ∈ [0,1] is the predicted tumor probability at voxel r, Y(r) is the corresponding binary ground-truth label, and λ = 1.0 balances the two terms. The Dice loss directly optimizes the target evaluation metric and addresses the extreme foreground–background imbalance inherent to meningioma segmentation, where tumor voxels typically constitute less than 0.5% of the total intracranial volume. The binary cross-entropy (BCE) term provides per-voxel supervision, stabilizing training in this imbalanced setting.

Training configuration

All segmentation models were optimized using AdamW with an initial learning rate of η₀ = 3 × 10⁻⁴, cosine annealing schedule, and weight decay of 5 × 10⁻⁴. The centralized baseline models were trained for 200 epochs on the 450-case BraTS2023 cohorts with a batch size of 16. To avoid ambiguity in the federated setting, we use the following terminology throughout this manuscript: a communication round denotes one complete FL cycle consisting of server broadcast, local client optimization, and server aggregation; a mini-batch iteration denotes one parameter update on one mini-batch; and the local optimization performed by each client in one round consisted of 250 mini-batch iterations. In other words, each client performed one local training stage per round, and that local stage contained 250 parameter-update iterations. Thus, the term 'round' (r = 3) in this study refers only to FL communication, not to an epoch across the entire centralized dataset. An online data augmentation pipeline was applied uniformly across all training configurations, including random horizontal and vertical flipping (p=0.5 per axis), random in-plane rotation (±15^circ), random intensity scaling (factor uniformly sampled from [0.85, 1.15]), and additive Gaussian noise (sigma ∼ U[0, 0.05]) computed per two-dimensional slice. All augmentation was disabled during inference and validation. The framework was implemented in PyTorch 2.0.1 with MONAI 1.3 for data loading, and all computations were performed on NVIDIA GeForce RTX 3090 (24 GB GDDR6X) graphics processing units (GPUs). Reproducibility was ensured by fixing all random seeds across PyTorch, NumPy, and the Python random module to identical values for all strategies.

Experimental strategies

Three primary training strategies were compared to evaluate cross-center generalization (Figure 3). The labels meningioma-pretrained FL and Glioma-pretrained FL are descriptive names introduced in this manuscript to distinguish the source of the initial global model before federated adaptation.

Figure 3. Experimental design overview. (A) Centralized training using 450 BraTS2023-Men cases followed by external evaluation on SPHS. (B) Meningioma-pretrained FL across three simulated clients. (C) Glioma-pretrained FL initialized from a centralized glioma model before federated adaptation. Please click here to view a larger version of this figure.

Strategy 1 (Centralized training)

A single UMamba 2D model was trained on all 450 BraTS2023-Men cases. This strategy represents the upper bound on in-domain performance and serves as the baseline for assessing cross-center degradation.

Strategy 2 (FL with MEN-domain initialization)

The 450 BraTS2023-Men cases were randomly assigned to three clients (|D_i| = 150) for equal partitioning. The three BraTS2023-Men clients were therefore simulated from the same public cohort rather than constructed using institution-specific metadata or scanner-origin labels. To initialize federated training in the meningioma domain, the initial global model was first trained on a fixed BraTS2023-Men client split (client 1 in this implementation) and then used to initiate the federated procedure. This seed client was held constant across experiments for reproducibility and should be interpreted as a within-domain warm start rather than an optimized or privileged source client. Because the three clients were simulated by partitioning the same BraTS2023-Men cohort, this initialization choice reflects an implementation-level starting point rather than a biologically distinct institutional prior. The purpose of this design was not to strictly replicate a real multi-institution launch process, but to provide a controlled comparison between centralized training on all available MEN data and federated collaboration starting from a partial meningioma-pretrained FL. The global federated optimization objective is defined as:

where D_i denotes the local training set of clients i, and ell is the Dice + cross-entropy loss. In communication round 1, aggregation weights were proportional to sample count (w_i⁽¹⁾ = 1/3). To down-weight underperforming clients in subsequent communication rounds, rounds 2 and 3 used validation Dice-weighted aggregation:

where DSC_i^(r) is the case-averaged DSC on the local validation partition of client i after local training in communication round r, and ε = 10⁻⁶ is a small constant to prevent division by zero.

Strategy 3 (FL with glioma-pretrained initialization)

To obtain a task-related cross-domain warm start, Centralized training on Glioma was first performed using 450 BraTS2023-Gli cases with the same three MRI modalities (T1, T1c, and T2) and the same binary tumor-versus-background formulation adopted in the meningioma experiments.

Specifically, glioma labels were harmonized by combining the non-enhancing tumor core and enhancing tumor into a single foreground class, while peritumoral edema was excluded, so that the source and target tasks shared the same binary output definition. Because both source and target tasks were formulated as binary segmentation, the pretrained weights were transferred directly to initialize the full UMamba 2D model, including the segmentation head, rather than discarding and reinitializing the final layer. The same FedAvg protocol was then applied, with the initial global model initialized by the converged Centralized training on Glioma weights (θ⁽°⁾ = θ_Gli^*). The subsequent federated adaptation is given by:

where Δθ^(r) represents the accumulated domain-specific parameter updates across r communication rounds. This strategy tests whether a Glioma-pretrained FL source model provides a better starting point for federated adaptation on meningioma clients.

Evaluation metrics

All strategies were evaluated on the SPHS cohort using a case-averaged protocol. The DSC and IoU were computed independently per case and then averaged, ensuring equal statistical weight across all subjects regardless of tumor volume:

 

Case-averaged error compositions—comprising mean true positive (TP), false positive (FP), false negative (FN), and true negative (TN) voxel counts per case—were also reported. This breakdown distinguishes over-segmentation from under-segmentation errors, providing insight into directional failure modes.

Results

The centralized UMamba 2D model (Strategy 1) achieved competitive in-domain performance on the BraTS2023-Men internal test set, with a case-averaged DSC of 0.8958, IoU of 0.8356, FN of 1,902, FP of 667, TN of 8,902,814, and TP of 22,616. When evaluated on the SPHS external validation cohort, the centralized model yielded a DSC of 0.7452 ± 0.1992 and an IoU of 0.6241 ± 0.1941, corresponding to a cross-center generalization gap of ΔDSC = 0.1506 (Table 1, ID 1). As an additional self-configuring baseline, nnUNet-v2 achieved a DSC of 0.7381 ± 0.1931 and an IoU of 0.6142 ± 0.1897 on the same SPHS external validation cohort without federated learning or glioma pretraining.

The error composition (Table 1, ID 1) showed a predominantly under-segmentation pattern, with a mean FN count of 14,873 voxels per case versus a mean FP count of 2,644, indicating that the centralized model identified the gross tumor region but missed peripheral and low-contrast boundary components under external acquisition conditions.

Meningioma-pretrained FL achieved a DSC of 0.7122 ± 0.1894 and an IoU of 0.5799 ± 0.1863 on the SPHS cohort (Table 1), remaining below the centralized baseline by 0.0330 in DSC. The corresponding error composition (Table 2) showed reduced false positives relative to the centralized model but increased false negatives, indicating that this strategy did not fully resolve the external generalization gap.

Glioma-pretrained FL achieved the best external performance among the three final comparison models, with a DSC of 0.7503 ± 0.1965 and an IoU of 0.6301 ± 0.1934 on the SPHS cohort (Table 1). This result was numerically slightly higher than centralized training (+0.0051 DSC), but the difference was not statistically significant after Holm correction (p = 0.1488). The corresponding error composition (Table 2) showed the highest TP count together with a more balanced FP/FN profile, suggesting that Glioma-pretrained FL provided a more transferable starting point for subsequent meningioma-side federated adaptation.

Model	DSC (mean±SD)	IoU (mean±SD)	p value (DSC)	p value (IoU)
Centralized	0.7452 ± 0.1992	0.6241 ± 0.1941	0.148827896	0.125240623
Meningioma-pretrained FL	0.7122 ± 0.1894	0.5799 ± 0.1863	5.57014×10⁻²¹	1.17404×10⁻²²
Glioma-pretrained FL	0.7503 ± 0.1965	0.6301 ± 0.1934	–	–
SD = standard deviation across cases. Best result in Red.

Table 1: External performance on SPHS (case-averaged, mean ± SD). P values were computed using pairwise case-level Wilcoxon signed-rank tests against the final Glioma-pretrained FL model, followed by Holm correction for multiple comparisons.

Model	FN	FP	TN	TP
Centralized	14,873	2,644	7,319,250	31,090
Meningioma-pretrained FL	16,038	2,068	7,319,827	29,925
Glioma-pretrained FL	12,709	4,278	7,317,617	33,254

Table 2: External error composition on SPHS (case-averaged voxel counts) for the three final comparison models.

Figure 4. Representative qualitative segmentation outputs on SPHS-Men cases. Rows represent three representative SPHS-Men cases. Columns display the T1c image, ground truth (GT) annotation, centralized model output, the final model from meningioma-pretrained FL, and the final model from Glioma-pretrained FL. All predictions were generated using the full three-channel input (T1, T1c, and T2); only the T1c slice is displayed for visual reference. Please click here to view a larger version of this figure.

Representative SPHS segmentation outputs are presented in Figure 4. The centralized model captured the bulk tumor volume but exhibited visible boundary under-segmentation, consistent with the FN-dominant error composition. The final model from meningioma-pretrained FL remained more conservative, with residual boundary omissions despite fewer spurious extensions. The final model from Glioma-pretrained FL produced contours with the strongest geometric agreement to ground truth, capturing tumor extent with improved precision while reducing both FP inclusions and FN boundary omissions.

Data Availability:

The public BraTS2023-Men and BraTS2023-Gli datasets used in this study are available through the official BraTS 2023 Challenge data-access platform on Synapse (https://www.synapse.org/brats2023), subject to the corresponding BraTS data-use agreements and access requirements.

The SPHS cohort contains clinical MRI data from human participants collected at the Second People’s Hospital of Shenzhen. This data is not publicly available because of patient privacy protections, institutional data-governance policies, ethical restrictions, and the scope of the approved institutional review board protocol. Access to de-identified SPHS-related data, including imaging data, segmentation annotations, and derived datasets, may be requested from the corresponding author. Requests should include a scientifically appropriate research proposal, documentation of institutional or ethics approval where applicable, and agreement to comply with institutional data-governance requirements and an appropriate data-use agreement. The source code, including scripts related to the federated learning and Glioma-pretrained FL workflow, has been deposited in a public GitHub repository: https://github.com/ChendongNi/FL_TL-FL-. The final trained model weights generated in this study have been deposited in a publicly accessible Google Drive folder: https://drive.google.com/drive/folders/1JSSd90zJqBxJNooJM4ygnVj54WabxYT_?usp=sharing.

Supplementary Table 1: Additional federated communication-round analysis. Case-averaged DSC and IoU obtained at different numbers of federated communication rounds for meningioma-pretrained FL and Glioma-pretrained FL on the external SPHS validation cohort.Please click here to download this file.

Discussion

The central finding of this study is that initialization quality—rather than the federated aggregation protocol itself—is the decisive factor in closing the cross-center generalization gap for privacy-preserving meningioma segmentation. While the centralized UMamba 2D model achieved strong in-domain performance on the BraTS2023-Men cohort (DSC = 0.8958), it suffered a substantial degradation when evaluated on the independent SPHS clinical dataset (Δ DSC = 0.1506). This performance drop underscores the vulnerability of models trained on curated benchmarks to compound domain shifts, encompassing both imaging-level discrepancies (e.g., scanner hardware, voxel spacing, and reconstruction protocols) and label shifts arising from divergent institutional conventions for defining tumor boundaries and dural tails. The centralized model’s error profile on the SPHS cohort was heavily skewed toward under-segmentation (FN = 14,873 vs. FP = 2,644), suggesting that it overfitted to BraTS-specific distributions and conservatively excluded peripheral or low-contrast tumor components that manifest differently in routine clinical acquisitions.

Our controlled comparison reveals that standard federated adaptation starting from a within-domain warm start is insufficient to overcome this generalization gap. Meningioma-pretrained FL yielded a lower external DSC (0.7122) than the centralized baseline, accompanied by a greater number of false-negative errors. From a systems perspective, this strategy was designed as a controlled comparison, loosely resembling a clinical consortium, in which one participating center uses an existing preliminary model to initiate federated collaboration. Its lower performance compared to centralized training indicates that the local-update-and-aggregation dynamics of standard FedAvg, when strictly seeded within the target domain, may produce an optimization trajectory that converges to a more conservative, less transferable local minimum rather than successfully mitigating cross-client label shifts. Conversely, Glioma-pretrained FL delivered the highest external performance among the final comparison models (DSC = 0.7503, IoU = 0.6301) and the most balanced error profile. By pretraining on a related brain-tumor source domain, the model effectively learned robust intermediate representations of multimodal contrast-enhancing tissue embedded within normal parenchyma. This cross-tumor feature reuse likely improved adaptation efficiency and provided a superior starting point for subsequent meningioma-side federated optimization, enabling the model to achieve external performance comparable to centralized training on fully unseen external data.

These results are consistent with and extend existing findings on federated learning and cross-tumor pretraining for brain tumor segmentation. Prior work has demonstrated the feasibility of FedAvg for multi-institutional glioma segmentation under true multi-site deployment³². The present results reveal that, even with multi-client architecture, standard FedAvg with a within-domain warm start does not automatically guarantee superior external generalization compared to centralized training. Furthermore, the observed benefit of glioma-to-meningioma pretraining extends recent cross-tumor pretraining findings, which showed improved meningioma segmentation when nnU-Net was pretrained on glioma data in limited-data settings²³. Our findings demonstrate that combining cross-tumor pretraining with federated adaptation on a full-cohort training set yields improved external performance relative to either strategy in isolation. This aligns with the domain-similarity framework³³, supporting the premise that representation transferability scales with the overlap of learned intermediate representations at the brain-parenchymal level.

The clinical and operational implications of these results are substantial, particularly for precision radiotherapy and longitudinal surveillance. In stereotactic radiosurgery (SRS), where steep dose gradients (30–50 Gy/cm) are delivered to the meningioma margin, systematic boundary under-segmentation poses a direct risk of marginal miss and adjacent critical structure toxicity⁴. Moreover, the false-positive burden observed in standard models would still be clinically problematic if such contours required substantial manual revision, thereby negating the practical efficiency benefit of automated segmentation. For conservatively managed grade-I meningiomas, the substantial inter-site volumetric bias (Δ DSC ≈ 0.15) would severely confound longitudinal growth assessments across different diagnostic centers². The Glioma-pretrained FL framework mitigates these clinical risks while strictly satisfying data residency regulations (e.g., GDPR, PIPL) by transmitting only model parameters. However, for eventual production deployment, this framework should be augmented with secure aggregation³⁴ or differential privacy³⁵ to prevent gradient inversion attacks³⁶.

Several methodological limitations of the current study must be acknowledged. First, to ensure a controlled comparison of initialization pathways, the federated clients were simulated via equal partitioning of the public BraTS2023-Men dataset rather than utilizing truly independent, institution-siloed datasets. Consequently, this experimental design does not capture the severe non-IID data heterogeneity typical of real-world multi-site deployments. Second, external validation was limited to a single clinical cohort (SPHS); validating across a broader array of heterogeneous scanner platforms and annotation protocols is necessary to confirm the robustness of Glioma-pretrained FL. Third, our federated pipeline used a standard FedAvg protocol with three communication rounds, and the exact mechanisms driving the suboptimal trajectory of meningioma-pretrained FL remain interpretable due to the lack of explicit client-wise feature distribution visualization. An additional sensitivity analysis examining the effect of increasing the number of communication rounds on external segmentation performance is provided in Supplementary Table 1. Finally, the harmonization-oriented preprocessing applied to the SPHS cohort (e.g., explicit brain extraction, rigid registration, and isotropic resampling) was necessary to bridge fundamental format discrepancies, but it may have partially masked native cross-site differences, complicating the isolation of intrinsic acquisition shifts from preprocessing-induced alignment.

Future investigations should address these limitations by deploying the Glioma-pretrained FL framework across genuinely heterogeneous, multi-national institutional nodes. From an algorithmic perspective, transitioning from standard FedAvg to more advanced aggregation schemes—such as FedProx³⁷, which penalizes local divergence, or SCAFFOLD²² for variance reduction—could further stabilize convergence under severe label shift. Additionally, incorporating modality-specific intensity harmonization prior to training, along with staged fine-tuning strategies that gradually unfreeze decoder layers, may alleviate initial under-segmentation transients. Finally, developing client outlier-detection mechanisms capable of identifying and dynamically downweighting institutions with highly divergent annotation conventions will be critical for scaling this privacy-preserving framework to broader neuro-oncology networks.

Meningioma-pretrained FL did not reliably improve cross-center generalization for multi-modal MRI meningioma segmentation and failed to reach the centralized baseline under the current federated training setup. Glioma-pretrained FL yielded the highest external DSC (0.7503) among the three final comparison models while preserving data residency. These findings establish initialization quality as a decisive factor for privacy-preserving cross-center brain tumor segmentation and support the use of Glioma-pretrained FL in future multi-institutional deployments.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
BraTS2023-Gli dataset	ASNR-MICCAI BraTS Challenge / Synapse (Sage Bionetworks)	N/A	Public glioma MRI dataset used for glioma-pretrained initialization. In this study, 450 cases were used under the binary tumor-versus-background formulation. https://www.synapse.org/Synapse:syn51156910/wiki/622351
BraTS2023-Men dataset	ASNR-MICCAI BraTS Challenge / Synapse (Sage Bionetworks)	N/A	Public meningioma MRI dataset used as the source cohort for model development and internal evaluation. In this study, 450 cases were used. https://arxiv.org/abs/2305.07642
DeepBet	Academic software by Fisch et al.	N/A	Brain extraction tool applied to the T1-weighted image to generate a brain mask for the SPHS cohort. https://arxiv.org/abs/2308.07003
MONAI	Project MONAI	Version 1.3	Medical imaging AI framework used for data loading and deep learning workflow implementation. https://docs.nvidia.com/monai/index.html
NVIDIA GeForce RTX 3090 GPU	NVIDIA	GeForce RTX 3090	Graphics processing unit with 24 GB GDDR6X memory used for model training and computation. https://www.nvidia.com/en-me/geforce/graphics-cards/30-series/rtx-3090-3090ti/
PyTorch	PyTorch Foundation, Linux Foundation	Version 2.0.1	Deep learning framework used to implement and train the UMamba 2D segmentation models. https://pytorch.org/foundation/
Python	Python Software Foundation	Version 3.10	Programming environment used for model training, preprocessing, evaluation, and statistical analysis. Replace with the exact Python version if known. https://www.python.org/psf-landing/
SPHS-Men clinical cohort	Second People’s Hospital of Shenzhen	N/A	External clinical multi-modal MRI cohort used for independent evaluation. A total of 174 cases were included.
UMamba 2D model implementation	Bowang-lab / U-Mamba authors	U-Mamba 2D configuration	State space model-based segmentation architecture used for binary meningioma segmentation. https://u-mamba.github.io/