Privacy-Preserving Federated Brain Tumor Segmentation using nnU-Net and Swin-UNETR across Multiple Institutions

Privacy-Preserving Federated Brain Tumor Segmentation using nnU-Net and Swin-UNETR across Multiple Institutions Mohammad Aman Ullah Khan Department of Computer Science and Engineering, School of Data and Sciences, BRAC University Dhaka, Bangladesh mohammad.aman.ullah.khan.pi @gmail.com Dr. Syed Khasim School of Computer Science and Engineering (SCOPE), VIT-AP University Amaravati, Andhra Pradesh, India [email protected] Irfan Sadiq Rahat School of Computer Science and Engineering (SCOPE), VIT-AP University Amaravati, Andhra Pradesh, India [email protected] Abstract—Medical image segmentation underpins computer-assisted monitoring, diagnosis, and treatment by delineating clinically salient regions of interest (e.g., tumors and lesions). We study brain-tumor segmentation in a Federated Learning (FL) setting to enable multi-institutional collaboration without exchanging raw data. Our framework evaluates two strong backbones: nnU-Net (an auto-configuring U-Net–family successor) and Swin-UNETR (a hierarchical transformer encoder with a U-Net–style decoder) within privacy-preserving training. Client data remain local; only model updates are aggregated, and dataset-level protections (e.g., anonymization) further reduce disclosure risk. We analyze the privacy–utility trade-off using standard segmentation metrics, with the Dice coefficient as the primary measure, and demonstrate that competitive performance can be achieved while adhering to stringent privacy constraints. Keywords—Federated Learning; nnU-Net; Swin-UNETR; Brain Tumor Segmentation; Dice coefficient; Privacy-preserving learning I. INTRODUCTION Medical image segmentation is fundamental to computer-assisted analysis in healthcare, underpinning monitoring, diagnosis, and treatment planning by accurately delineating regions of interest (ROI) such as tumors and lesions to characterize spatial extent and morphology that inform clinical decisions [1]. High-quality segmentation directly supports precise intervention strategies and better understanding of disease progression, thereby improving patient outcomes.Deep learning (DL) has substantially advanced segmentation by learning hierarchical representations directly from images [2]. Yet, performance is tightly coupled to the quality, diversity, and scale of training data [3]. Curating large, expertly annotated datasets is difficult and costly, since labeling demands specialized clinical expertise and sustained time investment [4]. Moreover, many state-of-the-art systems are trained and validated on data from a small number of sites or devices, which can introduce demographic and acquisition biases that degrade model reliability when deployed across different populations or scanners [5]. Incorporating diverse, multi-institutional datasets is therefore essential to capture heterogeneity in disease patterns including rare or complex presentations and to improve generalization and robustness of DL-based segmentation [6].However, centralizing sensitive medical data raises substantial legal, ethical, and technical concerns, including privacy, governance, and data-ownership barriers that often render conventional data pooling impractical [7]. Federated Learning (FL) addresses these constraints by training models locally at each institution and sharing only model updates with a coordinating server that aggregates them into a global model; raw data never leave the premises. This iterative train–aggregate–redistribute cycle preserves privacy while enabling collaborative learning across sites (the FL system architecture is shown in Figure 1) [8]. In this work, we study brain-tumor segmentation under FL using two strong modern backbones: nnU-Net (an automated, self-configuring successor within the U-Net family) and Swin-UNETR (a hierarchical transformer encoder paired with a U-Net–style decoder) to examine privacy–utility trade-offs and cross-site robustness under realistic, multi-institutional constraints. This study investigates Federated Learning (FL) for medical image segmentation with particular emphasis on brain-tumor segmentation. We evaluate two strong modern backbones: nnU-Net and Swin-UNETR within an FL framework and compare them against conventional centralized (pooled-data) training and single-site baselines. Using multiple multi-institutional datasets, we analyze convergence behavior, cross-site generalization, and the privacy–utility trade-off. To our knowledge, this constitutes one of the first systematic, multi-dataset evaluations of nnU-Net and Swin-UNETR under federated training for brain-tumor segmentation, providing a practical foundation for privacy-preserving, collaborative medical imaging. The paper is organized as follows: Section II reviews related work, Section III details the proposed federated framework, Section IV presents experimental results, and Section V concludes. XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE II. RELATED WORK Early work established the feasibility of Federated Learning (FL) for medical image segmentation while surfacing core systems challenges. Sheller et al. introduced FL for brain-tumor segmentation across many institutions and demonstrated privacy-preserving collaborative training using a U-Net–family baseline (referred to in our manuscript as nnU-Net) [9]. Subsequent analyses emphasized the communication overhead of frequent client–server synchronization and the need for compression/efficient aggregation, as well as the impact of non-IID client distributions on convergence and accuracy [10, 11].Privacy-preserving mechanisms have been integrated into FL pipelines for neuro-oncology segmentation. Differential privacy applied to BraTS-style workflows highlighted a clear privacy–utility trade-off, where stronger perturbations protect patients but degrade Dice and boundary fidelity [12, 13]. Beyond the brain, FL has been explored for abdominal organs. For pancreas segmentation, a coarse-to-fine NAS approach with a VAE branch improved cross-site generalization but increased compute on both clients and server, complicating deployment at scale [14, 15]. For prostate MRI, FL with anisotropic 3D hybrids achieved gains over single-site or naïve pooled baselines, yet concerns remained around scalability (few sites) and imbalanced client data [16, 17]. Liver segmentation studies combining FL with lightweight encoders reported accuracy drops when strong privacy noise was enforced underscoring the sensitivity of dense prediction to gradient perturbations [18].Domain generalization techniques have been adapted to FL to counter inter-site shift (scanner, protocol, population). Episodic or frequency-space curricula improved robustness but did not fully close the generalization gap under severe heterogeneity [19]. For multi-organ abdominal CT, federated multi-encoder designs leveraged partially labeled datasets across clients, yet client-contribution merging and label sparsity introduced instability and sensitivity to data quality [20, 21]. Addressing non-IID aggregation, cross-learning and related strategies improved averages but still struggled when client distributions diverged markedly (e.g., extreme class imbalance, device drift) [22]. Cross-hospital evaluations in cardiac CT confirmed FL’s practicality for privacy-aware collaboration while again revealing performance variability that tracks institutional differences [23]. Fairness-aware FL with client-contribution estimation reduced bias against under-represented clients, though residual privacy risks and communication cost persisted [24]. Synthesis and positioning. Taken together, prior work shows FL can enable privacy-preserving, multi-institutional segmentation but faces persistent obstacles: communication efficiency, non-IID robustness, fairness, and privacy-accuracy trade-offs. Moreover, most pipelines rely on classic U-Net baselines. In contrast, our study adopts two modern backbones: nnU-Net (self-configuring U-Net–family successor) and Swin-UNETR (hierarchical transformer encoder with a U-Net–style decoder, used here as the replacement for U-Net++-style designs) to test whether stronger architectures translate into better FL behavior under realistic heterogeneity [25].Gaps our work addresses. First, there has been no systematic, multi-dataset evaluation of nnU-Net and Swin-UNETR under federated training for brain-tumor segmentation; we provide such an analysis and report convergence, cross-site generalization, and privacy–utility trends [26]. Second, rigorous head-to-head comparisons between FL and centralized (pooled-data) training remain sparse in segmentation; we quantify the relative trade-offs to guide deployment decisions [27].Third, prior studies often involve few institutions or organ-specific cohorts; we broaden the scope and probe scalability and client heterogeneity, including the impact of model choice on client contribution and aggregation stability under non-IID data [28]. III. PROPOSED FEDERATED LEARNING FOR MULTI-INSTITUTIONAL MEDICAL IMAGE SEGMENTATION We propose a federated framework for medical image segmentation with emphasis on brain-tumor delineation across heterogeneous institutions. In place of legacy U-Net baselines, our study evaluates nnU-Net (a self-configuring U-Net–family successor) and Swin-UNETR (a hierarchical transformer encoder coupled to a U-Net–style decoder); Attention U-Net is included as an auxiliary comparator to probe the effect of attention gating under federated training. We benchmark these backbones in federated versus centralized (pooled-data) and single-site settings to assess convergence, cross-site generalization, and privacy–utility trade-offs. The overarching objective is to achieve segmentation quality that is competitive with centralized training while preserving patient confidentiality and demonstrating scalable, multi-institutional collaboration. A. Federated Learning Federated Learning paradigm. Unlike conventional centralized learning where collaborators upload all data to a hub for joint training Federated Learning (FL) keeps data in place. Each institution (client) trains the model locally on its private dataset and transmits only parameter updates (e.g., weight deltas or gradients) to a coordinating server fig 1. The server aggregates these updates into a new global model and redistributes it to clients for the next training round. This iterative train–aggregate–redistribute cycle continues until convergence. To minimize exposure, the server retains only the current global parameters; transient client updates are discarded once applied. A. Centralized Learning B. Federated Learning Fig. 1. Depicts the overall system System components and training cycle. In our setup, collaborators retain full control of their datasets and perform local optimization for multiple epochs per round using identical code and hyperparameter templates, with institution-specific batch sizes allowed for memory constraints. A central parameter server orchestrates rounds, initializes the global model, schedules clients, and disseminates parameters. The aggregator implements weighted federated averaging, where client contributions are proportionate to effective sample size; it also supports partial participation when some clients are offline. A round is defined as (i) local training at selected clients, followed by (ii) secure upload of model updates and (iii) server-side aggregation. This abstraction accommodates non-IID data, device variability, and intermittent connectivity typical of clinical partners. Privacy, trust, and practicality. FL reduces legal and operational burdens by eliminating raw-data transfers and restricting the attack surface to endpoints rather than a central repository. We assume an honest-but-curious coordinator (standard in clinical FL) and enforce encrypted transport and ephemeral storage for updates on the server. Where required by governance, differential-privacy noise or secure aggregation can be layered without altering the learning protocol; we report the privacy–accuracy trade-offs separately. Communication efficiency is addressed via periodic local epochs per round and optional update sparsification, balancing wall-clock time with statistical efficiency. Backbone choices in FL. nnU-Net is adopted for its data-driven configuration (preprocessing, patching, deep supervision) that tends to yield robust performance under domain shift without manual retuning, a desirable property in multi-site FL. Swin-UNETR contributes long-range context through shifted-window self-attention while preserving precise boundaries via a U-Net–style decoder; this combination is well-suited to heterogeneous scanners and protocols. Attention U-Net serves as a lightweight attention baseline to contextualize gains from transformer hierarchies and nnU-Net’s automated design. Collectively, these models allow us to test whether stronger architectural priors translate into improved federated behavior under realistic heterogeneity. By jointly analyzing federated versus centralized training across institutions, we aim to establish whether nnU-Net and Swin-UNETR can deliver centrally competitive Dice and boundary metrics while honoring privacy constraints, and to delineate the operational conditions (round length, client participation, weighting, and update policy) under which federated segmentation is both accurate and practical for large-scale, real-world deployment. B. Training Process At the start of each communication round r, the coordinator disseminates the current global parameters θ(r) to the available collaborators (partial participation allowed). Each site then performs E local epochs of stochastic optimization on its private dataset using a common training recipe 𝐷𝑖 (optimizer, scheduler, augmentations) and returns only model deltas or compressed weights over encrypted ∆(𝑟)𝑖 channels, no raw images or labels ever leave the institution. The server constructs the next global model by weighted aggregation (e.g., FedAvg with weights ; optional α |𝐷𝑖| momentum/corrective terms), applies update clipping and, where required, secure aggregation and/or differential privacy to bound leakage, and then broadcasts θ(r+1) back to clients. This train–aggregate–broadcast loop repeats until a cross-site validation criterion triggers early stopping, progressively distilling complementary signals from heterogeneous scanners and populations while preserving data locality and governance. Practicalities such as straggler tolerance, variable batch sizes, and round length are handled by allowing site-specific compute budgets and scheduling, which improves wall-clock efficiency without altering the privacy guarantees. C. nnU-Net Architectures nnU-Net is a self-configuring U-Net–family framework rather than a single fixed model: given a dataset “fingerprint” (voxel spacing, image size, class balance, anisotropy, memory budget), it automatically derives preprocessing (resampling/normalization), patch size and batch size, network depth and kernel sizes, and training/inference rules. Architecturally, it instantiates a 2D U-Net, a 3D full-resolution U-Net, or a 3D cascade (coarse low-res network feeding a full-res refiner), each built as an encoder–decoder with symmetric skip connections, strided-convolution downsampling and transposed-convolution upsampling, residual/conv blocks with InstanceNorm and LeakyReLU, and deep supervision via auxiliary outputs at multiple decoder scales to stabilize optimization and mitigate class imbalance. Channel widths typically double per resolution level (e.g., 32 → 64 → 128 → 256 → 512, dataset-dependent). Training uses a composite Dice + cross-entropy loss with strong spatial/intensity augmentations; inference is sliding-window with Gaussian blending and optional test-time ensembling across folds or across 2D/3D/cascade variants selected by the planner. This automated design consistently yields robust, high-quality segmentations across heterogeneous scanners and institutions without manual returning an advantage in federated, multi-site workflows. Contracting Path (Encoder) The contracting path is similar to a convolutional neural network (CNN) used for classification tasks. It applies a series of convolutional layers to extract high-level features, gradually reducing the spatial dimensions of the input image. This path includes repeated application of the following steps: ● Convolutions: Two 3x3 convolutions, followed by a rectified linear unit (ReLU) activation function, which can be represented mathematically as: (1) 𝑓𝑥()=𝑅𝑒𝐿𝑈(𝑊*𝑥+𝑏) where W is the filter, x is the input, b is the bias term, as well as ∗ denotes the convolution operation. ● Max-pooling: After each convolution, a 2x2 max-pooling operation is applied to downsample the image by reducing its spatial resolution: (2) 𝑃𝑥()=(𝑥𝑖,𝑗) which reduces the feature map size while retaining important features. Expansive Path (Decoder) The expansive path mirrors the contracting path but works in reverse, progressively upsampling the feature maps to reconstruct the spatial dimensions of the original input while refining the segmentation mask. This path consists of the following: Fig. 2. Up-convolutions (also called transposed convolutions or deconvolutions): These increase the spatial resolution by learning how to upsample, and are applied as: (3) 𝑓𝑥()=𝑊𝑇*𝑥 where represents the learned filters for 𝑊𝑇 upsampling. Fig. 3. Concatenation: For each upsampling step, the corresponding high-resolution feature map from the contracting path is concatenated with the upsampled output. This allows the model to retain high-level information while also incorporating spatial details. Fig. 4. Convolutions: Similar to the contracting path, the concatenated feature maps undergo two 3x3 convolutions with ReLU activation. Final Layer At the final stage, a 1x1 convolution is applied to map the multi-channel feature map to the desired number of classes (e.g., the segmentation mask), represented by: (4) 𝑓𝑥()=𝑊*𝑥 where W is a 1x1 convolutional filter. Symmetry and U-Shape The architecture’s symmetry resembling the letter “U” enables the integration of fine details from the input image with global contextual information, making it particularly effective in medical image segmentation tasks where precise localization of features (e.g., tumors) is critical. This combination of downsampling (contracting path) and upsampling (expansion path) enhances the model's ability to capture both high-level and detailed features, improving segmentation accuracy. D. SWIN UNETR Architectures Swin-UNETR is a hierarchical Transformer–U-Net hybrid for (primarily 3D, also 2D) medical image segmentation that couples a Swin Transformer encoder patch embedding followed by multi-stage shifted-window self-attention (W-MSA/SW-MSA) with patch-merging downsamples and channel doubling with a U-Net–style decoder that upsamples feature maps and fuses them via long skip connections originating from each encoder stage (UNETR pathway). The encoder uses LayerNorm and GELU, capturing long-range dependencies efficiently thanks to localized windows and relative position bias, while preserving inductive bias for locality. The decoder employs convolutional refinement (transpose-conv or upsample+conv blocks) to recover fine boundaries; optional deep supervision can stabilize training on class-imbalanced lesions. Typical training uses Dice + cross-entropy (or DiceFocal) with AdamW, cosine LR warm-up, mixed precision, and strong spatial/intensity augmentations; inference relies on sliding-window with overlap and blending. By combining global context modeling with U-Net skip fusion and scaling from Tiny/Base backbones Swin-UNETR delivers robust cross-site generalization and boundary accuracy under heterogeneous scanners and protocols, making it well-suited to federated, multi-institutional segmentation. These operations are: 1 Convolutions: (5) 𝑓𝑥()=𝑅𝑒𝐿𝑈𝑊*𝑥+𝑏( ) This applies to both the intermediate and original convolution layers, refining features progressively. 2 Skip Connection Formula: In U-Net++, the output of each skip connection is: ) (6) 𝑥𝑖,𝑗=𝑓(𝑥𝑖−1,𝑗−1,𝑥𝑖,𝑗−1 where is the feature map at the i-th layer and j-th node in 𝑥𝑖,𝑗 the skip connection grid, and f is the function applied via the convolutional layer.By reusing previous layers' features and adding dense layers, Swin-UNETR reduces semantic gaps between encoder and decoder paths, improving segmentation performance in scenarios where the original U-Net may struggle. E. Attention U-Net Architectures Attention U-Net extends the standard encoder–decoder U-Net with attention gates (AGs) on the skip connections to suppress background responses and amplify task-relevant structures, especially small or low-contrast lesions. At each decoder stage, a gating signal derived from high-level decoder features is compared with the corresponding encoder feature map to compute attention coefficients (via a lightweight additive gating module); these coefficients reweight the encoder activations so that only salient information is forwarded to the decoder for fusion. By filtering skip features before concatenation, Attention U-Net improves boundary fidelity and reduces false positives with modest parameter overhead and no changes to the overall training protocol yielding higher segmentation accuracy and efficiency in medical imaging contexts. Introduced in, it remains a strong baseline when fine-grained focus and robustness to clutter are required. Attention U-Net builds on the canonical encoder–decoder (U-Net–family) design: a contracting path of convolutional blocks with pooling progressively reduces spatial resolution while increasing channels, a high-capacity bottleneck captures compressed context, and an expansive path upsamples features and fuses them with encoder skip maps to recover detail. To suppress irrelevant regions, attention gates (AGs) are inserted on the skip connections of each decoder stage. Each AG receives a gating signal from the decoder (high-level context) and the corresponding encoder feature map; a lightweight gating mechanism computes attention coefficients that weight the encoder features, attenuating background and enhancing task-relevant structures. The filtered encoder features are then concatenated with the upsampled decoder features and refined by convolutions, allowing multi-scale, context-aware fusion that preserves salient boundaries while reducing noise. a) Datasets: The BraTS 2020 dataset comprises multi-institutional pre-operative brain MRI from 19 institutions, provided as NIfTI (.nii.gz) volumes per subject with four sequences: native T1, post-contrast T1Gd, T2, and T2-FLAIR. Expert annotations delineate three regions: GD-enhancing tumor (ET), peritumoral edema (ED), and the necrotic/non-enhancing tumor core (NCR/NET). The scans were acquired under diverse clinical protocols and scanner vendors, yielding realistic inter-site heterogeneity that is well suited for federated segmentation studies. b) Mechanism: Before federation begins, the dataset is split into disjoint train/validation/test partitions; the training split is then partitioned across n clients to represent collaborating institutions (with site-aware/non-IID shards when applicable). Figure 2: System Architecture of Federated Learning for Multi-Institutional Medical Image Segmentation We instantiate two modern backbones: nnU-Net (via its reference self-configuring pipeline) and Swin-UNETR (via MONAI) as the shared global architectures. The server initializes the global parameters and broadcasts them to available clients; each client trains locally for E epochs on its private shard using a common recipe (optimizer, scheduler, augmentations), returning only encrypted model deltas (no images/labels). The coordinator performs secure fig 2 FedAvg, weighting updates by effective sample size (with optional clipping or momentum/correction), updates the global model, and redistributes the new parameters to clients. This train–aggregate–broadcast loop repeats for R rounds with partial participation tolerated, and early stopping triggered by cross-site validation. Throughout, client updates are ephemeral on the server and transport is encrypted, preserving privacy while progressively distilling complementary signals from heterogeneous scanners and populations. Algorithm 1: Federated Learning 1. Initialize the global model. 2. Set the number of federated rounds n. 3. For each federated round (1 to n): ○ Retrieve the global model’s weights. ○ Create m clients. 4. For each client (1 to m): ○ Initialize the local model. Set the local model’s weights equal to the global model’s weights. ○ Train the local model using the client’s local data. ○ Store the updated local model’s weights. ○ Clear the session. ○ Aggregate the weights from all local models. ○ Update the global model with the aggregated weights. 5. End for. c) Benchmarking Metric: The quantitative performance evaluation metric for all the medical image segmentation tasks is the Dice Coefficient (DC). The Dice Similarity Coefficient or simply the Dice Coefficient is a similarity measure ranging from 0 which indicates no overlap and 1 which indicates a perfect overlap. IV. RESULT AND DISCUSSION We simulated a federated learning (FL) environment by implementing the client–server training loop in Python with Keras/TensorFlow, keeping all preprocessing and training code identical across sites. The experimental dataset is multi-institutional, introducing natural heterogeneity across contributors. Each dataset was partitioned into train/validation/test = 80/10/10 (patient-level separation to prevent leakage); the training portion was then split into three non-overlapping shards and assigned to three clients C1,C2,C3 to emulate collaborating institutions. Before sharding, the training data were randomly shuffled (with a fixed seed) and, where labels permitted, stratified to approximately preserve class proportions across clients. This setup yields realistic inter-site variation while ensuring that each client holds a representative local sample for federated optimization. A. Evaluation Metrics We analyze the FL setup by first partitioning the dataset into train/validation/test = 80%/10%/10% (patient-level to avoid leakage); for a total of NNN samples, the splits contain 0.80N0.80N0.80N, 0.10N0.10N0.10N, and 0.10N0.10N0.10N items (integer-rounded) respectively. The training split is then divided evenly into three client shards to emulate institutions; for example, with N=249N=249N=249 subjects, each client receives 83 subjects (249/3=83)(249/3=83)(249/3=83). Before sharding, we apply a fixed-seed random shuffle and, where labels permit, stratification to preserve class/site proportions. Representativeness is verified quantitatively by comparing each shard’s feature and label distributions to the global training distribution, e.g., monitoring class-proportion deltas, scanner/site frequency differences (χ² tests), and distribution distances for continuous features (two-sample KS tests or Jensen–Shannon divergence), using false-discovery-rate control for multiple comparisons to ensure each client holds a statistically comparable slice of the data. B. Performance Metrics Analysis We compare nnU-Net, Swin-UNETR, and Attention U-Net using Accuracy, optimization Loss, and Dice Coefficient (Figures 3-5). Accuracy (defined as correct predictions ÷ total predictions) is highest for Swin-UNETR: 97.15%, followed by nnU-Net: 96.71% and Attention U-Net: 96.32%. For the optimization loss (cross-entropy), nnU-Net attains the lowest value (0.015), edging Swin-UNETR: 0.017 and clearly outperforming Attention U-Net: 0.038. Segmentation overlap, measured by the Dice coefficient DC = 2∣P∩T∣/(∣P∣+∣T∣), is best for Swin-UNETR: 0.78, with Attention U-Net: 0.65 and nnU-Net: 0.57 trailing (Fig. 5). In short, Swin-UNETR delivers the strongest overlap (Dice) and top accuracy, whereas nnU-Net minimizes training loss; Attention U-Net remains competitive but consistently below the transformer-based backbone in overlap. Accuracy is defined as: 𝐴𝑐𝑐= 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐶𝑜𝑟𝑟𝑒𝑐𝑡 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠 𝑇𝑜𝑡𝑎𝑙 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠 𝐷𝐶=2×∣𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑∩𝑇𝑟𝑢𝑒| ∣𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑∣+∣𝑇𝑟𝑢𝑒∣ Figure 3. Accuracy by Model after 5 Federated Rounds (Brain-Tumor Segmentation) Figure 4. Optimization Loss by Model after 5 Federated Rounds (lower is better) Figure 5. Dice Coefficient by Model after 5 Federated Rounds (higher is better) C. Comparison of Results For brain-tumor segmentation, the 249-subject cohort was split 80/10/10 into train/validation/test and then partitioned evenly across three clients (249/3 = 83 per site), with an identical training recipe at each site (batch size = 32, learning rate = 1e-4, Adam). After five federated rounds, Swin-UNETR delivered the strongest segmentation quality 97.15% accuracy and 0.78 Dice consistent with its ability to capture long-range context while preserving fine boundaries; nnU-Net achieved competitive accuracy (96.71%) and the lowest loss (0.015), reflecting stable optimization and effective deep supervision, though its Dice (0.57) suggests less precise overlap under cross-site heterogeneity; Attention U-Net reached 96.32% accuracy, 0.038 loss, and 0.65 Dice, sitting between the two. Practically, when clinical delineation fidelity is paramount (e.g., tumor burden quantification or therapy planning), Swin-UNETR is the preferred backbone; when optimization stability and ease of configuration are prioritized (e.g., rapid deployment across new sites), nnU-Net is attractive. The divergence between loss and Dice underscores that minimizing the training objective does not always translate into superior overlap under non-IID data, a common reality in federated settings Table1-2. With additional rounds or modest client-specific tuning (e.g., augmentation strength, Dice-centric losses), we expect Swin-UNETR’s overlap advantage to persist, while nnU-Net may close some of the gap, leaving model choice to the application’s trade-off between segmentation accuracy and training/operational stability. TABLE I. HYPERPARAMETERS FOR BRAIN TUMOR SEGMENTATION Batch Size 32 Learning Rate 1e-4 Optimizer Adam TABLE II. COMPARISON OF U-NET VARIANTS PERFORMANCE AFTER 5 FEDERATED ROUNDS FOR BRAIN TUMOR SEGMENTATION Metric nnU-Net Swin-UNETR Attention U-Net Accuracy 96.71 97.15 96.32 Loss 0.015 0.017 0.038 Dice coefficient 0.57 0.78 0.65 V. CONCLUSION AND FURTHER WORK This study demonstrates that Federated Learning (FL) enables privacy-preserving, multi-institutional brain-tumor segmentation without centralizing patient data. Across three clients and five rounds, the transformer-based Swin-UNETR achieved the strongest segmentation quality (highest accuracy and Dice), while nnU-Net delivered the lowest loss and stable optimization highlighting a practical trade-off between overlap fidelity and training stability in federated settings. By aggregating updates rather than raw images, FL harnesses heterogeneity across institutions to improve generalization under non-IID data, while standard safeguards (encrypted transport, ephemeral server state) keep the attack surface small. Nevertheless, model updates can still leak information; thus, privacy protection must be treated as a first-class objective alongside utility.We will integrate formal privacy accounting with client-side differential privacy and secure aggregation, and study their utility cost on Dice/HD95 under realistic participation and network conditions. Robust aggregation and personalization under non-IID data (e.g., FedProx/ScaFFold-style corrections, FedBN, calibration/uncertainty estimation) merit systematic evaluation, as do communication-efficient schemes (update sparsification/quantization and adaptive round lengths). We also plan larger-scale, multi-organ studies with more clients and rounds, federated hyperparameter tuning, and prospective clinical validation to assess workflow impact. Finally, we will release code and standardized splits to support reproducibility and fair comparison of nnU-Net and Swin-UNETR under privacy-aware federation. ACKNOWLEDGEMENT The authors would like to thank all the institutions and collaborators who contributed to the collection and preparation of the brain tumor imaging data. We also acknowledge the computational resources and technical support that facilitated the development and evaluation of the federated-learning framework. Finally, we are grateful for the constructive feedback and guidance provided by colleagues and reviewers during the preparation of this manuscript. AUTHOR CONTRIBUTIONS Mohammad Aman Ullah Khan led the overall conceptualization and methodology of the study, designed the federated-learning framework, conducted model training and evaluation for nnU-Net and Swin-UNETR, coordinated the project, and drafted the initial manuscript. Dr. Sayed Khasim assisted with the software pipeline, data curation, and preparation of figures and visualizations. Irfan Sadiq Rahat provided supervision and resources, guided formal analysis and validation, and contributed critical revisions and editing. All authors contributed to the interpretation of results, revised the manuscript, approved the final version, and agreed to be accountable for all aspects of the work. DATA AVAILABILITY The dataset used in this study is not publicly available due to privacy and data-sharing constraints but can be provided by the corresponding author upon reasonable request. DECLARATIONS Clinical Trial Number: Not applicable. Human Ethics and Consent to Participate declarations: Not applicable. Consent to Publish declaration: Not applicable. Consent to Participate declaration: Not applicable. Ethics declaration: Not applicable. Funding Declaration: This research received no external funding. REFERENCES [1] A. Esteva et al., “A guide to deep learning in healthcare,” Nature Medicine, vol. 25, no. 1, pp. 24–29, Dec. 2018, doi: 10.1038/s41591-018-0316-z. [2] G. A. Kaissis, M. R. Makowski, D. Rückert, and R. F. Braren, “Secure, privacy-preserving and federated machine learning in medical imaging,” Nature Machine Intelligence, vol. 2, no. 6, pp. 305–311, Jun. 2020, doi: https://doi.org/10.1038/s42256-020-0186-1. [3] P. Kairouz et al., “Advances and Open Problems in Federated Learning,” arXiv:1912.04977 [cs, stat], Dec. 2019, Available: https://arxiv.org/abs/1912.04977 [4] G. Litjens et al., “A Survey on Deep Learning in Medical Image Analysis,” Medical Image Analysis, vol. 42, no. 1, pp. 60–88, Dec. 2017, doi: https://doi.org/10.1016/j.media.2017.07.005. [5] A. Selvikvåg Lundervold and A. Lundervold, “An overview of deep learning in medical imaging focusing on MRI,” Zeitschrift für Medizinische Physik, vol. 29, no. 2, Dec. 2018, doi: https://doi.org/10.1016/j.zemedi.2018.11.002. [6] N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan, “A Survey on Bias and Fairness in Machine Learning,” ACM Computing Surveys, vol. 54, no. 6, pp. 1–35, Jul. 2021, doi: https://doi.org/10.1145/3457607. [7] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” arXiv.org, May 18, 2015. https://arxiv.org/abs/1505.04597 [8] H. Guan, P.-T. Yap, A. Bozoki, and M. Liu, “Federated learning for medical image analysis: A survey,” Pattern recognition, pp. 110424–110424, Mar. 2024, doi: https://doi.org/10.1016/j.patcog.2024.110424. [9] M. J. Sheller et al., “Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data,” Scientific Reports, vol. 10, no. 1, p. 12598, Jul. 2020, doi: https://doi.org/10.1038/s41598-020-69250-1. [10] W. Li et al., “Privacy-Preserving Federated Brain Tumour Segmentation,” kclpure.kcl.ac.uk, Jan. 01, 2019. https://kclpure.kcl.ac.uk/portal/en/publications/privacyp reserving-federated-brain-tumour-segmentation(93e702 6d-927e-4b77-bd13-7463b161e910)/export.html (accessed Jan. 08, 2021). [11] P. Wang et al., “Automated Pancreas Segmentation Using Multi-institutional Collaborative Deep Learning,” arXiv.org, 2020. https://arxiv.org/abs/2009.13148 (accessed Dec. 03, 2025). [12] K. V. Sarma et al., “Federated learning improves site performance in multicenter deep learning without data sharing,” Journal of the American Medical Informatics Association: JAMIA, vol. 28, no. 6, pp. 1259–1264, Jun. 2021, doi: https://doi.org/10.1093/jamia/ocaa341. [13] X. Tang, A. Panda, V. Sehwag, and P. Mittal, “Differentially Private Image Classification by Learning Priors from Random Processes,” Journal of Privacy and Confidentiality, vol. 15, no. 1, Mar. 2025, doi: https://doi.org/10.29012/jpc.910. [14] Q. Liu, C. Chen, J. Qin, Q. Dou, and P.-A. Heng, “FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space,” Thecvf.com, pp. 1013–1023, 2021, Accessed: Dec. 03, 2025. [Online]. Available: https://openaccess.thecvf.com/content/CVPR2021/html/ Liu_FedDG_Federated_Domain_Generalization_on_M edical_Image_Segmentation_via_Episodic_CVPR_202 1_paper.html [15] X. Xu, H. H. Deng, J. Gateno, and P. Yan, “Federated Multi-Organ Segmentation With Inconsistent Labels,” IEEE Transactions on Medical Imaging, vol. 42, no. 10, pp. 2948–2960, Apr. 2023, doi: https://doi.org/10.1109/tmi.2023.3270140. [16] X. Xu et al., “Federated Cross Learning for Medical Image Segmentation,” arXiv.org, 2022. https://arxiv.org/abs/2204.02450 (accessed Dec. 03, 2025). [17] Thibaud Misonne and Sebastien Jodogne, “Federated learning for heart segmentation,” Digital Access to Libraries, pp. 1–5, Jun. 2022, doi: https://doi.org/10.1109/ivmsp54334.2022.9816345. [18] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A Nested U-Net Architecture for Medical Image Segmentation,” Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, vol. 11045, pp. 3–11, 2018, doi: https://doi.org/10.1007/978-3-030-00889-5_1. [19] O. Oktay et al., “Attention U-Net: Learning Where to Look for the Pancreas,” arXiv.org, 2018. https://arxiv.org/abs/1804.03999 [20] R. Wang, T. Lei, R. Cui, B. Zhang, H. Meng, and A. K. Nandi, “Medical image segmentation using deep learning: A survey,” IET Image Processing, vol. 16, no. 5, Jan. 2022, doi: https://doi.org/10.1049/ipr2.12419. [21] D. C. Nguyen et al., “Federated Learning for Smart Healthcare: A Survey,” arXiv:2111.08834 [cs, eess], Nov. 2021, Available: https://arxiv.org/abs/2111.08834 [22] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-Efficient Learning of Deep Networks from Decentralized Data,” proceedings.mlr.press, Apr. 10, 2017. https://proceedings.mlr.press/v54/mcmahan17a.html [23] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated Machine Learning: Concept and Applications,” ACM Transactions on Intelligent Systems and Technology, vol. 10, 2019, doi: https://repository.hkust.edu.hk/ir/Record/1783.1-95671. [24] Z. Liu et al., “Deep Learning Based Brain Tumor Segmentation: A Survey,” arXiv (Cornell University), Jan. 2020, doi: https://doi.org/10.48550/arxiv.2007.09479. [25] M. J. Sheller, G. A. Reina, B. Edwards, J. Martin, and S. Bakas, “Multi-institutional Deep Learning Modeling Without Sharing Patient Data: A Feasibility Study on Brain Tumor Segmentation,” Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, pp. 92–104, 2019, doi: https://doi.org/10.1007/978-3-030-11723-8_9. [26] V. Ummadi, “U-Net and its variants for Medical Image Segmentation : A short review,” arXiv (Cornell University), Apr. 2022, doi: https://doi.org/10.48550/arxiv.2204.08470. [27] B. H. Menze et al., “The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993–2024, Oct. 2015, doi: https://doi.org/10.1109/tmi.2014.2377694. [28] S. Bakas et al., “Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features,” Scientific Data, vol. 4, no. 1, Sep. 2017, doi: https://doi.org/10.1038/sdata.2017.117.