Two-Round Federated Learning Framework Classifies Brain Tumors With Over 98 Percent Accuracy

A research team has demonstrated a hybrid federated learning framework capable of classifying and segmenting brain tumors from magnetic resonance images with 98.73 percent diagnostic accuracy across simulated clinical institutions while completing distributed training in two communication rounds. Published in Frontiers in Artificial Intelligence by researchers from the Vellore Institute of Technology, the system, termed HOS-VFL, combines prototype-based one-shot aggregation with encrypted vertical feature space augmentation. This design reconciles the strict privacy demands of multi-center medical collaboration with the bandwidth limitations of distributed neural networks. By achieving diagnostic parity with centralized benchmarks that pooled all patient records, the protocol demonstrates that collaborative diagnostic models can overcome the conventional trade-off between institutional data isolation and distributed computational performance in medical image computing.

The clinical detection of brain tumors relies heavily on magnetic resonance imaging due to its superior non-invasive soft tissue contrast, yet interpreting complex intracranial scans remains a time-consuming, subjective task requiring specialized neuroradiological expertise. Primary intracranial neoplasms, of which gliomas represent roughly 45 percent alongside meningiomas and pituitary tumors, demand prompt detection to inform neurosurgical and oncological interventions. While deep neural networks have achieved diagnostic performance rivaling medical experts on benchmark collections, their conventional development requires pooling sensitive patient scans into centralized servers. In clinical settings, healthcare privacy legislation such as HIPAA in the United States and the GDPR in Europe places stringent restrictions on sharing raw patient records, creating institutional silos that hinder collaborative studies and restrict model generalizability.

Federated learning offers a decentralized paradigm that trains models across multiple institutions without exchanging underlying medical scans. In standard horizontal federated learning, participating centers possess datasets with similar feature spaces but different patient samples, training local models independently and repeatedly transmitting model updates to a central server. However, communicating high-dimensional neural network parameters over dozens or hundreds of iterations incurs severe bandwidth costs that challenge resource-constrained clinical networks. Vertical federated learning resolves a complementary operational challenge, operating where different healthcare facilities hold unique, non-overlapping feature spaces for an identical patient cohort. While vertical federated learning enriches model representations through feature-space augmentation, traditional implementations still require multi-round communication and present entity alignment vulnerabilities, leading the authors to develop an integrated protocol that unifies one-shot prototype compression with encrypted vertical feature aggregation.

The researchers validated their methodology using the Brain Tumor Image Segmentation and Classification dataset, designated BRISC2025. Containing 6,000 contrast-enhanced T1-weighted magnetic resonance imaging scans divided into 5,000 training scans and 1,000 test scans, the dataset addresses historical limitations of older benchmarks, including annotation inconsistencies and tumor class imbalances. The imagery is balanced equally across four diagnostic categories, allocating 25 percent each to gliomas, meningiomas, pituitary tumors, and non-tumorous cases across axial, coronal, and sagittal planes. Expert segmentations were performed and validated by certified radiologists, establishing an inter-rater agreement Cohen’s kappa coefficient of 0.89. The repository captures authentic imaging variations, including scanner field strengths of 1.5 Tesla and 3.0 Tesla, hardware from multiple manufacturers, and diverse pulse sequences, while maintaining strict separation between training and test sets without patient duplication.

To model multi-institutional collaboration, 5,000 training scans were partitioned across five simulated institutions under balanced and heterogeneous configurations. In the balanced distribution, each facility received 1,000 scans with 250 images per pathological category. To simulate real-world specialization where centers encounter uneven case distributions, Dirichlet sampling was evaluated at alpha values of 0.1, 0.5, and 1.0, representing high, moderate, and mild data skews. Simultaneously, the vertical federated configuration divided feature space along the depth of a shared EfficientNet-B4 backbone, assigning high-level semantic representations from layers one through three to the first institution, mid-level representations to the second institution, and low-level structural maps to the remaining three centers. Entity alignment across institutional feature spaces was conducted using keyed HMAC-SHA256 hashing of patient identifiers, matching overlapping records securely without exposing patient identities.

The training workflow of the HOS-VFL framework executes in two distinct communication phases designed to minimize bandwidth while enforcing rigorous cryptographic protections. In the initial one-shot round, each institution trains its local model independently using gradient clipping with a maximum norm of 1.0. Upon local convergence, institutions calculate class-wise prototype vectors by averaging the 512-dimensional activations of the penultimate layer across each tumor class. Transmitting these concise vectors—amounting to 2,048 scalars per institution—achieves a compression ratio exceeding 9,400 times compared to sending full 19.3-million-parameter models. Before transmission to the central coordinator, each institution injects calibrated Gaussian differential privacy noise at epsilon 1.0 and delta 1e-5. The central coordinator computes a global prototype through sample-weighted averaging, applying median-based anomaly detection to filter potentially poisoned submissions before initiating the second stage.

In the second communication round, participating institutions extract their assigned vertical feature subsets for aligned patient samples and encrypt them using 2,048-bit Paillier homomorphic encryption. The central coordinator receives these encrypted payloads and executes secure multiparty computation to derive attention-weighted feature representations and a vertical refinement vector without decrypting the underlying data. Query classifications are finalized by measuring the cosine distance between query feature embeddings and vertically refined global class prototypes. In classification benchmarks on BRISC2025, HOS-VFL achieved an overall accuracy of 98.73 percent, precision of 98.91 percent, recall of 98.68 percent, and an F1-score of 98.79 percent. This performance approached the centralized EfficientNet-B4 benchmark of 98.94 percent within 0.21 percentage points, while reducing communication overhead by 94.3 percent relative to multi-round federated methods that required 35 to 50 iterations.

Comparative evaluations showed that HOS-VFL substantially surpassed conventional federated baselines evaluated on identical Dirichlet partitions at alpha 0.5. The framework exceeded the 94.58 percent accuracy of Federated Averaging by 4.15 percentage points, FedProx at 95.42 percent by 3.31 points, FedBN at 95.87 percent by 2.86 points, and a specialized VGG16 federated model at 96.00 percent by 2.73 points, while surpassing centralized Swin-HAFNet at 97.83 percent. Per-class analysis demonstrated 99.6 percent accuracy on pituitary tumors, 99.2 percent on non-tumorous cases, 98.8 percent on meningiomas, and 97.6 percent on gliomas, yielding a macro-averaged area under the curve of 0.9981. Paired t-tests across five random seeds confirmed significance against Federated Averaging with a t-statistic of 27.84, while McNemar’s test yielded a chi-squared value of 41.27, and bootstrap resampling established a 95 percent confidence interval between 98.55 and 98.90 percent.

The investigation further assessed clinical reliability through confidence calibration and uncertainty quantification using test-time Monte Carlo dropout across 20 stochastic passes. The framework recorded an Expected Calibration Error of 0.0287 across 15 bins, with high-confidence predictions exceeding 0.95 probability attaining 99.1 percent accuracy. Predictive entropy averaged 0.0456 nats for correct classifications versus 0.8234 nats for incorrect assignments, allowing an uncertainty threshold at 0.5 nats to flag 11 of 13 misclassifications with 84.6 percent sensitivity and a 2.3 percent false-positive rate. Detailed failure analysis revealed that six misclassifications occurred in lesions smaller than 500 cubic millimeters where pixel information was limited, four were compromised by partial volume artifacts near the cranial boundary, and three exhibited ambiguous imaging features where inter-rater diagnostic agreement was low.

Tumor segmentation was implemented using an independent 24.1-million-parameter U-Net++ architecture with ResNet-34 encoder blocks, attention gates, and deep supervision heads, maintaining separate parameters from the classifier to eliminate multi-task gradient conflict. Across tumor-containing scans, HOS-VFL achieved an unrefined mean Dice similarity coefficient of 0.9147 and a mean Intersection over Union of 0.8423, approaching centralized U-Net at 0.9218 while outperforming federated U-Net at 0.8740. Mean Dice scores reached 0.9245 for gliomas, 0.9185 for pituitary tumors, and 0.9012 for meningiomas. By lesion volume, tumors larger than 5,000 cubic millimeters achieved a Dice score of 0.9456, whereas tumors under 1,000 cubic millimeters reached 0.8523. Applying conditional random field post-processing raised overall mean Dice to 0.9223, while morphological filtering reduced false-positive rates to 1.2 percent, producing segmentations requiring no corrections in 92 percent of cases.

External cross-dataset benchmarks and privacy stress testing verified the generalizability and security defenses of the framework across diverse settings. On the BraTS 2021 glioma segmentation benchmark, the zero-shot framework recorded a whole-tumor Dice score of 0.8842, advancing to 0.9087 after five fine-tuning epochs, while classification accuracy reached 97.84 percent on the Cheng dataset. In privacy defense assessments, differential privacy noise and gradient clipping suppressed gradient inversion reconstruction quality to 8.34 decibels, rendering reconstructed images unrecognizable. Membership inference accuracy remained at 51.2 percent, model inversion success stayed below 5 percent, and median-based robust aggregation eliminated 94.7 percent of simulated poisoning vectors. Under extreme Dirichlet non-identical class distributions at alpha 0.1, classification accuracy remained at 96.84 percent, confirming operational resilience under acute data heterogeneity.

Despite these favorable performance metrics, the authors emphasized substantial methodological constraints that must be addressed prior to prospective clinical deployment. The experimental results were derived from an artificially partitioned single dataset across five simulated environments rather than genuine hospital installations equipped with heterogeneous electronic health record platforms, varying picture archiving systems, or asynchronous transmission networks. Moreover, the threat model did not encompass adaptive adversarial strategies, property inference exploits, or direct attacks against the vertical aggregation step, and clinical acceptability was assessed through a retrospective review of only 50 cases. Future research initiatives outlined by the team must therefore focus on validating the framework within prospective multi-center clinical trials, extending vertical alignment protocols to multi-sequence imaging modalities including T2, FLAIR, and diffusion-weighted scans, and refining privacy-preserving identifier reconciliation across non-standardized hospital infrastructure.

Subject of Research: Hybrid federated learning for privacy-preserving brain tumor classification and segmentation

Article Title: Hybrid one-shot and vertical federated learning framework for privacy-preserving brain tumor classification and segmentation using the BRISC2025 dataset

Article References: M., A. S., & Chowdhary, C. L. (2026). Hybrid one-shot and vertical federated learning framework for privacy-preserving brain tumor classification and segmentation using the BRISC2025 dataset. Frontiers in Artificial Intelligence, 9, Article 1967588. https://doi.org/10.3389/frai.2026.1967588

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1967588

Keywords: brain tumor classification, medical image segmentation, federated learning, BRISC2025 dataset, one-shot federated learning, vertical federated learning, magnetic resonance imaging, homomorphic encryption, differential privacy, glioma segmentation, meningioma, pituitary tumor, EfficientNet-B4, U-Net, entity alignment, Dirichlet distribution, uncertainty quantification