BraTS Challenge

Updated 7 January 2026

BraTS Challenge is an international benchmark that standardizes mpMRI datasets and evaluation metrics for automated brain tumor segmentation.
It promotes the adoption of advanced deep learning and transformer-based methods across diverse CNS tumor types including gliomas and pediatric cases.
The challenge emphasizes rigorous annotation, stratified data splits, and lesion-wise evaluation to enhance clinical treatment planning and research reproducibility.

The Brain Tumor Segmentation (BraTS) Challenge is a series of international, multi-institutional community benchmarks focused on advancing automated brain tumor analysis and segmentation, primarily using multi-parametric magnetic resonance imaging (mpMRI). Since inception, BraTS has provided standardized datasets, annotation protocols, and rigorously defined evaluation metrics to drive method development, compare algorithms, and calibrate clinical readiness for automated segmentation across a range of central nervous system (CNS) tumor types including gliomas, meningiomas, brain metastases, and, more recently, pediatric high-grade gliomas and radiotherapy-planning cases.

1. Historical Context, Motivation, and Expansion

The initial BraTS challenge emerged in 2012 as a response to the lack of public, multi-site, expertly annotated mpMRI data for brain glioma segmentation. Over a decade, the challenge has catalyzed adoption of deep learning, standardized clinical nomenclature for regions (enhancing tumor, non-enhancing core, edema, resection cavity), and supported clinical endpoints such as radiogenomic prediction. The series has since expanded to cover adult and pediatric CNS tumors, meningiomas, brain metastases, post-treatment scenarios, image synthesis for missing modalities, healthy tissue inpainting, and longitudinal pre/post-op registration (Baid et al., 2021, Kazerooni et al., 2023, Li et al., 2023, Adewole et al., 2023, LaBella et al., 2024, Kazerooni et al., 2024, LaBella et al., 2024).

The central motivation remains establishing robust, reproducible benchmarks for tumor segmentation tasks critical to treatment planning, response evaluation, and multi-center clinical trials. The design philosophy emphasizes rigorous annotation, stratified data splits, lesion-wise evaluation, and clinically relevant metrics to ensure real-world applicability.

2. Dataset Composition, Annotation Protocols, and Preprocessing

BraTS datasets are characterized by large sample sizes, institutional diversity, harmonized imaging protocols, and multi-label expert annotation. Representative features include:

Imaging Modalities: Four co-registered 3D MRI sequences: native T1, post-contrast T1 (T1Gd/T1CE), T2, and T2-FLAIR, each resampled to 1 mm³ isotropic voxels and skull-stripped using standard pipelines (e.g., FeTS, HD-BET).
Region-of-Interest Definitions: Tumor subregions align with clinical practice: enhancing tumor (ET), non-enhancing/necrotic core (NETC/NCR), surrounding non-enhancing FLAIR hyperintensity (SNFH/edema), cystic component (in pediatrics), and resection cavity (post-treatment cases).
Annotation Workflow: Initial automated segmentation (e.g., nnU-Net, DeepScan ensembles with STAPLE), followed by iterative manual refinement and expert neuroradiologist approval, incorporating structured annotation protocols and inter-rater quality control (Kazerooni et al., 2023, LaBella et al., 2023, Kazerooni et al., 2024, LaBella et al., 2024).
Preprocessing: DICOM-to-NIfTI conversion, rigid co-registration to an atlas (SRI24/MNI), skull-stripping (with attention to preserving edge voxels and extra-axial lesions), intensity normalization, and data harmonization across centers (Kazerooni et al., 2023, LaBella et al., 2024).

Stratified Data Splits: Typically, datasets are partitioned into training (70%), validation (10%), and hidden test (20%) sets, with training labels released and validation/test labels withheld for objective ranking.

3. Segmentation Tasks, Metrics, and Evaluation Frameworks

BraTS segmentation tasks are defined per clinical question and tumor type:

Primary Segmentation Tasks: Delineation of tumor subregions—ET, TC (tumor core), WT (whole tumor)—and in select tasks, RC (resection cavity), Cystic component, or SNFH.
Lesion-wise Evaluation: To address multifocality and small lesion detection, especially in brain metastases and meningiomas, lesion-wise metrics are now standard (LaBella et al., 2024, Moawad et al., 2023).

Key Evaluation Metrics:

Metric	Mathematical Definition	Clinical Relevance
Dice Similarity Coefficient (DSC)	$\mathrm{DSC}(A,B) = \frac{2\,\|A \cap B\|}{\|A\| + \|B\|}$	Volumetric overlap
95th-percentile Hausdorff Distance (HD95)	$\mathrm{HD}_{95}(A,B) = \max\{ h_{95}(A,B), h_{95}(B,A)\}$	Boundary accuracy, outlier-tolerant
Sensitivity (Recall)	$\mathrm{Sens} = \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$	Detection rate, false negatives
Specificity	$\mathrm{Spec} = \frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$	Over-segmentation, false positives

For lesion-wise scoring, any false negative or false positive lesion is penalized by assigning DSC = 0 and HD95 = 374 mm. Lesions <50 voxels are excluded to minimize noise from uninformative small structures (LaBella et al., 2024, Moawad et al., 2023).

Ranking is determined by aggregating each team’s rank across metrics and subregions, typically by averaging ranks for DSC and HD95 in ET, TC, and WT or for all defined classes. Permutation testing is employed to assess statistical significance of team differences (LaBella et al., 2024, Kazerooni et al., 2024).

4. Representative Challenge Tracks and State-of-the-art Methods

4.1 Intracranial Meningioma (BraTS 2023)

Dataset: Largest multi-institutional, expert-annotated meningioma mpMRI dataset (1000 train, 141 validation, 283 test cases), with three labels: ET, NET, SNFH (LaBella et al., 2024).
Top Method: Auto3DSeg (NVAUTO, SegResNet backbone) with five-fold ensembling, achieved median DSCs of 0.976 (ET), 0.976 (TC), and 0.964 (WT); mean HD95 values of 23.9 mm (ET), 21.8 mm (TC), 31.4 mm (WT). Systematic failure modes included loss of tumor at brain mask boundary (90.3% of cases affected) due to skull-stripping, and poor performance on calcified NET subtypes.
Benchmark Impact: Establishes new standards: lesion-wise DSC ≈ 0.98 (ET, TC), 0.96 (WT); HD95 medians ≈ 1 mm (LaBella et al., 2024).

4.2 Pediatric High-Grade Glioma (BraTS-PEDs 2023)

Unique Considerations: Pediatric HGGs exhibit more frequent infiltrative, non-enhancing, cystic, and midline tumors, complicating subregion annotation and segmentation (Kazerooni et al., 2024, Kazerooni et al., 2023).
Segmentation Schema: Four subregions (ET, NET, CC, ED), merged into three evaluation labels (ET, NC, ED).
Best Methods: Ensembles of nnU-Net and Swin UNETR, or transformer-augmented networks. CNMCPMI2023’s ensemble reached mean Dice of 0.65 (ET), 0.81 (TC), 0.83 (WT); NVAUTO Auto3DSeg (SegResNet) achieved 0.55/0.78/0.84 (Kazerooni et al., 2024).
Performance Trends: Substantial drop in ET segmentation accuracy compared to adult BraTS (Dice ~0.55–0.65), reflecting the scarcity and small volumes of enhancing regions in DMG/DIPG. WT accuracy (Dice ~0.83–0.84) matched adult benchmarks (Kazerooni et al., 2024).

4.3 Other Notable Tracks (Selection)

BraTS-MEN-RT 2024: Focuses on expert-annotated GTV segmentation for radiotherapy planning on post-contrast T1 MRIs, using Dice and HD95 without skull-stripping to preserve extracranial margins (LaBella et al., 2024).
BraTS-METS 2023: Brain metastasis segmentation with rigorous lesion-wise metrics, emphasizing sensitivity to small lesions (≥ 40% false negatives <5 mm diameter) (Moawad et al., 2023).
Adult Glioma (BraTS 2023): Ensemble approaches (e.g., SegResNet + MedNeXt) using channel-wise postprocessing and deep supervision are among top performers; third-place achieved average Dice of 0.8313 and HD95 of 36.38 mm (Maani et al., 2024).

5. Methodological Advances and Failure Modes

The BraTS series has been instrumental in driving technical innovation in multi-modal 3D convolutional neural networks, self-configuring architectures (nnU-Net), heavy data augmentation, ensembling, transformer-based encoders, lesion-wise postprocessing, and uncertainty quantification. Recent advances include:

Self-configuring architectures (nnU-Net, Auto3DSeg): Dynamic adaptation of patch size, learning rate, data augmentation, and ensembling to maximize performance across tasks (LaBella et al., 2024, Kazerooni et al., 2024).
Transformer-based Encoders: Swin UNETR and masked autoencoder pretraining yield improvement on complex, heterogeneous data (Kazerooni et al., 2024).
Dedicated Postprocessing: Connected component filtering and false-positive suppression at lesion level are crucial for optimizing under the BraTS 2023–2024 lesion-wise metric regime (Maani et al., 2024, LaBella et al., 2024).

Common Failure Modes:

Skull-stripping artifacts: Aggressive brain masking often results in loss of extra-axial (meningeal) tumor tissue, with up to 90% of meningioma cases exhibiting abutment or cropping at the brain mask boundary (LaBella et al., 2024).
Non-enhancing and calcified lesions: Poor representation in training data leads to systematic under-segmentation of these biologically distinct subtypes (LaBella et al., 2024).
Small lesion detection: Sensitivity for sub-5 mm metastases remains a major challenge, with most false negatives in BraTS-METS and meningioma tracks arising from this cohort (Moawad et al., 2023).

6. Clinical and Research Implications, Limitations, and Recommendations

Clinical Impact: Automated, reproducible 3D volumetry of glioma and meningioma subregions enables more objective treatment planning in radiotherapy (GTV/CTV definition), neurosurgical guidance (volumetry and spatial mapping), and longitudinal assessment (progression versus treatment effect). In pediatric settings, such tools support standardized response criteria (RAPNO), multi-center trials, and centralized reading (LaBella et al., 2024, Kazerooni et al., 2024, Kazerooni et al., 2024).

Limitations:

Unbalanced representation of rare histologies (e.g., non-enhancing, calcified, and cystic variants) hampers generalizability.
Skull-stripping and anonymization strategies are a major source of error for extra-axial tumors and need refinement (e.g., mri_reface or edge-preserving defacing) (LaBella et al., 2024).
Segmentation of subtle subregions (enhancement in pediatric tumors; small metastases) is limited by class imbalance and annotation ambiguity.
Transferability across global centers and scanner types (e.g., low- and middle-income settings) remains problematic without robust domain adaptation (Adewole et al., 2023).

Recommendations:

Augment training datasets to include non-enhancing, calcified, and small lesions for all tumor types.
Adopt lesion-wise metrics and robust postprocessing for multifocal or heterogeneously labeled cases.
Redesign skull-stripping/anonymization workflows to preserve extra-axial space while securing privacy.
Continuous benchmarking leveraging federated learning or cross-center cohort expansion will improve model generalizability and clinical translation (LaBella et al., 2024, Kazerooni et al., 2024, Adewole et al., 2023).

7. Future Directions

Expansion to multi-modal (e.g., CT, PET), multi-phase (longitudinal, post-operative), and multi-task (segmentation + molecular prediction) settings (Baid et al., 2021).
Direct benchmarking of radiotherapy planning targets and integration with commercial auto-contouring solutions (LaBella et al., 2024).
Extension of BraTS evaluation to uncertainty quantification and confidence-driven workflows for clinical review (Mehta et al., 2021).
Enhancement of pediatric and global coverage with increased international, domain-heterogeneous datasets to foster equitable AI tool deployment (Kazerooni et al., 2023, Kazerooni et al., 2024, Kazerooni et al., 2024, Adewole et al., 2023).

By maintaining high standards of annotation, robust metric design, and a transparent, reproducible evaluation ecosystem, the BraTS Challenge remains the definitive global benchmark for brain tumor segmentation, continually setting state-of-the-art standards and guiding research toward clinical robustness and reproducibility (LaBella et al., 2024, Kazerooni et al., 2023, Kazerooni et al., 2024, LaBella et al., 2024).