BraTS 2021 Challenge Overview

Updated 3 February 2026

BraTS 2021 Challenge is a benchmark for advanced brain tumor segmentation and MGMT status prediction, using a large, multi-parametric MRI dataset.
The challenge emphasizes standardized preprocessing, precise segmentation metrics (DSC, HD95), and containerized evaluation to ensure reproducibility.
Innovative approaches, including U-Net variants, ensembling, and custom loss functions, achieved performance at or above clinical standards.

The Brain Tumor Segmentation (BraTS) 2021 Challenge is a large-scale, community-driven benchmark for the development and evaluation of computational algorithms in brain tumor segmentation and radiogenomic classification using multi-parametric MRI. Organized in conjunction with RSNA, ASNR, and MICCAI, BraTS 2021 marks the tenth anniversary of the initiative, furthering its impact by providing the largest multi-institutional dataset to date and introducing dual tasks: the segmentation of glioma sub-regions and the prediction of MGMT promoter methylation status from pre-operative mpMRI. Submissions are evaluated on rigorously curated reference standards by containerized testing on held-out data, ensuring reproducibility and clinical relevance (Baid et al., 2021).

1. Dataset Composition and Preprocessing

BraTS 2021 comprises 2,040 pre-operative glioma cases drawn from a federated consortium of public and >70 private sources, systematically partitioned into 1,251 training, 219 validation, and 570 (sometimes reported as 530) test cases. Each case includes four rigidly co-registered, skull-stripped, and resampled 3D MRI sequences (T1, T1Gd, T2, T2-FLAIR) at 1×1×1 mm³, with volumes sized approximately 240×240×155 voxels. Preprocessing is standardized: DICOM to NIfTI conversion, rigid registration to the SRI24 atlas, brain extraction, and per-volume z-score intensity normalization.

Segmentation ground truth for Task 1 consists of expert-refined labels for necrotic/non-enhancing core (NCR/NET, label 1), peritumoral edema (ED, label 2), and enhancing tumor (ET, label 4), with derived evaluation targets: whole tumor (WT = {1,2,4}), tumor core (TC = {1,4}), and ET = {4} (Baid et al., 2021, Zeineldin et al., 2021). For Task 2, a subset of 585 subjects is labeled with binary MGMT methylation status, curated from clinical assays.

2. Segmentation and Classification Task Definitions

BraTS 2021 comprises two primary tasks:

Task 1: Sub-region Segmentation — Delineation of histologically distinct glioma compartments (ET, TC, WT) on mpMRI, evaluated using the Dice Similarity Coefficient (DSC) and 95th-percentile Hausdorff distance (HD95) for each region. Official rankings are derived from averaged subject-level ranks across these metrics and subregions (Baid et al., 2021, Siddiquee et al., 2021).
Task 2: Radiogenomic Classification — Prediction of MGMT promoter methylation status from mpMRI, scored by ROC AUC, accuracy, precision, recall, and F₀.₅. The challenge provides only preprocessed images and labels at the patient level; segmentation-derived or deep-learned features may be used (Pálsson et al., 2021, Kollias et al., 2023).

3. Algorithmic Approaches and Model Architectures

BraTS 2021 catalyzed advances across U-Net variants, deep ensembling, augmentation, and learning paradigms:

Ensemble and Self-configuring Frameworks: nnU-Net, a self-adapting 3D U-Net architecture with tailored normalization, aggressive data augmentation, and hybrid region-wise loss (BCE + batch Dice), formed a core of multiple top entries (Zeineldin et al., 2021). Custom U-Net derivatives such as DeepSeg and E1D3 U-Net (a one-encoder, three-decoder variant) provide architectural diversity and efficiency, enabling robust segmentation without reliance on multi-model ensembles (Bukhari et al., 2021, Zeineldin et al., 2021).
Loss Innovations: Several teams incorporated custom losses to address specific challenges. Barlow-Twins–style redundancy reduction was used to decorrelate feature embeddings under perturbations, improving ET and TC Dice (Siddiquee et al., 2021). Generalized Wasserstein Dice loss (GWDL) with anatomically-informed class distance matrices penalized biologically implausible misclassifications less harshly, yielding superior boundary precision and outlier suppression (Fidon et al., 2021).
Test-Time Augmentation and Confidence Ensembling: Extensive inference-time augmentation (flips, scaling, Gaussian blending) proved routine, with leading submissions averaging predictions over spatial transforms and folds. Confidence-based ensemble selection, quantified by region-wise average probability, supplanted naïve averaging, favoring segmentations with internally consistent probability maps (Siddiquee et al., 2021, Fidon et al., 2021).
Adversarial and Robustness-Enhancing Strategies: Reciprocal adversarial learning combined a U-Net generator, virtual adversarial perturbations (encouraging local smoothness), and a voxel-wise PatchGAN critic for high-order anatomical consistency, yielding improved segmentation, especially in low-SNR regimes (Peiris et al., 2022).
Radiogenomic Pipelines: Approaches for MGMT classification included multi-stream CNN-RNNs with dynamic routing/length masking (BTDNet), radiomic/shape feature ensembles using variational autoencoders, and late multimodal fusion. Effective augmentation (MixAugment in 3D, TTA) and advanced loss formulations (focal loss, SAM optimization) were instrumental in boosting generalization (Kollias et al., 2023, Pálsson et al., 2021).

4. Quantitative Results and Technical Benchmarks

Performance converged at or above human inter-rater thresholds:

Method/Paper	Dice ET	Dice TC	Dice WT	HD95 ET (mm)	HD95 TC	HD95 WT
Ensemble CNN (Zeineldin et al.) (Zeineldin et al., 2021)	87.6	87.5	91.9	12.1	6.3	14.9
Redundancy Reduction (NVAUTO) (Siddiquee et al., 2021)	86.0	88.7	92.7	9.1	5.8	3.6
Wasserstein Ensemble (Fidon et al., 2021)	87.4	87.8	92.9	10.1	15.8	4.1
Reciprocal Adversarial (Peiris et al., 2022)	84.6	85.3	90.5	13.5	17.0	6.3
E1D3 U-Net (Bukhari et al., 2021)	86.5	86.7	91.8	9.5	17.4	5.7

The BTDNet classifier delivered state-of-the-art MGMT F1, achieving 66.2 ± 3.1 across validation folds—exceeding prior published CNN-RNN runners-up by over 18 percentage points in F1 (Kollias et al., 2023).

5. Methodological Themes and Insights

Prevailing strategies in BraTS 2021 included:

Heavy, Diverse Augmentation: Standardized spatial and intensity perturbations at both training and test-time were critical for inter-site generalization (Zeineldin et al., 2021, Fidon et al., 2021).
Ensembling as Standard Practice: All leading segmentation solutions employed model and data-split ensemble procedures, often combining distinct architectures or loss configurations (e.g., U-Net plus DeepSeg, or confidence-weighted ensembles) (Zeineldin et al., 2021, Siddiquee et al., 2021).
Loss and Consistency Optimization: Superior performance on challenging regions (notably ET) resulted from specialized objectives that enforced invariance (Barlow-Twins loss), anatomical class structure (GWDL), or statistical robustness (virtual adversarial loss) (Siddiquee et al., 2021, Fidon et al., 2021, Peiris et al., 2022).
Dimensionality and Mask Management in MGMT: Handling non-uniform volume lengths and integrating multi-modal temporal features via LSTM/Routing blocks (as in BTDNet) significantly improved radiogenomic prediction (Kollias et al., 2023).

6. Clinical and Benchmarking Implications

High-fidelity automated segmentation supports neurosurgical planning, therapy monitoring, and individualized prognosis by reducing manual annotation burdens and inter-observer variability. The challenge’s infrastructure—containerized evaluation, centralized curation, and standardized metrics—ensures reproducibility and fair comparison across methods (Zeineldin et al., 2021, Baid et al., 2021).

Radiogenomic prediction establishes benchmarks for noninvasive molecular stratification, setting the stage for future large-scale, cross-institutional studies (Baid et al., 2021, Pálsson et al., 2021, Kollias et al., 2023).

7. Open Problems and Future Directions

Challenges persist regarding:

Annotation Variability: Single-rater annotations with expert approval preclude direct quantification of inter-rater agreement; future challenges may include multi-expert annotation subsets (Baid et al., 2021).
External Validation and Generalization: Expansion to federated learning, continuous molecular targets, and domain adaptation is anticipated (Kollias et al., 2023).
Advances in Architectural Diversity: Transformer integration and further meta-learned ensembling are emerging research directions, though overparameterization and limited training data currently constrain their utility (Fidon et al., 2021).
Radiomic–Deep Feature Synthesis: Hybrid pipelines integrating shape, radiomic, and latent deep features are under examination for further gains in radiogenomic classification (Pálsson et al., 2021, Kollias et al., 2023).

The BraTS 2021 Challenge thus serves as a milestone for both methodological innovation and comparative benchmarking within neuroimaging and computational oncology research (Baid et al., 2021).