RSNA-ASNR-MICCAI BraTS 2021 Dataset

• RSNA-ASNR-MICCAI BraTS 2021 dataset is a comprehensive, publicly available MRI resource featuring expertly curated glioma cases with multi-parametric imaging modalities.
• It provides voxelwise annotations for distinct tumor sub-regions and binary MGMT promoter methylation status for robust segmentation and radiogenomic challenges.
• The dataset’s standardized preprocessing and strict validation protocols promote reproducibility and fair benchmarking across diverse automated analysis workflows.

The RSNA-ASNR-MICCAI BraTS 2021 Challenge Dataset is the largest, publicly released, multi-institutional collection of expertly curated pre-operative, multi-parametric brain MRI studies designed for benchmarking automated glioma segmentation and radiogenomic prediction algorithms. Developed under the auspices of the Radiological Society of North America (RSNA), American Society of Neuroradiology (ASNR), and Medical Image Computing and Computer Assisted Interventions Society (MICCAI), it provides a robust experimental framework for two principal tasks: (1) the delineation of histologically distinct glioma sub-regions, and (2) inference of O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status from imaging. This resource aggregates MRI acquisitions and manual expert annotations from over eighty centers worldwide, encompassing real-world variations in imaging protocols, scanner manufacturers, and clinical workflows (Baid et al., 2021).

1. Dataset Composition and Structure

The BraTS 2021 dataset comprises 2,040 glioma patients with histological grades II–IV, stratified as 1,251 cases in the segmentation and radiogenomic training cohort, 219 in the validation cohort, and 570 held out for testing. For segmentation, each case includes voxelwise, four-way annotations—background, necrotic/cystic core (NCR), peritumoral edema/invasion (ED), and gadolinium-enhancing tumor (ET). The radiogenomic component provides binary MGMT methylation status for all training cases, determined by site-specific pyrosequencing (>10% of CpG 74–81) or bisulfite sequencing (≥2% of 17 CpG sites) (Baid et al., 2021, Pálsson et al., 2021).

Data splits are strictly enforced for benchmarking: training sets expose labels, validation and testing sets release only images, and any augmentation with external data is disallowed for leaderboard or final ranking (Baid et al., 2021).

2. Imaging Modalities, Acquisition, and Preprocessing

Each subject is represented with four co-registered, skull-stripped 3D MRI volumes:

• T1-weighted (pre-contrast)
• T1-Gadolinium (post-contrast)
• T2-weighted
• T2-FLAIR

Raw data originated from diverse clinical 1.5T and 3T scanners—including both public (e.g., TCGA-GBM, TCGA-LGG, IvyGAP) and private institutional repositories—yielding wide-ranging in-plane resolutions (0.4–1.0 mm), slice thicknesses (1–7 mm), and spatial orientations. Standardized preprocessing involved DICOM-to-NIfTI conversion, rigid co-registration to SRI24 template space, isotropic resampling to 1×1×1 mm³, deep learning–driven skull-stripping, N4 bias-field correction, and per-volume z-score intensity normalization (Baid et al., 2021).

For the radiogenomic classification challenge, the curated NIfTI volumes were remapped back into their native DICOM space and anonymized to facilitate standardized algorithm deployment (Baid et al., 2021).

3. Annotation Protocol and Ground Truth

Tumor region annotations were generated via STAPLE fusion of top-performing BraTS 2020 algorithms (nnU-Net, DeepScan, DeepMedic), then subjected to iterative editing by neuroradiology experts and final approval by board-certified specialists (≥15 years experience). Annotation definitions:

• ET: all T1-Gd hyperintense regions outside normal white matter
• NCR: T1-Gd hypointense necrotic/cystic regions
• ED: T2-FLAIR hyperintense peritumoral infiltrative and vasogenic edema

Compound regions are derived as Whole Tumor (WT = NCR ∪ ED ∪ ET) and Tumor Core (TC = NCR ∪ ET) (Baid et al., 2021). Only the MGMT status (binary) is available as subject-level clinical metadata for radiogenomic predictions (Baid et al., 2021, Pálsson et al., 2021, Kollias et al., 2023).

4. Algorithmic Workflows and Methodological Benchmarks

Segmentation

Segmentation methodologies employ architectures common to top BraTS entrants, including ensembles of 3D U-Net–style CNNs and variants such as nnU-Net (Pálsson et al., 2021, Peiris et al., 2022). For instance, workflows include:

• Four-channel 240×240×155 input: modalities as channels, softmax output for background and three subregions (Pálsson et al., 2021).
• Network losses: combined voxelwise cross-entropy and soft Dice loss; augmentations: on-the-fly rotations, scaling, elastic deformations, intensity perturbations (Pálsson et al., 2021).
• Reciprocal adversarial training pairs a 3D U-Net segmenter with a 3D PatchGAN critic, optimizing a composite loss combining Dice, virtual adversarial training (VAT), and adversarial consistency terms (Peiris et al., 2022).
• VAT augments robustness by enforcing local distributional smoothness under input perturbations via an additional KL divergence–based loss (Peiris et al., 2022).

Performance is assessed quantitatively via mean Dice (ET: 84.6%, TC: 85.3%, WT: 90.5% on test data), and 95th percentile Hausdorff Distance (ET: 13.5 mm, TC: 17.0 mm, WT: 6.3 mm) (Peiris et al., 2022). These results are on par with top BraTS 2020/2021 methods.

Radiogenomic Classification

Radiogenomic pipelines operate in both feature-engineering and end-to-end deep learning regimes:

• Hand-crafted radiomic features (PyRadiomics; first-order, shape, texture GLCM, GLRLM, GLSZM, GLDM, NGTDM) and learned latent shape descriptors (3D VAE μ-vectors) are extracted from network-derived segmentations (Pálsson et al., 2021).
• Random forest classifiers with Fisher’s exact test–screened features outperform latent-only baselines; radiomics alone achieved validation AUC 0.632, latent-only 0.488, and the combination 0.598. Improved feature selection and larger hold-out sets are recommended (Pálsson et al., 2021).
• Deep learning approaches (e.g., BTDNet (Kollias et al., 2023)) directly process all slices per modality with a CNN-LSTM backbone, a routing mask for variable slice counts, and cross-modality fusion. Geometric and mixup-style augmentations, focal loss (with α, γ parameters), and test-time ensembling enhance robustness and generalizability. Macro F₁ scores exceeded prior challenge winners, achieving 0.662±0.031 (five-fold cross-validation) (Kollias et al., 2023).

Radiogenomic evaluation employed only volume-level MGMT labels; no voxel- or lesion-level localization is required for classification entries.

5. Evaluation Metrics and Ranking

Segmentation is scored per-case and per-region using Dice Similarity Coefficient

$$\mathrm{Dice}(X,Y) = \frac{2\,|X \cap Y|}{|X| + |Y|}$$

and 95th percentile Hausdorff Distance, together with sensitivity and specificity. Teams are ranked per-metric and per-region on each of the 570 test sets, and Final Ranking Score (FRS) is computed by averaging these ranks. Permutation testing determines statistical significance (Baid et al., 2021).

Radiogenomic MGMT classification is assessed via Area Under the ROC Curve (AUC; primary), accuracy, macro F₁-score, and Matthew’s correlation coefficient. The leaderboard prioritizes AUC (Baid et al., 2021, Kollias et al., 2023).

6. Data Access and Usage Policy

BraTS 2021 data is distributed subject to strict data use agreements via Sage Bionetworks Synapse and Kaggle. All data released are de-identified under HIPAA/GDPR. Participants may not supplement with external data for official submissions. Preprocessing code and annotation tools are open-source (e.g., CaPTk, FeTS toolbox), supporting reproducibility and standardization. The radiogenomic classification dataset remains accessible at the RSNA-MICCAI Kaggle portal (Kollias et al., 2023).

7. Limitations and Prospects

The main limitation of the BraTS 2021 dataset is its restriction to MGMT methylation as the sole released clinical label—patient age, performance status, and additional molecular data are absent, constraining multi-modal prediction. The challenge structure—hidden test sets and strict data partitioning—ensures fair comparison but precludes external data augmentation.

Emerging directions include:

• Multi-task learning uniting segmentation and molecular characterization;
• Advanced feature selection (e.g., LASSO, mutual information maximization);
• Integration of clinical and genomic covariates if/when available;
• National image harmonization efforts to further mitigate cross-site effects.

Continued development of robust segmentation architectures (such as reciprocal adversarial training with VAT), feature-engineered and deep radiogenomic models, and consensus metrics for evaluation defines the current research paradigm for automated neuro-oncological image analysis utilizing the RSNA-ASNR-MICCAI BraTS 2021 resource (Baid et al., 2021, Pálsson et al., 2021, Peiris et al., 2022, Kollias et al., 2023).