BraTS Brain Tumor Segmentation

• Brain Tumor Segmentation (BraTS) is a standardized framework that provides curated multi-parametric MRI datasets with expert annotations for benchmarking segmentation algorithms.
• It employs rigorous annotation protocols combining automated pre-segmentation and iterative expert reviews to delineate subregions like enhancing tumor, tumor core, cystic components, and edema.
• State-of-the-art techniques including 3D U-Nets, transformer models, and graph-based methods achieve high Dice scores, advancing clinical applications in therapy planning and survival prediction.

Brain Tumor Segmentation (BraTS) refers to a series of standardized challenges, curated datasets, and community benchmarks focused on the automatic, quantitative analysis of brain tumor subregions using multi-parametric magnetic resonance imaging (mpMRI). The BraTS framework provides rigorously annotated MRI data for a range of cohorts—including adult glioma, pediatric glioma, meningioma, and brain metastases—enabling comparative assessment of state-of-the-art deep learning segmentation algorithms. Targets include not only the segmentation of biologically and clinically relevant subregions (such as enhancing tumor, non-enhancing core, cystic components, resection cavities, and peritumoral edema) but also derived quantitative endpoints critical for therapy response, radiotherapy planning, and survival prediction (Verdier et al., 2024, Kazerooni et al., 2024, Capellán-Martín et al., 16 Dec 2025).

1. Evolution and Rationale of BraTS Datasets and Cohorts

BraTS originated from the need for high-quality, multi-institutional datasets with voxelwise ground truth, catalyzing benchmark-driven segmentation research in neuro-oncology. Recent editions have expanded the cohort range and refined annotation protocols:

• Adult Diffuse Glioma: Historically, BraTS focused on pre-operative high- and low-grade gliomas, providing four mpMRI sequences per case (T1, T1-Gd, T2, FLAIR). Core subregions defined included enhancing tumor (ET), non-enhancing core (NET/necrosis), and peritumoral edema (ED/SNFH) (Zeineldin et al., 2022).
• Post-treatment Glioma (BraTS 2024): The 2024 challenge established a new standard with the largest post-treatment glioma dataset to date (2,200 expertly labeled cases), incorporating ET, NETC, SNFH, and resection cavity (RC) (Verdier et al., 2024). This design addresses the complexity of therapy-induced parenchymal changes.
• Pediatric Brain Tumor (BraTS-PEDs): Datasets from multiple international consortia (e.g., CBTN, DMG/DIPGR, BCH, Yale) aggregate over 450 pediatric HGGs, annotated for ET, NET, cystic component (CC), and edema, per RAPNO recommendations (Kazerooni et al., 2024, Kazerooni et al., 2024).
• Meningioma Radiotherapy Planning (MEN-RT): MEN-RT isolates the radiotherapy target volume (GTV/CTV) in post-contrast T1w images, adhering to consensus clinical protocols for both pre- and postoperative cases, and is further stratified by tumor status (LaBella et al., 2024).
• Sub-Saharan Africa/Africa Dataset: Recognizing domain shift and health inequity, BraTS-Africa includes data reflecting low-field MRI, variable quality, and advanced-stage presentations unique to LMIC environments (Adewole et al., 2023).

Recent BraTS editions use stratified cross-validation and radiomic-guided clustering to maximize representativeness across the biological and image-acquisition heterogeneity (Capellán-Martín et al., 16 Dec 2025, Jiang et al., 2024).

2. Annotation Protocols and Tumor Subregion Definitions

Expert-driven annotation standards are central to BraTS. All segmentation protocols are informed by clinical practice and reviewed by board-certified neuroradiologists and, in radiotherapy subsets, by radiation oncologists:

• Enhancing Tissue (ET): Defined as nodular/thick non-vascular enhancement on T1Gd (excluding vessels, linear dural enhancement), further clarified by T1Gd – T1 subtraction maps (Verdier et al., 2024).
• Non-Enhancing Tumor Core (NET/NETC/NC): Includes necrosis, cystic change, and intrinsic T1 hyperintensity within the core (delineated from RC and non-enhancing FLAIR hyperintensity) (Verdier et al., 2024, Kazerooni et al., 2024).
• Cystic Component (CC) [PED]: Identified as intra-tumoral cysts, T2-bright and T1CE-dark, preserving CSF-like signal (Kazerooni et al., 2024, Kazerooni et al., 2024).
• Surrounding Non-Enhancing FLAIR Hyperintensity (SNFH/ED): Captures FLAIR signal abnormalities due to edema, infiltrative tumor, or post-therapy gliosis, excluding chronic small-vessel disease (Verdier et al., 2024, Kazerooni et al., 2024).
• Resection Cavity (RC) [Post-treatment]: Defined as the fluid-filled, air, blood, or proteinaceous post-surgical bed (Verdier et al., 2024).
• Gross Tumor Volume (GTV/CTV) [MEN-RT]: For preoperative cases, includes all enhancing tumor plus nodular dural tails; for postoperative, merges resection cavity with institutional CTV margins (LaBella et al., 2024).

Annotation typically combines automated pre-segmentation fusion (e.g., STAPLE across nnU-Net, MONAI SegResNet, proprietary CNNs) with meticulous manual refinement and iterative expert review, using standardized, tool-assisted platforms (ITK-SNAP, CaPTk/FeTS pipelines).

3. Core Segmentation Algorithms and Network Architectures

Segmentation methods assessed in BraTS are predominantly based on three-dimensional convolutional networks, with increasing adoption of transformers and graph-based methods:

• 3D U-Net Family: The canonical encoder–decoder architecture with skip connections, often enhanced with self-configuring parameters (nnU-Net), deep supervision, and combinations such as MedNeXt (CNN+attention hybrids) (Zeineldin et al., 2022, Verdier et al., 2024, Jiang et al., 2024).
• Transformer-based Models: Swin UNETR and similar architectures inject hierarchical local self-attention and long-range context modeling, frequently ensemble-fused with CNNs (Kazerooni et al., 2024, Jiang et al., 2024, Capellán-Martín et al., 2024).
• Multi-branch and Fusion Models: Architectures such as multiPI-TransBTS employ modality-specific encoders with adaptive feature fusion (AFF), separating modality signals in early layers to enhance subregion learning, and task-specific decoders (TSFI) (Zhu et al., 2024).
• Graphical and Hybrid Approaches: Joint GNN–CNN schemes encode supervoxel-level global structure via GraphSAGE-pooling, followed by local CNN refinement of tumor boundaries (Saueressig et al., 2021).
• GAN-augmented Pipelines: Synthetic high-contrast MR channels generated by conditional GANs can supplement or replace real modalities, enhancing contrast for segmentation (Hamghalam et al., 2019).
• Patchwise Multiscale and Attentional Networks: Patch-based CNNs with multi-scale input and deep supervision, as well as scale-attention modules (SA-Net, GLIMS), support efficient context aggregation and robust feature extraction (Stawiaski, 2017, Yuan, 2020, Yazıcı et al., 2024).

Ensembling remains a central paradigm, often with region-wise or lesion-wise fusion strategies that leverage complementary strengths of different network backbones and fusion of probabilistic softmax outputs (Jiang et al., 2024, Capellán-Martín et al., 16 Dec 2025).

4. Loss Functions, Training, and Evaluation Protocols

Losses and learning strategies are tailored for severe class imbalance and intricate tumor morphology:

• Dice Loss and Generalizations: Used ubiquitously, either on its own or combined with cross-entropy, for direct overlap optimization. Dice is often employed per subregion, sometimes with class weights inverse to region prevalence.
• Auxiliary and Compound Losses: Additions include focal loss (emphasizing hard voxels), active contour terms (volumetric and length penalties), edge losses (matching gradient magnitude), and task-specific weighting (for small ET or NET) (Myronenko et al., 2020, Ren et al., 2024).
• Data Preprocessing: Standardized steps are conversion to NIfTI, skull-stripping, rigid or affine registration to SRI24 or MNI templates, resampling to 1 mm³ voxels, intensity normalization (z-score), and (optionally) N4 bias field correction (Capellán-Martín et al., 16 Dec 2025, Kazerooni et al., 2024).
• Data Augmentation: Spatial (random flipping, rotation, scaling, elastic deformation) and intensity (shift, scale, gamma, Gaussian noise) augmentations are used extensively for robustness.
• Cross-validation and Stratification: Five-fold or stratified cross-validation—by key radiomic subtypes or tumor clusters—ensures generalization and balanced sampling (Capellán-Martín et al., 16 Dec 2025).
• Adaptive Processing: Radiomic subtype clustering (PCA + k-means on shape and intensity features) controls pre- and post-processing thresholds, improving segmentation across tumor phenotypes (Jiang et al., 2024, Capellán-Martín et al., 16 Dec 2025).
• Ranking and Metrics: Lesion-wise Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD₉₅) are standard. Adjustments such as lesion-wise aggregation penalize failures on small or rare lesions, and specificity/sensitivity are also monitored (Verdier et al., 2024, Kazerooni et al., 2024).

Metric	Formula	Role
Dice (DSC)	$\frac{2|X \cap Y|}{|X| + |Y|}$	Overlap between prediction and ground truth
Hausdorff	$$\max\{\sup_{a\in A}\inf_{b\in B}d(a,b), \sup_{b\in B}\inf_{a\in A}d(a,b)\}_{95\%}$$	Boundary error; 95th percentile
Sensitivity	$\frac{\rm TP}{\rm TP + FN}$	True positive rate (recall)
Specificity	$\frac{\rm TN}{\rm TN + FP}$	True negative rate

5. Quantitative Results and Comparative Performance

Recent BraTS editions have codified multi-center performance improvement:

• Top Ensemble Results (Adult BraTS): On multi-institutional validation/testing, nnU-Net/Swin UNETR/MedNeXt ensembles, with adaptive post-processing, yield lesion-wise Dice scores of WT 0.926, 0.918 for TC, and 0.692 for ET in pediatric (BraTS 2024), MEN-RT GTV 0.801, and MET WT 0.688 (Jiang et al., 2024). For SSA cohorts, the best approaches maintain WT Dice ≈ 0.87–0.97 and ET ≈ 0.82–0.90, even under domain shift (Zeineldin et al., 2022, Adewole et al., 2023).
• Pediatric BraTS Performance: On the BraTS-PEDs 2023 test set, ensembles reach WT Dice ≈ 0.84, TC ≈ 0.81, and ET ≈ 0.65 across diffuse midline glioma. Ensembles consistently outperform single models, especially in detecting small enhancing foci. Lesion-wise ranking penalizes missing or false positives in subregions, a critical adjustment for small-volume ET/CC in DIPG (Kazerooni et al., 2024).
• Post-treatment Glioma: Performance in post-treatment ET and SNFH is generally lower than in pre-operative cohorts, highlighting the increased segmentation complexity of therapy-induced changes, as anticipated in BraTS 2024 (Verdier et al., 2024).
• MEN-RT and Metastases: MEN-RT top entries reach GTV Dice ≈ 0.80–0.87; for brain metastases, WT Dice ≈ 0.57–0.69 reflecting smaller lesion size and multiplicity (Jiang et al., 2024, Capellán-Martín et al., 16 Dec 2025).
• Ablation and Robustness: Adaptive ensemble weighting by internal cross-validated ranking and radiomic subtyping yields ~1–2% Dice gain over any backbone, and subtype-aware post-processing is essential to reduce over/under-segmentation (Capellán-Martín et al., 16 Dec 2025).

6. Clinical and Research Implications

BraTS segmentation frameworks underpin a broad translational pipeline in neuro-oncology:

• Response Assessment: Automated volumetric measurements, standardized via community benchmarks, advance objective assessment in clinical trials and therapy monitoring, reducing inter-reader variability and supporting RAPNO (Response Assessment in Pediatric Neuro-Oncology) recommendations (Kazerooni et al., 2024, Kazerooni et al., 2024).
• Radiotherapy Planning: MEN-RT and post-treatment glioma datasets foster integration of automated GTV/CTV contouring, accelerating planning and enabling individualized dose optimization (LaBella et al., 2024).
• Radiomics and Survival Prediction: Accurate, robust subregion delineation enables extraction of shape, texture, and volumetric biomarkers that feed prognostic models and drive precision medicine (Isensee et al., 2018, Hamghalam et al., 2019).
• Deployability and Equity: SSA- and LMIC-focused branches address the need for robust, lightweight, containerized AI tailored to variable scan quality and late-stage disease, with federated learning pathways preserving data sovereignty (Adewole et al., 2023).
• Methodological Innovation: Radiomic subtyping, multi-path encoding, adaptive feature fusion, hybrid CNN-transformer architectures, and lesion-wise model selection represent recent methodological advances confirmed to generalize across tumor histologies, age groups, and imaging protocols (Capellán-Martín et al., 16 Dec 2025, Zhu et al., 2024).

7. Ongoing Challenges and Future Directions

Despite advances, critical issues persist:

• Domain and Cohort Shift: Heterogeneous acquisition protocols, rare subtypes, and small pediatric enhancing regions pose persistent generalization challenges. Stratified CV and radiomics-driven subtyping mitigate, but do not fully resolve, these issues (Capellán-Martín et al., 16 Dec 2025, Kazerooni et al., 2024, Kazerooni et al., 2024).
• Ground Truth and Labeling: Manual annotation remains labor-intensive and depends on expert consensus, especially for ill-defined or therapy-altered subregions. There is a push for deep-learning–assisted active labeling and uncertainty quantification (Kazerooni et al., 2024).
• Longitudinal Analysis and Outcome Prediction: Future benchmarks will incorporate serial scans, multimodal (diffusion, perfusion, PET) input, and histologically confirmed ground truth to bridge segmentation with real-world clinical trial endpoints (Verdier et al., 2024, Capellán-Martín et al., 16 Dec 2025).
• Federated Learning and Data Harmonization: Cross-site intensity normalization, harmonization protocols, and privacy-preserving collaborative training are under evaluation to further democratize the technology (Adewole et al., 2023).
• Scalability: Efficient inference and knowledge distillation are priorities for clinical deployment, as ensembles of large models increase computational cost (Jiang et al., 2024).

A plausible implication is that as BraTS datasets and protocols expand in scope and rigor, coupled with methodological advances in radiomic-guided adaptive segmentation and transformer-based architectures, the field is positioned to transition from proof-of-concept to real-world implementation in both high-resource and resource-limited healthcare environments.