BraTS 2025 Segmentation Challenge

• The BraTS 2025 Segmentation Challenge is a global initiative that benchmarks automated brain tumor segmentation on curated, multi-institutional MRI datasets.
• The challenge employs a robust, multi-phase annotation pipeline combining AI pre-segmentation and expert reviews to generate high-fidelity volumetric data.
• Evaluation leverages metrics like DSC and Hausdorff Distance, promoting clinically meaningful outcomes and streamlining radiological workflows.

The BraTS 2025 Segmentation Challenge refers to the comprehensive suite of international benchmarking initiatives under the Brain Tumor Segmentation (BraTS) umbrella for the year 2025, encompassing both the BraTS-METS 2025 Lighthouse Challenge (targeting pre- and post-treatment brain metastases on MRI) and related pediatric and adult tumor segmentation tracks. Its primary objective is to catalyze clinically meaningful, automated, and volumetrically precise brain tumor segmentation methods, evaluated on rigorously curated, multi-institutional, annotated datasets spanning a broad spectrum of disease types and patient demographics.

1. Clinical Motivation and Challenge Objectives

Brain metastases (BM) comprise a major complication across cancer subtypes, affecting 20–40% of adult cancer patients and conferring unfavorable prognosis (Maleki et al., 16 Apr 2025). Traditional clinical response assessment (e.g., via RANO-BM criteria) relies on unidimensional diameter measurements, largely due to the impracticality of routine manual 3D segmentation in clinical workflows. However, volumetric criteria—such as ≥30% volume increase denoting progression and ≥20% decrease indicating partial response—yield superior sensitivity and outcome correlation, yet lack widespread adoption due to the absence of validated, automated tools.

The BraTS-METS 2025 Lighthouse Challenge was established to:

• Provide a multi-institutional brain MRI dataset with high-quality, multi-rater ground truth for both pre- and post-treatment BM cases.
• Quantify human annotation variability at scale, including both "from scratch" and AI-seeded workflows.
• Benchmark and rank automated segmentation algorithms using harmonized, clinically relevant metrics.
• Release all annotated data (2023–2025) to the public domain, promoting open-science standards and supporting clinical translation (Maleki et al., 16 Apr 2025).

2. Dataset Composition and Annotation Protocol

MRI Modalities and Cohort Characteristics:

Mandatory input sequences include pre-contrast T1-weighted (T1), post-contrast T1-weighted (T1-CE), and T2-FLAIR. In 2025, T2 (non-FLAIR) is optional. The 2025 release comprises 1,778 cases (1,046 pre-treatment, 732 post-treatment) sourced from eight international institutions. Standardized spatial preprocessing (1 mm³ isotropic resampling, SRI-24 atlas registration) is applied, except for UCSF/UCSD data retained in native space (Maleki et al., 16 Apr 2025).

Annotation Pipeline:

A five-phase, hybrid annotation workflow ensures both scalability and reproducibility:

Phase	Description	Actors
Phase I – AI Pre-segmentation	Three nnU-Net models trained on heterogenous BM and glioma datasets; outputs fused via STAPLE voting	Automated (nnU-Net ensemble)
Phase II – Student/Neurorad Review	Medical student edits, followed by supervised review by 52 neuroradiologists	Students, Neuroradiologists
Phase III – Automated QC	Voxels outside brain, orientation, and mask coverage checks	Automated scripts
Phase IV – Secondary Review	Second neuroradiologist approves segmentations	Neuroradiologist
Phase V – Final Approval	Senior neuroradiologist finalizes labels	Senior neuroradiologist

For a 75-case "Lighthouse set," each underwent 4×2 expert annotations (two "from scratch," two AI-refined), with 7-day washout to minimize recall bias and all sessions video-recorded to robustly quantify both inter- and intra-rater variability. These multi-annotator labels define the reference standard for the challenge test phase (Maleki et al., 16 Apr 2025).

3. Evaluation Metrics and Benchmarking Framework

Submitted models must be Docker-packaged (MLCube format) and executed on the Synapse platform. The evaluation metrics combine classical image overlap and boundary precision with clinical volumetric change endpoints:

• Dice Similarity Coefficient (DSC): $\frac{2|X \cap Y|}{|X| + |Y|}$ quantifies segmentation overlap.
• Hausdorff Distance (95%): $H_{95}(A,B) = \max\{h_{95}(A,B), h_{95}(B,A)\}$ captures boundary outliers robustly.
• Normalized Surface Distance (NSD): Quantifies the proportion of boundary voxels within a set tolerance.
• Sensitivity, Specificity, Precision: Assessed per lesion, supporting robust clinical interpretation.

Final rankings aggregate ranks across all metrics (DSC, $H_{95}$, NSD, sensitivity, specificity, precision) per case, with DELPHI-inspired statistical testing for robustness. While segmentation metrics drive the leaderboard, volumetric response thresholds (≥30% progression, ≥20% response) following RANO-BM are used in post-challenge analyses for clinical applicability (Maleki et al., 16 Apr 2025).

As a baseline, nnU-Net ensembles on multi-institutional BM data achieved average DSC of approximately 0.75–0.80 for whole-tumor and core subregions in the 2023/2024 editions (Maleki et al., 16 Apr 2025).

4. Dataset Release, Challenge Phases, and Open Science

The challenge unfolds over three main phases:

• Training: 1,296 cases with ground-truth masks.
• Validation: 179 cases without segmentation masks for unbiased hyperparameter tuning.
• Test: 303 hidden cases plus 75 Lighthouse multi-annotator cases.

Key timeline steps:

• June–July 2025: Data release via Synapse.
• August 2025: Submission deadline.
• September 2025: Leaderboard and workshop.
• Post-challenge: Public release of Lighthouse set and previous years' datasets.

All 2023 and 2024 pre-treatment BM datasets will be released at the challenge's conclusion, rendering the public resource the largest and most diverse reference for BM segmentation research (Maleki et al., 16 Apr 2025).

5. Impact, Clinical Significance, and Future Directions

The BraTS-METS 2025 Lighthouse Challenge delivers the first large-scale, multi-annotator, pre- and post-treatment BM MRI dataset with fully transparent evaluation and high-fidelity annotation protocols, specifically tailored to bridge the gap between algorithmic innovation and clinical deployment (Maleki et al., 16 Apr 2025). Rigorously characterizing human raters' intra- and inter-observer variability establishes realistic performance targets for AI-driven systems. Consistent, publicly benchmarked volumetric tracking of metastases is expected to streamline radiological workflows and catalyze personalized therapeutic decision-making.

The open release of multi-edition, multi-institutional, multi-annotator data is positioned to accelerate research on generalizable MRI segmentation, promote the adoption of volumetric response metrics in daily practice, and support broader clinical outcome studies leveraging automated image analysis (Maleki et al., 16 Apr 2025).