BrEaST Dataset for Breast MRI AI Research

• BrEaST dataset is a harmonized, multi-center breast MRI collection that includes 1,000 studies from six European institutions with standardized lesion labels.
• It features comprehensive MRI protocols and preprocessing pipelines ensuring uniformity across vendors and clinical sites for reliable AI model benchmarking.
• The dataset facilitates advanced research in breast cancer detection, classification, and cross-site generalization, driving progress in radiologic AI applications.

The BrEaST dataset refers to a suite of publicly available, rigorously annotated resources supporting AI research in breast cancer imaging, with the most prominent being the European Multi-Center Breast Cancer MRI dataset released by the ODELIA consortium. This dataset comprises multi-institutional, harmonized breast MRI examinations with standardized lesion labels, enabling comprehensive algorithmic benchmarking for breast cancer detection, classification, and cross-site generalization using magnetic resonance imaging (Müller-Franzes et al., 31 May 2025).

1. Dataset Composition and Cohort Structure

The BrEaST dataset includes approximately 1,000 contrast-enhanced breast MRI studies, contributed by six European centers: Cambridge University Hospitals (UK), Mitera Hospital (Greece), Ribera Salud (Spain), Radboud UMC (Netherlands), University Hospital RWTH Aachen (Germany), and UMC Utrecht (Netherlands). Each case provides per-breast binary or ternary class labels: no lesion (normal), benign lesion, or malignant lesion. Lesion subtyping (e.g., DCIS, invasive, unknown) is collapsed at the dataset level into the “malignant” category for benchmarking.

Demographics such as age are provided, recorded in days (converted to years as $$\mathrm{Age}_{\mathrm{years}} = \mathrm{Age}_{\mathrm{days}}/365.25$$), with full site-wise distributions available in accompanying CSV files. The case mix yields approximately 25% malignant, 10% benign, and 65% no-lesion breasts, stratified appropriately for cross-validation splits. All major MRI vendors are represented (GE, Siemens, Philips), with field strengths primarily at 1.5 T, and a subset at 3.0 T.

2. MRI Protocols and Data Standardization

MRI protocols encompass dynamic contrast-enhanced (DCE) T1-weighted sequences and T2-weighted acquisitions, with image parameters harmonized retrospectively. Standard DCE-MRI used 3D spoiled gradient echo (SPGR/GR) pulse sequences with or without fat suppression, echo times ~1.7–4.6 ms, repetition times 4.8–7.1 ms, and flip angles of 7–16°, with 3–5 post-contrast phases per exam. T2-weighted imaging employed 2D/3D spin echo with echo times up to 365 ms and in-plane field-of-view ~200–456 mm. All volumes are distributed as 16-bit NIfTI files at consistent voxel sizes (post-resampling to 0.7×0.7×3.0 mm is recommended for T1w and matching for T2w), facilitating uniform model ingestion regardless of source scanner.

3. Annotation Protocol and Metadata

Lesion annotations are performed by expert radiologists at each site, based on histopathological confirmation or minimum two-year imaging follow-up in ambiguous cases. Only single-label, per-breast annotations are provided (“case-level classification codes” with no pixel masks or bounding boxes). The original data schema separates left and right breasts, while the standardized “unilateral” format merges these into per-breast records. Lesion classes comprise:

• 0: no lesion (control/unremarkable breast),
• 1: benign lesion,
• 2: malignant lesion (including all subtypes).

Ground-truth consensus, label harmonization across sites, and annotation disagreements are not formally quantified in the ODELIA release; Cohen’s $\kappa$ may be computed for multi-reader experiments as $\kappa = \frac{p_o - p_e}{1 - p_e}$, where $p_o$ is observed, $p_e$ expected inter-reader agreement.

Metadata includes fold-level splits (train/val/test), patient identifiers, acquisition center, and age. Data folders encode provenance (e.g., root/CAM/data_unilateral/).

4. Preprocessing, Data Formats, and Model Development

Recommended preprocessing involves:

• Resampling images to 0.7×0.7×3.0 mm grid,
• Masking and patching to 256×256×32 voxels per breast using Otsu thresholding and median axial splits,
• Per-volume z-score normalization (alternative: min-max per sequence).

The images are provided in 16-bit NIfTI format, with detailed metadata in CSV. Examples of recommended file organization and naming conventions are described (e.g., Pre.nii.gz for pre-contrast T1w, Post_1.nii.gz for post-contrast, Sub_1.nii.gz for subtractions, T2.nii.gz). Sites provide their own conversion pipelines for DICOM to NIfTI manipulation. Stratified five-fold cross-validation is provided, balancing per-class and per-center distributions.

The ODELIA release includes a baseline model: a Medical Slice Transformer (MST) using 3-channel input (pre-contrast T1w, first post-contrast subtraction, T2w). The training setup uses random 224×224×32 crops, rotations, flips, Gaussian noise, AdamW with learning rate $10^{-6}$, batch size 1, and cross-entropy loss.

5. Benchmark Tasks and Performance Metrics

Recommended tasks include:

• Multi-class breast classification (no lesion / benign / malignant),
• Binary tasks (malignant vs. non-malignant, benign vs. non-benign),
• Lesion detection or segmentation (if manually annotated; not provided in base dataset),
• Cross-center generalization—crucial for robust model validation.

Evaluation relies on:

• Accuracy: $$\mathrm{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}$$,
• Sensitivity (Recall): $\mathrm{Sensitivity} = \frac{TP}{TP + FN}$,
• Specificity: $\mathrm{Specificity} = \frac{TN}{TN + FP}$,
• Area under ROC curve (AUC), computed by standard trapezoidal integration.

Baseline results (mean over classes/folds): in-distribution accuracy ≈ 85–90%; RUMC as unseen out-of-distribution, accuracy drop ≈ 3–5%. Both sensitivity at 90% specificity and specificity at 90% sensitivity are typically 80–88%.

6. Comparative Context and Best-Practice Recommendations

Compared to prior public datasets, notably CBIS-DDSM (≈2,500 mammograms, no tomosynthesis) and single-institution MRI cohorts (limited scale, less clinical heterogeneity), BrEaST is the first European multi-center, harmonized, large-scale breast MRI dataset with per-breast lesion labels suitable for benchmarking model robustness and generalization (Müller-Franzes et al., 31 May 2025). A plausible implication is that the dataset enables extensive research into domain adaptation and federated learning scenarios, given its cross-vendor and cross-country diversity.

Best practices based on the ODELIA experience include:

• Strict patient-level splits to avoid leakage,
• Use of AUC, sensitivity, specificity at multiple operating points; reporting cross-site and cross-fold metrics,
• Handling class imbalance via class-weighted loss functions or focal loss; possible oversampling strategies,
• Maintaining center stratification in cross-validation due to institutional bias,
• External validation where possible, particularly by holding out one entire center.

7. Data Access, Licensing, and Applications

The BrEaST dataset is distributed under a CC BY-NC 4.0 license. It is accessible from Hugging Face (https://huggingface.co/datasets/ODELIA/ODELIA2025). Citation of Müller-Franzes et al. and the ODELIA consortium is required. The data is suitable for tasks such as:

• Developing and benchmarking deep learning classifiers for breast MRI,
• Researching generalizability and robustness across scanners and clinical sites,
• Pre-training and transfer learning for related radiologic applications,
• Potential future expansion to lesion-level segmentation or annotation challenges.

The dataset’s scale, rigor in curation, and inclusion of diverse acquisition protocols position it as a central resource for breast MRI AI model development and validation (Müller-Franzes et al., 31 May 2025).