MAMA-MIA Dataset: Breast MRI AI Benchmark

Updated 9 May 2026

MAMA-MIA Dataset is an extensively curated multi-center DCE-MRI resource designed for breast tumor segmentation and neoadjuvant treatment response prediction.
It includes gold-standard voxel-wise tumor masks, detailed patient demographics, and rigorous annotation protocols to support fairness audits and performance benchmarking.
The dataset underpins benchmark tasks with metrics like Dice Similarity Coefficient and balanced accuracy, emphasizing both technical performance and demographic equity.

The MAMA-MIA (Multi-center dAta on breast Magnetic resonance for AI: Multicenter Imaging Analysis) dataset is an extensively curated dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) resource designed to benchmark, audit, and advance methodologies for primary breast tumor segmentation and neoadjuvant treatment response prediction. Developed in conjunction with the MAMA-MIA Challenge, the dataset emphasizes cross-institutional generalizability and demographic fairness in algorithmic performance. Its composition, annotation protocols, and evaluation frameworks reflect state-of-the-art standards for robust, equitable benchmarking in medical imaging AI (Parikh et al., 31 Oct 2025, Garrucho et al., 1 Mar 2026, Musah, 3 Aug 2025).

1. Dataset Composition and Patient Demographics

The MAMA-MIA dataset comprises 1,506 pre-treatment T1-weighted DCE-MRI studies aggregated from four major multi-center cohorts hosted on TCIA: DUKE, I-SPY1, I-SPY2, and NACT. Each study consists of a 3D DCE-MRI series, typically containing 50–150 slices per exam, acquired after intravenous contrast administration. The images are delivered as de-identified DICOM volumes encompassing approximately 0.6–1.0 mm in-plane resolution and 2–4 mm slice thickness. Annotation and demographic metadata are key distinguishing features:

Each case is annotated with gold-standard voxel-wise tumor masks by a panel of 16 expert breast radiologists, supported by silver-standard (automated) masks evaluated for quality (Good, Acceptable, Poor, Missed) (Parikh et al., 31 Oct 2025).
Patient-level demographics include ethnicity (74.9% Caucasian, 16.0% African-American, 5.7% Asian, 3.4% other), age (23.2% <40 years, 50.1% 40–55 years, 26.6% >55 years), menopausal status, and BI-RADS breast density, with additional clinical variables (e.g., pathologic complete response, tumor subtype) available for method development and audit (Garrucho et al., 1 Mar 2026).
The external test and validation cohorts consist of 574 cases sourced from three independent European centers (GUMED, KAUNO, HCB), annotated by local radiologists following harmonized semi-automatic protocols (Garrucho et al., 1 Mar 2026).

Cohort	U.S. Training Set	European External Set	Imaging Protocol
DUKE, I-SPY1/2, NACT	1,506 cases	—	T1w DCE-MRI, ~0.6–1.0 mm in-plane
GUMED, KAUNO, HCB	—	574 cases	T1w DCE-MRI, typically 512×512×80

2. Annotation Protocols and Labeling Structure

Manual tumor segmentation is performed on the first post-contrast (p1) DCE-MRI volume, with annotations reviewed and corrected by a consensus of 16 specialized radiologists. All external test data annotations are created with the same rigor by four radiologists per center using semi-automatic tools. The dataset also provides automated segmentation masks from deep learning models (e.g., nnU-Net), accompanied by multi-expert quality ratings to facilitate performance, error, and fairness analysis (Parikh et al., 31 Oct 2025, Garrucho et al., 1 Mar 2026, Musah, 3 Aug 2025).

Each case carries a binary pathologic complete response (pCR) ground-truth (presence/absence of residual invasive disease after neoadjuvant therapy), enabling pCR prediction tasks (Garrucho et al., 1 Mar 2026, Musah, 3 Aug 2025). Both segmentation and pCR classification tasks leverage detailed domain annotations and metadata.

3. Benchmark Tasks and Evaluation Framework

The MAMA-MIA dataset supports two primary tasks:

Task 1: Tumor Segmentation
- Input: p1 DCE-MRI volume (optionally multi-phase).
- Output: voxel-wise binary mask of the primary tumor.
- Reference: expert-corrected annotation.
- Metrics: Dice Similarity Coefficient (DSC), normalized Hausdorff Distance (NormHD).
Task 2: pCR Prediction
- Input: all or selected DCE-MRI phases.
- Output: binary label (pCR/non-pCR).
- Reference: clinical pathology report.
- Metrics: Balanced accuracy ( $\mathcal S_p^\mathrm{cls}$ ), subgroup fairness.

A unified scoring protocol computes the overall challenge rank as a convex combination ( $\lambda=0.5$ ) of performance ( $\mathcal{S}_p$ ) and fairness ( $\mathcal{S}_f$ ):

$\mathcal S \;=\;(1-\lambda)\mathcal S_p+\lambda\mathcal S_f$

Segmentation fairness ( $\mathcal S_f^\mathrm{seg}$ ) and pCR prediction fairness ( $\mathcal S_f^\mathrm{cls}$ ) are explicitly quantified by subgroup consistency across age, menopausal status, and breast density, using maximal subgroup performance gap and equalized odds metrics, respectively (Garrucho et al., 1 Mar 2026).

4. Subgroup Fairness and Bias Audit

Comprehensive audits reveal consistent, reproducible sources of bias intrinsic to the dataset and the automated pipelines:

Age-related bias: Automated segmentation quality (Dice) is systematically lower for <40-year-old patients, persisting even in age-rebalanced cohorts or after controlling for data source (OLS regression $R^2=0.0104$ , $p=1\times10^{-4}$ ). Fairness gap remains (e.g., S=0.0399, $p=0.026$ ) (Parikh et al., 31 Oct 2025).
Ethnic bias: Aggregate ethnic disparities are mild for Dice (DIR~0.89), but boundary errors (HD95) show significant drop for Asian subgroups (p=0.0046; DIR=0.52). Site-specific analysis uncovers extreme disparities within individual centers, notably in I-SPY2 (Dice DPD=10%) (Parikh et al., 31 Oct 2025).
Confounding by data source: Adjusting for site reduces apparent HD95-based disparities by up to 64%, suggesting that volume-based errors (Dice) are more intrinsic, while boundary errors are predominantly site-driven (Parikh et al., 31 Oct 2025).

Mitigation strategies include careful cohort balancing, maintaining granular per-site metadata, and explicit fairness penalties or constraints in model training and evaluation.

5. Challenge Outcomes, Leaderboards, and Methodological Advances

The MAMA-MIA Challenge operationalizes the dataset for broad technical evaluation:

Segmentation Leaderboard: Top Dice scores on the external test set approximate 0.72–0.74 ( $\lambda=0.5$ 0), surpassing the nnU-Net baseline (0.687). Leading methods demonstrate fairness scores ( $\lambda=0.5$ 1) up to 0.96, indicating minimized demographic gap in the test split (Garrucho et al., 1 Mar 2026).
pCR Prediction Leaderboard: Balanced accuracy rarely exceeds 0.54; almost all teams struggle to improve over random or class-balanced baselines, with only a single submission demonstrating clinically significant gains ( $\lambda=0.5$ 2). Highest fairness scores are sometimes obtained by naïve classifiers due to low prevalence and weak imaging signal (Garrucho et al., 1 Mar 2026, Musah, 3 Aug 2025).
Technical advances: MedNeXt architectures with expanded receptive fields (5×5×5) via UpKern, radiomics-based self-normalizing networks, and ensembling strategies provide incremental improvements in both segmentation (Dice up to 0.67, NormHD=0.24) and subgroup pCR prediction (balanced accuracy up to 75% in specific subgroups) (Musah, 3 Aug 2025).

6. Practical Implications and Clinical Significance

Lower performance in segmentation for young and certain ethnic subgroups may result in missed or imprecise tumor contouring, potentially impacting radiotherapy planning and risk stratification, and amplifying downstream healthcare disparities when deployed as clinical decision support tools. The dataset and associated challenge explicitly advocate for continuous fairness auditing, subgroup-stratified performance reporting, and integration of expert quality ratings in model validation (Parikh et al., 31 Oct 2025, Garrucho et al., 1 Mar 2026).

7. Availability, Resources, and Future Directions

Complete MAMA-MIA dataset access, expert annotations, Dockerized leaderboard codebase, and benchmark documentation are available for download and reproducible evaluation via:

Future research directions highlighted include: (1) improved architectures for small/low-contrast lesions, (2) integrating additional physiological variables (e.g., breast density, menstrual cycle), (3) combining domain adaptation with fairness-aware training, and (4) developing hybrid AI–expert consensus pipelines to flag error cases and address annotation ambiguities (Garrucho et al., 1 Mar 2026, Parikh et al., 31 Oct 2025, Musah, 3 Aug 2025).

MAMA-MIA establishes new standards for multicenter, demographically balanced benchmarking in breast MRI AI and provides a rigorous foundation for the study of both predictive performance and algorithmic equity.