MAMA-MIA DCE-MRI Breast Imaging Dataset
- MAMA-MIA DCE-MRI is a large-scale, multi-center breast imaging dataset featuring pre-treatment T1-weighted MRI with expert tumor segmentations and harmonized clinical metadata.
- It combines 1,506 US cases and 574 European cases, integrating imaging, clinical, and demographic variables to support primary tumor segmentation and pCR prediction.
- The dataset’s deliberate protocol heterogeneity and standardized processing workflow enable robust evaluation of domain generalization and subgroup fairness in clinical research.
The MAMA-MIA DCE-MRI dataset is a large-scale, multi-center benchmark of pre-treatment T1-weighted dynamic contrast-enhanced breast MRI created to support primary tumor segmentation, treatment-response modeling, and robustness analysis under substantial protocol heterogeneity. It integrates four collections in The Cancer Imaging Archive—DUKE, I-SPY1, I-SPY2, and NACT—into a harmonized resource of 1,506 cases with expert segmentations and 49 harmonized clinical and demographic variables (Garrucho et al., 2024). Within the MAMA-MIA Challenge framework, the same 1,506-patient United States cohort served as the public training set for primary tumor segmentation and prediction of pathologic complete response, while external evaluation was conducted on 574 European cases to probe cross-continental generalization and subgroup fairness (Garrucho et al., 1 Mar 2026).
1. Origin, scope, and benchmark role
MAMA-MIA was introduced to address the limited availability of large-scale public breast DCE-MRI datasets with expert-labeled segmentations (Garrucho et al., 2024). The dataset focuses on patients with breast cancer who underwent neoadjuvant chemotherapy, and the selected imaging corresponds to pre-treatment DCE-MRI acquired before therapy initiation (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026).
The resource was designed around two closely related uses. First, it supports automatic segmentation of the primary breast tumor on pre-treatment DCE-MRI volumes. Second, in the challenge setting, it supports prediction of pathologic complete response from pre-treatment MRI alone (Garrucho et al., 1 Mar 2026). The challenge formulation is explicitly intended to evaluate both generalizability and fairness, rather than only aggregate internal performance (Garrucho et al., 1 Mar 2026).
At the dataset level, MAMA-MIA is the public United States training cohort of the challenge and consists of cases drawn from TCIA collections. At the challenge level, this training cohort is paired with a private European validation and test cohort, which makes the benchmark unusual in that it combines public development data with external, cross-institutional, cross-continental evaluation (Garrucho et al., 1 Mar 2026).
2. Cohort composition and imaging structure
MAMA-MIA contains 1,506 DCE-MRI cases from four source collections (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026).
| Collection | Cases | Notes in the benchmark |
|---|---|---|
| DUKE | 291 | Included in the public US training cohort |
| ISPY1 | 171 | Included in the public US training cohort |
| ISPY2 | 980 | Largest constituent subset |
| NACT | 64 | Included in the public US training cohort |
The dataset was assembled from open-access TCIA breast DCE-MRI studies in which patients had breast cancer, underwent neoadjuvant chemotherapy, and had pre-treatment DCE-MRI together with either pathologic complete response or five-year survival information (Garrucho et al., 2024). Cases with insufficient contrast enhancement or severe artifacts that impeded segmentation were excluded after expert quality control (Garrucho et al., 2024).
The imaging modality is T1-weighted DCE-MRI with one pre-contrast phase and multiple post-contrast phases (Garrucho et al., 2024). Across the full dataset, the number of phases has mean 6 and range 3–11 (Garrucho et al., 2024). The constituent collections differ markedly in temporal sampling: ISPY1 usually has 3–6 phases, ISPY2 often 7+ phases, DUKE typically 3–6 phases, and NACT 3–7 phases (Garrucho et al., 2024). In nnU-Net-style organization, phases are commonly indexed as 0000, 0001, 0002, and so forth; challenge submissions frequently treated these as separate input channels (Zayim et al., 22 Dec 2025, Musah, 3 Aug 2025).
Acquisition heterogeneity is a central property of the dataset. In the public training cohort, 1,271 cases are axial and 235 sagittal; 1,086 are acquired at 1.5T and 420 at 3T; scanner manufacturers are Siemens, GE, and Philips (Garrucho et al., 1 Mar 2026). The dataset also exhibits substantial geometric variability: slice thickness spans bins from submillimeter to at least 2.5 mm, pixel spacing is predominantly 0.5–0.99 mm, and image sizes vary widely across collections (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026). This heterogeneity is deliberate and is repeatedly framed as central to the benchmark’s value for domain generalization and robustness research (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026).
3. Annotations, labels, and metadata
MAMA-MIA provides voxel-wise expert segmentations of primary tumors and non-mass-enhanced regions (Garrucho et al., 2024). The annotation workflow combined preliminary automatic segmentations with expert correction and verification, yielding a fully annotated release (Garrucho et al., 2024). The released data also include automatic segmentations and expert quality ratings for those automatic masks, creating a dual-label structure often described as “gold” manual labels and “silver” automated labels in later auditing work (Parikh et al., 31 Oct 2025).
The annotation process used a preliminary nnU-Net model to generate initial masks, which were subsequently corrected and verified by 16 experts (Garrucho et al., 2024). Experts worked on the first post-contrast image, and the final mask is defined on that phase and mapped to other phases if needed (Garrucho et al., 2024). Annotation guidelines included segmenting only the primary tumor if secondary tumors were not within the functional tumor volume or volume of interest, avoiding inclusion of normal tissue in non-mass enhancement cases, excluding clips and intramammary lymph nodes, including necrotic tumor areas, and ensuring 3D consistency across planes (Garrucho et al., 2024).
The dataset includes 49 harmonized variables grouped into demographic, clinical, and imaging/acquisition variables (Garrucho et al., 2024). Examples explicitly described in the source materials include age, ethnicity, body mass index, implants, tumor subtype, bilateral and multifocal disease, pathologic complete response, acquisition plane, field strength, fat suppression, scanner manufacturer, matrix size, number of phases, number of slices, slice thickness, pixel spacing, and phase timing (Garrucho et al., 2024). Challenge papers further emphasize age, menopausal status, and breast density because those variables are used in subgroup fairness analysis (Garrucho et al., 1 Mar 2026).
For the public training cohort, pCR labels are distributed as 440 “Yes,” 1051 “No,” and 15 “N.A.” (Garrucho et al., 1 Mar 2026). Tumor subtype is aggregated into Luminal, Triple Negative, HER2-enriched, HER2-pure, and N.A. categories (Garrucho et al., 1 Mar 2026). These metadata are available for training, but at challenge test time the inference setting is imaging-only (Garrucho et al., 1 Mar 2026).
4. Harmonization, file organization, and access
The original source data are DICOM, and the harmonized MAMA-MIA release converts them to NIfTI using pycad (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026). Orientation is standardized by reorienting sagittal images to PSR and axial images to LAS (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026). A uniform folder structure and naming convention were enforced, with phase-specific filenames such as ISPY1_1221_0000.nii.gz and ISPY1_1221_0001.nii.gz (Garrucho et al., 2024).
The benchmark deliberately does not enforce a single downstream preprocessing pipeline. The challenge description states that “No additional preprocessing steps, such as intensity normalization, bias field correction, or voxel resampling, were enforced,” allowing participants to apply task-specific preprocessing (Garrucho et al., 1 Mar 2026). By contrast, the released baseline nnU-Net associated with the dataset uses per-case z-score normalization and isotropic resampling to mm (Garrucho et al., 2024). This distinction is important: format and orientation are standardized at the dataset level, whereas intensity normalization and resampling are method-dependent choices in downstream studies.
Access is provided through Synapse, with DOI 10.7303/syn60868042 and challenge-facing documentation also available through the official challenge website and CodaBench (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026). The public release contains images, final expert segmentations, automatic segmentations, harmonized metadata tables, quality-control tables, and pretrained baseline weights for nnU-Net (Garrucho et al., 2024).
5. Tasks, baselines, and representative uses
The dataset underpins two benchmark tasks in the MAMA-MIA Challenge: primary tumor segmentation and prediction of pathologic complete response (Garrucho et al., 1 Mar 2026). For segmentation, the unified challenge score combines overlap and boundary accuracy with subgroup consistency:
with a task-specific performance term based on DSC and normalized Hausdorff distance, and a fairness term based on disparities across age, menopausal status, and breast density (Garrucho et al., 1 Mar 2026).
The dataset paper released a baseline nnU-Net trained on all 1,506 cases with 5-fold cross-validation and evaluated on the first post-contrast phase, achieving mean validation Dice coefficient (Garrucho et al., 2024). Challenge evaluation on external European data showed that top methods outperformed the baseline nnU-Net under external testing, with the reported example of MIC reaching DSC 0.7360 versus a baseline of 0.6871 (Garrucho et al., 1 Mar 2026).
Subsequent methods have treated MAMA-MIA as both a segmentation benchmark and a substrate for downstream pCR modeling. The MedNeXt-based study “Large Kernel MedNeXt for Breast Tumor Segmentation and Self-Normalizing Network for pCR Classification in Magnetic Resonance Images” selected the pre-contrast image and the first two post-contrast images as a 3-channel input, resampled to isotropic 1.0 mm spacing, used patches, and reported unseen-validation performance of Dice 0.67 and NormHD 0.24 for an ensemble of large-kernel models (Musah, 3 Aug 2025). The same work used radiomic features extracted from predicted segmentations and first post-contrast MRI for pCR classification and reported balanced accuracy of 57% on the unseen validation set, with subgroup values up to 75% in some groups (Musah, 3 Aug 2025).
A separate challenge submission centered on data curation rather than architecture. “Selective Phase-Aware Training of nnU-Net for Robust Breast Cancer Segmentation in Multi-Center DCE-MRI” used the Synapse MAMA-MIA release to analyze center-specific variability, temporal phase selection, and image quality effects on nnU-Net segmentation (Zayim et al., 22 Dec 2025). In that study, a quality-filtered DUKE+NACT subset and early phases 0000–0002 yielded the best validation Dice, approximately 0.72, whereas naive inclusion of the full heterogeneous training pool could reduce performance (Zayim et al., 22 Dec 2025). This does not redefine the dataset itself; rather, it documents a characteristic way in which dataset heterogeneity interacts with modeling choices.
6. Heterogeneity, quality control, and fairness findings
A recurring theme in work on MAMA-MIA is that size alone does not eliminate domain shift. The dataset is explicitly multi-center and heterogeneous with respect to scanner characteristics, temporal sampling, acquisition plane, spatial resolution, and artifact prevalence (Garrucho et al., 2024, Garrucho et al., 1 Mar 2026). One consequence is that method performance can depend strongly on which constituent cohorts are included and how phases are selected.
The phase-aware nnU-Net study reported that DUKE and NACT generally exhibited clearer tumor-to-background contrast and fewer motion artifacts, whereas ISPY1 and ISPY2 more frequently exhibited motion artifacts, reduced contrast, and blurry or inconsistent tumor boundaries (Zayim et al., 22 Dec 2025). In that analysis, including low-quality ISPY scans could impair segmentation performance, and applying CLAHE globally introduced banding and over-sharpening artifacts that deteriorated training stability, particularly for ISPY1 and ISPY2 (Zayim et al., 22 Dec 2025). A plausible implication is that MAMA-MIA is especially informative for data-centric studies of quality-aware training and center-aware sampling, not only for architecture comparison.
Fairness analyses have extended this view from technical heterogeneity to demographic disparities. “Who Does Your Algorithm Fail? Investigating Age and Ethnic Bias in the MAMA-MIA Dataset” audited the automated silver labels distributed with MAMA-MIA and found an age-related bias against younger patients (Parikh et al., 31 Oct 2025). In pooled analysis, Dice performance increased with age, and the reported Dice Disparate Impact Ratio between young and older patients was 0.699, indicating that younger patients were markedly less likely to fall into the top quartile of segmentation quality (Parikh et al., 31 Oct 2025). Importantly, the same paper reports that the age effect remained significant after adjusting for data source and also persisted in a controlled nnU-Net experiment trained on an age-balanced cohort using gold labels (Parikh et al., 31 Oct 2025).
The same audit showed that ethnic disparities are more difficult to interpret from pooled metrics alone. Global Dice differences across ethnic groups were mild, but HD95 showed a significant disparity against the Asian subgroup, with DIR 0.52 (Parikh et al., 31 Oct 2025). Moreover, the study found that site-specific ethnic disparities could be substantially larger than pooled disparities, with an example of Dice demographic parity difference of 10.0% within ISPY2 versus 3.0% globally (Parikh et al., 31 Oct 2025). This suggests that MAMA-MIA’s multi-source composition can both reveal and mask subgroup bias, depending on the level of stratification.
Within the challenge itself, fairness is not an auxiliary consideration but part of the official evaluation protocol. Subgroup consistency across age, menopausal status, and breast density is built directly into the composite score for both segmentation and pCR prediction (Garrucho et al., 1 Mar 2026). As a result, MAMA-MIA functions simultaneously as a segmentation dataset, an outcome-prediction dataset, and a benchmark for robustness and fairness under clinically realistic variation (Garrucho et al., 1 Mar 2026).