This dataset is a redistribution of the DUKE portion of MAMA-MIA, whose images come from the TCIA Duke-Breast-Cancer-MRI collection. Both are licensed CC BY-NC 4.0, and use is subject to the TCIA Data Usage Policy (https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/). Requests are reviewed manually.

MAMA-MIA — DUKE split

The DUKE subset of MAMA-MIA, a large multicenter breast cancer DCE-MRI benchmark with expert tumor segmentations. MAMA-MIA combines four TCIA collections; this repository contains only the cases from Duke-Breast-Cancer-MRI, as distributed by MAMA-MIA on Synapse (syn60868042).

The data is unmodified. Image, segmentation, and patient-info files are byte-identical to the MAMA-MIA release. The clinical spreadsheet was exported cell for cell to CSV with no columns or rows dropped, and the train/test split was reshaped to one row per patient. Synapse download manifests and the generic nnUNet pretrained weights are not included; the weights are available from MAMA-MIA at syn61247992.

At a glance


Patients	291 (200 train / 91 test)
Modality	Bilateral axial DCE-MRI, fat-suppressed (290/291)
Phases per patient	3–5 (3: 4 patients, 4: 159, 5: 128)
Segmentations	Expert-corrected primary tumor masks + preliminary automatic masks (binary, 0/1)
Scanners	GE (176), Siemens (115); 1.5 T (136), 3 T (155)
Clinical variables	50 columns (pCR, tumor subtype, receptor status, age, menopause, imaging parameters, …)
Format	NIfTI (`.nii.gz`, int16 images), CSV, JSON
Size	~44 GB, 2,165 data files

Repository layout

images/
  DUKE_001/
    DUKE_001_0000.nii.gz        # phase 0: pre-contrast
    DUKE_001_0001.nii.gz        # phase 1..N: post-contrast
    ...
segmentations/
  expert/DUKE_001.nii.gz        # expert-corrected tumor mask (ground truth)
  automatic/DUKE_001.nii.gz     # preliminary automatic mask (nnUNet), before expert correction
patient_info_files/
  DUKE_001.json                 # per-patient clinical, lesion and imaging summary
metadata/
  clinical_and_imaging_info.csv               # 291 rows x 50 clinical/imaging variables
  clinical_and_imaging_info_legend.csv        # description and allowed values of each variable
  train_test_splits.csv                       # patient_id,split
  preliminary_automatic_segmentations_quality_scores.csv

All phases of a patient and both of its masks share the same voxel grid (shape and affine), so they can be stacked directly. Acquisition times of each phase are in the acquisition_times column of the clinical CSV.

`patient_info_files/*.json`

{
  "patient_id": "DUKE_001",
  "clinical_data": {"age": 41, "menopausal_status": "pre", "breast_density": ""},
  "primary_lesion": {
    "pcr": 0,
    "tumor_subtype": "her2_pure",
    "breast_coordinates": {"x_min": 1, "x_max": 158, "y_min": 5, "y_max": 359, "z_min": 224, "z_max": 445}
  },
  "imaging_data": {
    "bilateral": true, "dataset": "DUKE", "site": "DUKE",
    "scanner_manufacturer": "SIEMENS", "scanner_model": "Avanto",
    "field_strength": 1.5, "echo_time": 1.36, "repetition_time": 4.12
  }
}

breast_coordinates is the voxel bounding box of the breast containing the primary lesion. It is the only field not also in the clinical CSV. The axis order follows MAMA-MIA's own convention; see the MAMA-MIA repository for how it is used for cropping.

`metadata/preliminary_automatic_segmentations_quality_scores.csv`

Quality of the automatic masks, before expert correction:

Column	Meaning
`expert_1_qs`, `expert_2_qs`	Two experts' ratings of the automatic mask: `Good`, `Acceptable`, `Poor`, `Missed`
`dice`, `iou`	Overlap between automatic and expert masks
`hd95`, `msd`	95th-percentile Hausdorff distance and mean surface distance between automatic and expert masks

Median Dice of the automatic masks against the expert masks is 0.906.

Splits

metadata/train_test_splits.csv is the official MAMA-MIA train/test assignment for DUKE.

	Patients	pCR = 1	pCR = 0	pCR missing
train	200	34	155	11
test	91	30	61	0

Tumor subtype	train	test
luminal	83	40
triple_negative	61	24
her2_enriched	31	19
her2_pure	22	8
missing	3	0

Missing data and quirks

Values are kept exactly as in the MAMA-MIA release; no preprocessing has been applied.

Empty for all DUKE patients: nac_agent, days_to_metastasis, mammaprint, acquisition_date.
Largely missing: oncotype_score (92%), patient_size (80%), breast_density (79%), nottingham_grade (34%), weight and bmi_group (20%).
pcr is missing for 11 training patients; tumor_subtype for 3; menopause and ethnicity for a few.
pixel_spacing and acquisition_times are list-valued strings, e.g. "[0.8036, 0.8036]".
Some scanner_model values have trailing spaces (e.g. "SIGNA HDx "), and the legend contains non-breaking spaces and a header column literally named Unnamed: 3, all as in the original spreadsheet.
Four automatic masks have Dice 0 against the expert mask. Three (DUKE_040, DUKE_470, DUKE_891) are rated Missed by both experts; DUKE_847 is rated Good by both despite Dice 0, as in the original release. The expert masks are all non-empty.

Usage

Access is gated: request access on this page, then log in with hf auth login.

hf download Pitt-ICCI-Lab/MAMA-MIA-DUKE --repo-type dataset --local-dir MAMA-MIA-DUKE

import nibabel as nib
import numpy as np
import pandas as pd

root = "MAMA-MIA-DUKE"
clinical = pd.read_csv(f"{root}/metadata/clinical_and_imaging_info.csv")
splits = pd.read_csv(f"{root}/metadata/train_test_splits.csv")

pid = "DUKE_001"
n_phases = int(clinical.loc[clinical.patient_id == pid, "num_phases"].item())
phases = np.stack([
    nib.load(f"{root}/images/{pid}/{pid}_{i:04d}.nii.gz").get_fdata(dtype=np.float32)
    for i in range(n_phases)
])  # (n_phases, X, Y, Z)
mask = nib.load(f"{root}/segmentations/expert/{pid}.nii.gz").get_fdata() > 0

The tabular files can also be loaded with datasets:

from datasets import load_dataset
clinical = load_dataset("Pitt-ICCI-Lab/MAMA-MIA-DUKE", "clinical_and_imaging_info", split="train")

License and terms

MAMA-MIA: CC BY-NC 4.0.
Duke-Breast-Cancer-MRI (TCIA): CC BY-NC 4.0, subject to the TCIA Data Usage Policy.

Non-commercial research use only. Do not attempt to re-identify patients.

Citation

If you use this data, cite the MAMA-MIA paper and the Duke-Breast-Cancer-MRI collection and paper.

@article{garrucho2025,
  title   = {A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations},
  author  = {Garrucho, Lidia and Kushibar, Kaisar and Reidel, Claire-Anne and Joshi, Smriti and Osuala, Richard and
             Tsirikoglou, Apostolia and Bobowicz, Maciej and Riego, Javier del and Catanese, Alessandro and
             Gwoździewicz, Katarzyna and Cosaka, Maria-Laura and Abo-Elhoda, Pasant M and Tantawy, Sara W and
             Sakrana, Shorouq S and Shawky-Abdelfatah, Norhan O and Salem, Amr Muhammad Abdo and Kozana, Androniki and
             Divjak, Eugen and Ivanac, Gordana and Nikiforaki, Katerina and Klontzas, Michail E and
             García-Dosdá, Rosa and Gulsun-Akpinar, Meltem and Lafcı, Oğuz and Mann, Ritse and Martín-Isla, Carlos and
             Prior, Fred and Marias, Kostas and Starmans, Martijn P A and Strand, Fredrik and Díaz, Oliver and
             Igual, Laura and Lekadir, Karim},
  journal = {Scientific Data},
  volume  = {12},
  number  = {1},
  pages   = {453},
  year    = {2025},
  doi     = {10.1038/s41597-025-04707-4}
}

@misc{saha2021duke,
  title     = {Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations},
  author    = {Saha, Ashirbani and Harowicz, Michael R. and Grimm, Lars J. and Weng, Jingxi and Cain, Elizabeth H. and
               Kim, Connie E. and Ghate, Sujata V. and Walsh, Ruth and Mazurowski, Maciej A.},
  publisher = {The Cancer Imaging Archive},
  year      = {2021},
  doi       = {10.7937/TCIA.e3sv-re93}
}

@article{saha2018radiogenomics,
  title   = {A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features},
  author  = {Saha, Ashirbani and Harowicz, Michael R. and Grimm, Lars J. and Kim, Connie E. and Ghate, Sujata V. and
             Walsh, Ruth and Mazurowski, Maciej A.},
  journal = {British Journal of Cancer},
  volume  = {119},
  number  = {4},
  pages   = {508--516},
  year    = {2018},
  doi     = {10.1038/s41416-018-0185-8}
}

Downloads last month: 4

Total file size:

44.2 GB