MAMA-MIAAI - Here We Go Again: An Extended Breast MRI Dataset with Bias Field Correction and Automated Breast Tissue Segmentation for Radiomics and AI Research

@article{PhysioNet-mama-miaai-herewegoagain-1.0.0,
  author = {Schwarzhans, Florian and Handa, Palak and {Villazan Garcia}, Laura Atenea and Kartal, Serenay and Fraller, Astrid and Zhero, Svyatoslav and Zaric, Olgica and Woitek, Ramona},
  title = {{MAMA-MIAAI - Here We Go Again: An Extended Breast MRI Dataset with Bias Field Correction and Automated Breast Tissue Segmentation for Radiomics and AI Research}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0.0},
  doi = {10.13026/w6wc-jh81},
  url = {https://doi.org/10.13026/w6wc-jh81}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

The publicly available MAMA-MIA breast dynamic contrast-enhanced magnetic resonance imaging (MRI) dataset has become an important resource for artificial intelligence (AI) research on breast cancer, particularly for tumor detection and predicting pathological complete response (pCR) to neoadjuvant chemotherapy (NAC).

Of all the datasets, only the DUKE subset provides 100 whole-breast tissue masks together with a Breast U-Net AI model capable of automatically generating additional masks. These masks are essential for ensuring consistent preprocessing, a critical prerequisite for reliably analyzing tumors, blood vessels, and fibroglandular tissue (FGT) downstream. Signal i ntensity inhomogeneity is caused by external and local magnetic field non-uniformities that necessitate bias-field correction, which is most effective when applied within the tissue region of interest to avoid confounding factors from background voxels. 

To facilitate reproducible analysis in this field, we are developing and releasing an extended version of the "MAMA-MIAAI — Here we go again" dataset. This dataset includes expert-verified, AI-generated, automated, and manually created whole-breast tissue masks, as well as the corresponding bias-field-corrected DCE-MRI scans. The AI-generated masks were produced using a dedicated 3D U-Net model (MRI_BreastSeg), and N4 bias-field correction was performed using the SimpleITK library. The extended dataset contains 1,506 cases, each including a set of bias-field-corrected DCE-MRI scans and its corresponding breast tissue mask, for a total of 7,868 images and 1,506 automatic masks. All files are stored as 3D NIfTI files. Additionally, we are releasing the 94 manual masks used to train the AI model, along with the AI model itself. The AI model is available via a GitHub repository. This resource lowers the barrier to preprocessing, improves comparability between studies, and enables more robust radiomics and AI research in breast cancer.

Background

Breast magnetic resonance imaging (MRI) is a powerful diagnostic tool for tumor diagnosis, staging, and neoadjuvant chemotherapy (NAC) efficacy assessment. As the use of dynamic contrast-enhanced (DCE) MRI becomes more widespread, radiomics and AI approaches are being used to derive quantitative imaging biomarkers from this imaging modality [1,2,3,4]​. However, in multicenter settings, breast MRI data is highly heterogeneous due to differences in scanner hardware, imaging sequences, and acquisition protocols. While this variability reflects real-world practice, it can also limit the generalizability of radiomics features and AI models ​[5,6]. The main contributor to this variability is the bias field, a signal intensity inhomogeneity caused by factors related to the scanner or coil rather than true tissue differences. Bias-field effects degrade image quality and reduce the robustness of radiomic features, especially when combining images from different vendors and sequence types. The N3 and N4 algorithms are the most widely used correction methods ​[7]​. Systematic studies have demonstrated that applying standardized normalization and bias-field correction enhances the stability of features and improves model performance in multi-scanner datasets ​[5,6,8]​. 

The MAMA-MIA dataset is the first publicly available multicenter DCE-MRI cohort of breast cancer patients treated with NAC, providing detailed response data and expert segmentations [9]​. It has become an important benchmark for radiomics and AI research. Many preprocessing steps, including bias-field correction and fibroglandular tissue (FGT) characterization, rely on whole-breast masks. Currently, only the DUKE subset includes 100 whole-breast, FGT, and vessel segmentations, together with a deep learning model (Breast U-Net) ​[10]​, which may limit transferability to other MAMA-MIA sub-cohorts or external datasets ​[6]​. 

To overcome these constraints, we introduce MAMA-MIAI, an extension of MAMA-MIA that incorporates manual whole-breast masks and automated whole-breast segmentations from a dedicated 3D U-Net (MRI_BreastSeg), along with the corresponding N4 bias-field-corrected DCE-MRI scans of 1,506 cases. Our goal is to provide a standardized preprocessing methodology that supports reproducible radiomics and AI research on a large, heterogeneous breast MRI cohort. 

Methods

Data Description and Sources

The previously published MAMA-MIA dataset includes 1,506 pre-treatment T1-weighted (T1w) DCE-MRI scans of breast cancer patients with manual and automated segmentations of breast cancers [9]. The dataset is sourced from four TCIA collections: DUKE, ISPY1, ISPY2, and NACT. This combination of cohorts reflects real-world clinical variability and encompasses bilateral and unilateral breast MRIs acquired across multiple centers with heterogeneous scanning protocols, thereby enhancing its generalizability. No additional preprocessing, such as bias field correction, signal intensity normalization, or voxel resampling, was performed on the MAMA-MIA dataset. Manual and automated segmentations of the whole breast, FGT, and blood vessels were published for a random subset of 100 subjects from the DUKE cohort for pre-contrast T1w MRI sequences [10]​.

To prepare the dataset for our deep learning experiment, the authors selected 291 subjects from the DUKE cohort. They assigned 281 subjects to the training set and 10 subjects to the holdout set. No additional expert-generated masks were created for this cohort. The ISPY1 cohort contains 171 subjects for which no manual masks were previously available. Our experts generated masks for 33 of these subjects, including 23 in the training set and 10 in the holdout set. The ISPY2 cohort contains 980 subjects without previously available manual masks. Expert-generated masks were created for 37 subjects, 27 of which were used for training and 10 of which were used for the holdout set. The NACT cohort includes 64 subjects for which no manual masks were previously available. Expert annotations were generated for 24 subjects, with 14 assigned to the training set and 10 to the holdout set. Additionally, two private datasets called Private Set A and Private Set B were included. These were obtained via a data transfer agreement with the Medical University of Vienna (ERB number 1131/2015, agreement number 2023/030). These two cohorts remain private and are not made available here. Private Set A contains ten subjects, all of which were manually annotated by our expert and are reserved entirely for the holdout set. Private Set B contains 45 subjects, all of which were manually annotated by our expert. Thirty-five subjects were included in the training set, and 10 were included in the holdout set.

Generation of manual segmentation masks

As discussed above, manual 3D segmentation masks for all sources were created in ITK-SNAP software by our expert reader (a biomedical analyst who was specially trained in breast MRI by a board-certified breast radiologist with over 10 years of experience; the analyst had already successfully segmented over 150 breast cancers in the ISPY1 cohort before performing whole-breast segmentations for this work), except for the DUKE dataset, for which previously published whole-breast segmentations were only visually evaluated.

Generation of automated segmentation masks

Automated 3D segmentation masks were generated for all sources using the dedicated 3D U-Net model, MRI_BreastSeg. Training was performed in-house on a virtualized NVIDIA A40 GPU with 12 GB of VRAM running Ubuntu. The only exception was the DUKE dataset, for which automated masks were obtained using the existing Breast U-Net model [10]. 

The evaluation of Breast-U-Net on other datasets revealed suboptimal performance. This motivated the development of MRI_BreastSeg, which aims to achieve more accurate and consistent whole-breast segmentation across all cohorts. 

During training, only T1-weighted pre-contrast MRI scans were used because contrast-enhancing lesions were deemed irrelevant for whole-breast segmentation. The data were split into training and holdout sets. Ten subjects from each cohort were reserved for the holdout set, as described above. The training set was further divided into 80% for training and 20% for testing. MRI_BreastSeg was configured with channel dimensions of [32, 64, 128, 256, 512], stride dimensions of [2, 2, 2, 2], a kernel size of 3, leaky rectified linear unit (ReLU) activation, instance normalization, and a dropout rate of 0.2. The model was trained for up to 200 epochs with a batch size of eight, using early stopping to prevent overfitting. Grad-CAM visualizations were used for interpretability. Training was optimized using focal loss, and performance was evaluated with the Dice similarity coefficient (DSC) across the training, testing, and holdout sets. Focal loss was used to help the model focus on regions that were difficult to segment because the cohort was collected from multiple sources and exhibited substantial heterogeneity in tumor conspicuity and type. The average DSC reported (0.881) was achieved on the holdout set of each cohort using the best version of the final model (61st). 

The ability of MRI_BreastSeg to handle any orientation (e.g., RAS, LPA, LAS, and LPS) and view (axial, sagittal, and coronal) is unique. This capability is enabled by robust, on-the-fly augmentations during training using the MAMA-MIA dataset and supplementary private data. Because these cohorts are collected from diverse sources and heterogeneous clinical settings, MRI_BreastSeg demonstrates strong generalizability. 

Bias Field Corrections

In MRI, signal inhomogeneity, also known as intensity inhomogeneity or a bias field, refers to gradual, spatially varying changes in image intensity caused by factors related to the MRI scanner or coil rather than true tissue differences. In radiomics, where quantitative features are extracted from medical images, these variations can introduce unwanted variability and reduce feature reproducibility. They can also potentially bias predictive models. Therefore, bias field correction is a crucial preprocessing step that ensures radiomic features accurately reflect tissue characteristics rather than scanner-related artifacts.  

In this work, we corrected the bias field by applying the N4 algorithm from the SimpleITK library. We set all parameters to their default values and performed a maximum of 50 iterations at each of five resolution levels. We used the MAMA-MIA dataset as input, together with the automatically generated segmentation masks from MRI_BreastSeg. For each subject, we generally estimated the bias field from the pre-contrast scan. For nine subjects in the ISPY2 cohort where the pre-contrast scan was missing from the source data, the bias field was instead estimated from the first available post-contrast scan (i.e., _0001). The estimated bias field was then applied to all available contrast-enhanced T1w sequences within that pre-treatment session. This ensured uniform intensity correction across the entire dynamic series. We computed the coefficient of variation (CV) as a quality metric to evaluate the improvement in image uniformity before and after the bias correction. The CV was calculated for voxel intensities within each subject’s breast mask. Five percent of the dataset was sampled, and the images were processed in five-image batches with multithreaded loading to reduce computational load. For each case, we computed the CV as the standard deviation divided by the mean of the masked voxel intensities before and after bias correction. The overall CV summary showed a mean CV of 0.6578 before bias correction and 0.5202 after correction. This corresponds to a mean reduction of 0.1376, representing a 20.92% decrease.

Data Description

The MAMA-MIAAI "Here we go again" dataset is organized into two primary top-level folders: Breast Segmentation and Bias Field Correction. The Bias Field Correction folder contains bias-field-corrected pre-treatment T1w MRI scans in compressed NIfTI (.nii.gz) format, organized by data collections (DUKE, ISPY1, ISPY2, and NACT) and their subjects (e.g., DUKE_001, DUKE_002, ISPY_1001, and ISPY_1002). Each subject folder is further divided by the different scanning time points during DCE acquisition (e.g., DUKE_001_0000 = pre-contrast, DUKE_001_0001 = first post-contrast DCE acquisition, and DUKE_001_0002 = second post-contrast DCE acquisition). Most folders use the suffix _0000; however, some use the suffix _0001 to maintain consistency with the original MAMA-MIA dataset naming convention.

The Breast Segmentation folder contains two subfolders: expert-annotated, manual whole-breast segmentation masks in NIfTI (.nii) format for 94 subjects randomly selected from cohorts ISPY1, ISPY2, and NACT, and automated whole-breast segmentation masks for all 1,506 cases in MAMA-MIA. This dual-folder structure allows users to access the bias-field-corrected MRI data and the corresponding segmentation masks. Consistent subject naming across folders enables straightforward mapping between images and annotations, as in the original MAMA-MIA dataset. In total, we are releasing 94 manual whole-breast segmentation masks, 1,506 automated whole-breast segmentation masks, and 1,506 bias-field-corrected DCE-MRI sequences.

Usage Notes

Intended Use of Resource: This resource provides high-quality automated and manual whole-breast tissue masks and N4 bias-field–corrected images to support reproducible radiomics and AI research in breast DCE-MRI. This resource is for research purposes only and is not intended for clinical diagnosis or decision-making.

Limitations: The extended dataset presented here is subject to minor limitations. Whole breast segmentation masks were generated using a single deep learning segmentation model. While visual inspection was performed, minor residual segmentation errors or inaccuracies may have affected the accuracy of the bias field correction. Additionally, the dataset is still limited by the size and demographic characteristics of the original MAMA-MIA dataset. Users should consider these factors when using this dataset for downstream analyses.

Software/Libraries Used: The MAMA-MIAAI dataset was managed and processed using a combination of Python libraries and frameworks. MONAI was used to implement MRI_BreastSeg. PyTorch and PyTorch Lightning facilitated model training, inference, and distributed data handling. TorchMetrics provided evaluation metrics, such as Dice scores. SciPy, NumPy, and Pandas were used for numerical computations and structured data handling, including organizing metadata and tracking training logs. SimpleITK was used for reading and writing NIfTI images, typecasting, and applying the N4 bias field correction. MRI scans and segmentations in NIfTI format were handled with Nibabel, and file and directory operations were managed with os and glob to navigate the dataset structure, create folders, and organize subject-specific data. Visualization and plotting were performed using Matplotlib. Configuration management was handled using PyYAML and JSON. IT-SNAP software was used to draw the segmentation masks. The method for reusing and redeveloping the AI segmentation model is available on GitHub [11].

Licencing of the MAMA-MIA dataset: The MAMA-MIA data set has been released under a CC BY-NC 4.0 license. Our segmentations and bias field corrected MRI scans qualify as “Adapted Material” from MAMA-MIA and are also released under a CC BY-NC 4.0 license. CC BY-NC 4.0 explicitly grants the right to produce, reproduce, and share Adapted Material for non commercial purposes only.

Release Notes

Version 1.0.0: Initial release.

The MAMA-MIAAI dataset, hosted via PhysioNet, includes automatically generated breast tissue segmentation masks and corresponding N4 bias-corrected MRI scans for all 1,506 subjects in the original MAMA-MIAAI collection. The dataset is provided in compressed NIfTI format. The segmentation masks are stored as .nii files, and the bias-corrected scans are stored as .nii.gz files. This format allows researchers to easily download, extract, and use the data across different operating systems and platforms. Additional documentation, usage instructions, and related resources for the dataset are available on GitHub [12]. The AI segmentation model used to generate the masks, including the pretrained weights, has also been publicly released and is available on GitHub [11].

Ethics

This work used the publicly available, de-identified MAMA-MIA breast MRI dataset, as well as two private datasets. No new human subjects were recruited, and no additional clinical data were collected for this work. Ethical approval and informed consent were not required for the work presented here, as only the automatically generated breast tissue segmentation masks, together with the corresponding N4 bias-corrected MRI scans for all 1,506 subjects from the original MAMA-MIA collection, were released. Ethical approval for the original MAMA-MIA collection was obtained by the dataset providers, as described in [9]. The private A and B cohorts were collected for other research projects and were used solely for training and holdout data analysis during the model development stage. These cohorts were included in the data transfer agreement with the Medical University of Vienna (ERB number 1131/2015, agreement number 2023/030), and they remain private and are not being released here.

Acknowledgements

We thank the original investigators of the MAMA-MIA dataset for making their data publicly available, which enabled this work [9].  Supported by the Federal Ministry for Women, Science and Research of the Republic of Austria and Austria’s Agency for Education and Internationalisation (OEAD).

Conflicts of Interest

The authors declare that they have no competing financial interests or personal relationships that could have influenced the work reported in this paper.

References

1. Jiang Y, Edwards AV, Newstead GM. Artificial intelligence applied to breast MRI for improved diagnosis. Radiology. 2021;298(1):38-46.
2. Ye DM, Wang HT, Yu T. The application of radiomics in breast MRI: a review. Technol Cancer Res Treat. 2020;19:1533033820916191.
3. Militello C, et al. 3D DCE-MRI radiomic analysis for malignant lesion prediction in breast cancer patients. Acad Radiol. 2022;29(6):830-840.
4. McCague C, et al. Introduction to radiomics for a clinical audience. Clin Radiol. 2023;78(2):83-98.
5. Saint Martin MJ, et al. A radiomics pipeline dedicated to Breast MRI: validation on a multi-scanner phantom study. Magnetic Resonance Materials in Physics, Biology and Medicine. 2021;34(3):355-366.
6. Schwarzhans F, et al. Image normalization techniques and their effect on the robustness and predictive power of breast MRI radiomics. Eur J Radiol. 2025;187:112086.
7. Tustison NJ, et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010;29(6):1310-1320.
8. Vovk U, Pernus F, Likar B. A review of methods for correction of intensity inhomogeneity in MRI. IEEE Trans Med Imaging. 2007;26(3):405-421.
9. Garrucho L, et al. A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations. Sci Data. 2025;12(1):453.
10. Lew CO, et al. A publicly available deep learning model and dataset for segmentation of breast, fibroglandular tissue, and vessels in breast mri. Sci Rep. 2024;14(1):5383.
11. Schwarzhans F. MRI_BreastSeg. GitHub. Available from: https://github.com/FloXyPython/MRI_BreastSeg
12. Schwarzhans F. MAMA-MIAAI. GitHub. Available from: https://github.com/FloXyPython/MAMA-MIAAI