Deep learning-ready dataset of pretreatment 3D dynamic contrast-enhanced MRI (DCE-MRI) scans from 2,072 breast cancer patients across three clinical trials: I-SPY2 (n=982), I-SPY1 (n=172), and Duke (n=918). Features standardized 3D NIfTI volumes, tumor segmentations, and harmonized clinical metadata designed for reproducible research in breast cancer imaging analysis. Nifti files preserve original DICOM intensity values (converted from uint16 to float64) without any preprocessing, normalization, or manipulation. All medical data processing code is available in the dataset-specific subdirectories for full transparency.

Demo: https://colab.research.google.com/github/naomifridman/BreastDCEDL/blob/main/BrestDCEDL_zenodo_demo.ipynb

For Full methodology, Read on arXiv

@article{fridman2025breastdcedl,
  title={BreastDCEDL: A Comprehensive Breast Cancer DCE-MRI Dataset and Transformer Implementation for Treatment Response Prediction},
  author={Fridman, Naomi and Solway, Bubby and Fridman, Tomer and Barnea, Itamar and Goldstein, Anat},
  journal={arXiv preprint arXiv:2506.12190},
  year={2025},
  doi={10.48550/arXiv.2506.12190}
}

The MinCrop version provides three methodicaly selected DCE-MRI time points (pre-contrast, early post-contrast, late post-contrast) cropped to 256×256 pixels around the main tumor. This version has been successfully used to train deep learning models achieving state-of-the-art results for pCR (AUC 0.94) and HER2 (AUC 0.74) prediction in published research.

The dataset is suitable for a wide range of computational approaches including radiomics feature extraction, traditional machine learning, and deep learning. The standardized CSV metadata format with numerically encoded clinical variables and 3D NIfTI imaging files ensures compatibility with all major programming languages and medical imaging software. Examples in this repository demonstrate implementation in Python using PyTorch and HuggingFace infrastructure, though the data format supports any computational framework.

Key features:

• 3 tumor-centered scans per patient: Pre-contrast, early post-contrast, late post-contrast
• Standardized size: All scans cropped to 256×256 pixels around the main tumor
• Fully available on Zenodo: Download MinCrop Dataset
• Clinical relevance: These three time points are specifically selected by radiologists for tumor identification, characterization, and segmentation in clinical practice

The Full version preserves complete spatiotemporal information for researchers investigating advanced temporal dynamics, developing novel preprocessing methods, or requiring the full field of view. Contains all acquired time points (3-12) at original resolution, enabling deep exploration of contrast enhancement kinetics and spatial relationships beyond the tumor region.

Availability:

• I-SPY1 (Full): Download from Zenodo
• I-SPY2 (Full): Download from breastdcedl-ispy2 TCIA
• Duke: Download from TCIA and convert using provided code

DCE MRI 3D Data Organization

The dataset was split into training, validation, and test sets using stratified random sampling to ensure balanced distributions of pCR, HER2, and HR status across all splits. This multi-label stratification maintains consistent class proportions, enabling fair model comparison and preventing bias from imbalanced sampling. Reported performance metrics are calculated on the held-out test set.

*N = number of patients with available labels for each biomarker

Three standardized classification tasks are provided for benchmarking:

¹Results from Fridman et al., 2025 - BreastDCEDL
²Results from Fridman & Goldstein, 2025 - Dual-Attention ResNet

• Sequences: 3-12 time points (typically 7)
• Annotations: Full 3D tumor segmentations

• Sequences: 3-5 usable DCE scans
• Annotations: Full 3D tumor segmentations

• NAC subset: 298 patients with pCR labels (only 31% received neoadjuvant chemotherapy)
• Sequences: 1 pre-contrast + 2-4 post-contrast scans
• Annotations: Bounding box of largest tumor (no full segmentation)

Open BreastDCEDL_demo.ipynb to visualize and explore example data included in this repository.

1. Download the MinCrop dataset from Zenodo
2. Open BreastDCEDL_demo_on_local_data_min_crop.ipynb
3. Follow the notebook to explore and analyze the data

Dynamic Contrast-Enhanced MRI (DCE-MRI) is a key imaging technique for breast cancer evaluation. It captures tissue perfusion dynamics through sequential 3D scans before and after contrast agent administration. In breast cancer:

• Enhancement patterns: Malignant tumors typically show rapid initial enhancement followed by washout or plateau
• Clinical protocol: Radiologists analyze pre-contrast and multiple post-contrast phases (typically 3-12 time points)
• Predictive value: Enhancement dynamics correlate with treatment response and can predict pathological complete response (pCR) to neoadjuvant therapy

For detailed methodology, see Fridman et al., 2025

If you use the BreastDCEDL dataset or code in your research, please cite the article:

@article{fridman2025breastdcedl,
  title={BreastDCEDL: A Comprehensive Breast Cancer DCE-MRI Dataset and Transformer Implementation for Treatment Response Prediction},
  author={Fridman, Naomi and Solway, Bubby and Fridman, Tomer and Barnea, Itamar and Goldstein, Anat},
  journal={arXiv preprint arXiv:2506.12190},
  year={2025},
  doi={10.48550/arXiv.2506.12190}
}

@dataset{fridman2025breastdcedl_dataset,
  author       = {Fridman, Naomi and others},
  title        = {BreastDCEDL: Curated DCE-MRI Dataset},
  year         = {2025},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15627233}
}

Directories: I-SPY1/, I-SPY2/, and DUKE/ contain dataset-specific code for preprocessing metadata, DICOM exploration, and conversion to NIfTI format.

• Zenodo Repository: https://zenodo.org/records/15627233

• I-SPY1: Newitt, D., Hylton, N., on behalf of the I-SPY 1 Network and ACRIN 6657 Trial Team (2016). Breast DCE-MRI Data and Segmentations from Patients in the I-SPY 1/ACRIN 6657 Trials [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/K9/TCIA.2016.HdHpgJLK

• I-SPY2: Li, W., Newitt, D. C., Gibbs, J., Wilmes, L. J., Jones, E. F., Arasu, V. A., ... & Hylton, N. M. (2022). I-SPY 2 Breast Dynamic Contrast Enhanced MRI Trial (ISPY2) (Version 1) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.D8Z0-9T85

• Duke: Saha, A., Harowicz, M. R., Grimm, L. J., Kim, C. E., Ghate, S. V., Walsh, R., & Mazurowski, M. A. (2021). Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations (Duke-Breast-Cancer-MRI) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.e3sv-re93

Please refer to the original data sources for licensing information.