---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: volume_id
    dtype: string
  - name: slice_id
    dtype: int32
  - name: pre
    dtype: image
  - name: post1
    dtype: image
  - name: pre_series
    dtype: string
  - name: post1_series
    dtype: string
  - name: mask
    dtype: image
  - name: is_tumorous
    dtype: bool
  - name: menopause
    dtype: int32
  - name: race
    dtype: int32
  - name: er
    dtype: int32
  - name: pr
    dtype: int32
  - name: her2
    dtype: int32
  - name: tubule_tumor_grade
    dtype: int32
  - name: nuclear_tumor_grade
    dtype: int32
  - name: mitotic_tumor_grade
    dtype: int32
  - name: nottingham_grade
    dtype: int32
  - name: breast_density
    dtype: int32
  - name: recurrence
    dtype: int32
  splits:
  - name: train
    num_bytes: 27873069458
    num_examples: 156202
  download_size: 30304389430
  dataset_size: 27873069458
license: cc-by-nc-4.0
task_categories:
- image-classification
- image-segmentation
- object-detection
language:
- en
tags:
- mri
- medical
- breast
- cancer
- tumor
pretty_name: Duke Breast Cancer MRI
size_categories:
- 100K<n<1M
---
# DUKE-BREAST-CANCER-MRI — 2D Slice-Based Multisequence Breast MRI Dataset with Lesion Segmentation

This dataset is derived from the **Duke Breast Cancer MRI** collection, a large-scale breast MRI dataset with pathologically confirmed lesions, hosted on **The Cancer Imaging Archive (TCIA)**. The original collection provides multi-sequence 3D breast MRI volumes with bounding-box lesion annotations and comprehensive clinical, pathological, and molecular metadata.
Here, we provide a **2D slice-based version**, extracted from two of the available MRI sequences per patient — **pre-contrast T1-weighted (pre)** and **post-contrast T1-weighted (post1)** — paired slice-for-slice with the corresponding lesion segmentation mask and patient-level clinical/pathological metadata.
The dataset is designed for multimodal breast lesion segmentation, classification, and prognostic modeling on breast MRI.


## 📦 Dataset Structure

The dataset has a single `train` split, with each row corresponding to a single 2D axial slice from a patient's coregistered pre / post1 volumes:

| Field | Description |
|---|---|
| `volume_id` | Unique patient/case identifier (e.g., `Breast_MRI_001`) |
| `slice_id` | Index of the axial slice within the volume |
| `pre` | 2D pre-contrast T1-weighted MRI slice |
| `post1` | 2D post-contrast T1-weighted MRI slice (first post-contrast acquisition) |
| `pre_series` | Series UID identifier for the pre-contrast sequence |
| `post1_series` | Series UID identifier for the post1-contrast sequence |
| `mask` | 2D binary lesion segmentation mask (bounding box annotation) |
| `is_tumorous` | Boolean — whether a lesion is present on this slice |
| `menopause` | Menopausal status: `0` (pre-menopausal), `1` (post-menopausal) |
| `race` | Patient race/ethnicity (encoded categorical) |
| `er` | Estrogen receptor status: `0` (negative), `1` (positive) |
| `pr` | Progesterone receptor status: `0` (negative), `1` (positive) |
| `her2` | HER2 status: `0` (negative), `1` (positive), `2` (equivocal) |
| `tubule_tumor_grade` | Tubule formation grade (Nottingham): `1`, `2`, or `3` |
| `nuclear_tumor_grade` | Nuclear pleomorphism grade (Nottingham): `1`, `2`, or `3` |
| `mitotic_tumor_grade` | Mitotic count grade (Nottingham): `1`, `2`, or `3` |
| `nottingham_grade` | Combined Nottingham histologic grade: `1`, `2`, or `3` |
| `breast_density` | Breast tissue density (BI-RADS): `1`, `2`, `3`, or `4` |
| `recurrence` | Binary recurrence outcome: `0` (no recurrence), `1` (recurrence) |

`menopause`, `race`, `er`, `pr`, `her2`, `tubule_tumor_grade`, `nuclear_tumor_grade`, `mitotic_tumor_grade`, `nottingham_grade`, `breast_density`, and `recurrence` are patient-level attributes and are therefore identical across all slices belonging to the same `volume_id`.


## ⚙️ Preprocessing

- For each patient, preprocessed **pre-contrast T1** and **post1-contrast T1** DICOM series were extracted and converted to image format
- Volumes were sliced into 2D axial slices; the bounding-box lesion segmentation mask was created per slice from annotated lesion regions
- Patient-level clinical and pathological metadata (`menopause`, `race`, `er`, `pr`, `her2`, tumor grades, `breast_density`, `recurrence`) were broadcast from the Duke Breast Cancer MRI clinical database to every slice belonging to that patient
- All slices (lesion-containing and background) were exported; use `is_tumorous` to filter or balance
- The following patient cases were excluded during processing: `Breast_MRI_065`, `Breast_MRI_120`, `Breast_MRI_279`, `Breast_MRI_596`, `Breast_MRI_700`, `Breast_MRI_433`, `Breast_MRI_627`

---

## 🚀 Usage

```python
from datasets import load_dataset
import matplotlib.pyplot as plt
import numpy as np

ds = load_dataset("YOUR_USERNAME/DUKE_BREAST_CANCER_MRI", split="train")

# Grab a lesion-containing slice and view both sequences plus the mask overlay
sample = next(s for s in ds if s["is_tumorous"])

fig, axes = plt.subplots(1, 3, figsize=(14, 4))

# Pre-contrast with mask
axes[0].imshow(sample["pre"], cmap="gray")
axes[0].imshow(np.array(sample["mask"]), cmap="jet", alpha=0.4)
axes[0].set_title("Pre-Contrast + Mask")
axes[0].axis("off")

# Post1-contrast with mask
axes[1].imshow(sample["post1"], cmap="gray")
axes[1].imshow(np.array(sample["mask"]), cmap="jet", alpha=0.4)
axes[1].set_title("Post1-Contrast + Mask")
axes[1].axis("off")

# Enhancement map (post1 - pre)
enhancement = np.array(sample["post1"]) - np.array(sample["pre"])
axes[2].imshow(enhancement, cmap="RdBu_r")
axes[2].set_title("Enhancement (Post1 - Pre)")
axes[2].axis("off")

# Clinical and pathological metadata
info_text = (
    f"{sample['volume_id']} | Slice {sample['slice_id']}\n"
    f"ER: {sample['er']}, PR: {sample['pr']}, HER2: {sample['her2']}\n"
    f"Nottingham Grade: {sample['nottingham_grade']} | "
    f"Breast Density: {sample['breast_density']}\n"
    f"Recurrence: {sample['recurrence']}"
)
fig.suptitle(info_text, fontsize=10)
plt.tight_layout()
plt.show()
```

---

## 🏷️ Feature Descriptions

### Receptor Status
Molecular markers predictive of treatment response and prognosis:
- **ER (Estrogen Receptor)**: `0` = Negative, `1` = Positive
- **PR (Progesterone Receptor)**: `0` = Negative, `1` = Positive
- **HER2 (Human Epidermal Growth Factor Receptor 2)**: `0` = Negative, `1` = Positive, `2` = Equivocal

### Nottingham Histologic Grading System
Combined assessment of tumor differentiation and proliferation. Comprises three components, each graded 1–3:
- **Tubule Formation** (`tubule_tumor_grade`): Percentage of tumor forming tubular structures
- **Nuclear Pleomorphism** (`nuclear_tumor_grade`): Degree of variation in nuclear size and shape
- **Mitotic Count** (`mitotic_tumor_grade`): Number of mitotic figures per high-power field

**Combined Grade** (`nottingham_grade`): Sum of components; `1–3` (favorable), `4–5` (intermediate), `6–9` (unfavorable)

### Breast Density
BI-RADS density classification (tissue composition affecting lesion detectability):
- `1` = Almost entirely fat
- `2` = Scattered fibroglandular densities
- `3` = Heterogeneously dense
- `4` = Extremely dense

### Menopausal Status
- `0` = Pre-menopausal
- `1` = Post-menopausal

### Recurrence Outcome
- `0` = No recurrence during follow-up
- `1` = Recurrence during follow-up


## 🧪 Use Cases

- Multimodal (pre / post1-contrast T1) breast lesion segmentation and detection
- Breast cancer receptor status (ER/PR/HER2) classification from imaging
- Pathologic grade prediction from MRI
- Prognostic modeling: predicting recurrence risk from imaging and clinical features
- Enhancement kinetics analysis: examining lesion enhancement patterns across pre/post acquisitions
- Multi-task learning combining segmentation with receptor status, grade, and recurrence prediction
- Representation learning and self-supervised pretraining for breast cancer MRI
- Breast density characterization and tissue composition analysis


## ⚠️ Important Notes

- **Class imbalance**: most slices, particularly toward the superior/inferior extremes of a volume, contain no lesion. Use `is_tumorous` to filter or apply weighted sampling.
- **Volume-level splits**: always split train/val/test at the `volume_id` level, never at the slice level, to avoid data leakage between sets.
- **Lesion annotations**: lesion regions are provided as bounding-box masks; voxels within the bounding box may include surrounding tissue rather than tumor proper.
- **Mask-image alignment**: `mask` is in the same 2D slice space as `pre` and `post1` for a given row, so no additional registration is needed to overlay it.
- **Missing data**: a small number of cases may have incomplete clinical or pathological information; refer to the original TCIA collection for metadata details.
- **Patient exclusions**: seven patient cases were excluded from this release.


## 📚 Citation

This dataset is derived from data made available on **The Cancer Imaging Archive (TCIA)** under a **CC BY-NC 4.0** license.
If you use this dataset, please acknowledge [chehablab.com](Chehab Lab) and cite the original Duke collection:

```bibtex
@misc{Whissel2022,
  author = {Whissel, C. M. and Georgiade, G. S. and Sinha, S. and Sinha, U. P.},
  title  = {Duke Breast Cancer MRI [dataset]},
  year   = {2022},
  publisher = {The Cancer Imaging Archive},
  doi    = {10.7937/tcia.d8xv-ke92}
}
```

---

## 📜 License

This dataset is released under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license, consistent with the source data on TCIA.

You may copy, modify, distribute, and use the data **for non-commercial purposes only**, **provided that appropriate credit is given** to the original authors, TCIA, and the Duke Breast Cancer MRI collection.

[![CC BY-NC 4.0](https://licensebuttons.net/l/by-nc/4.0/88x31.png)](https://creativecommons.org/licenses/by-nc/4.0/)

**Chehab Lab @ 2026**