---
license: other
task_categories:
  - image-segmentation
tags:
  - medical
  - ultrasound
  - breast
  - segmentation
  - busi
size_categories:
  - n<1K
---

# BUSI — Breast Ultrasound Images Dataset

Re-hosted mirror of the Breast Ultrasound Images Dataset (Al-Dhabyani et al., 2020),
collected at Baheya Hospital, Cairo, Egypt and originally distributed via the
Cairo University Scholar page of co-author Aly Fahmy.

This mirror is intended for use with [EasyMedSeg](https://github.com/) and provides:

- A single canonical source (the original Cairo Univ. URL has unstable TLS).
- Parquet schema with one row per image and a single binary mask.
- For images with multifocal lesions, the multiple `_mask_N.png` files in the
  original release are merged into one binary mask via per-pixel logical OR
  (the convention adopted by the paper and by BUS-Set / Thomas et al. 2023).

## Composition

| Class       | Images | Multifocal cases |
|-------------|-------:|-----------------:|
| benign      | 437    | 16               |
| malignant   | 210    | 1                |
| normal      | 133    | 0                |
| **Total**   | **780**| **17**           |

Image dimensions are variable (~190–719 px, average ~500×500 px). Source release
is from 2018; 600 female patients, ages 25–75; GE LOGIQ E9 / E9 Agile B-mode.

## Splits

BUSI has **no official train/val/test split**. This release ships a single
`train` split. Downstream code is expected to define its own splits
(BUS-Set provides reproducible ones).

## Schema

| Column                   | Type      | Description                                                |
|--------------------------|-----------|------------------------------------------------------------|
| `image`                  | `Image`   | Source PNG (RGB)                                           |
| `mask`                   | `Image`   | Binary mask (`L` mode, 0/255) — OR-merged across all mask files |
| `class_label`            | `string`  | `"benign"` / `"malignant"` / `"normal"`                    |
| `image_id`               | `string`  | e.g., `"benign (100)"`                                     |
| `has_multifocal_lesion`  | `bool`    | `True` if the source had >1 mask file                      |

## Known data-quality caveats

Per Pawłowska et al. (2023, *Data in Brief*, [PMC10293973](https://pmc.ncbi.nlm.nih.gov/articles/PMC10293973/)):
- ~235 duplicated images (~19%); 8 lesions appear in both benign and malignant folders (label leakage).
- ~70 axilla (non-breast) images mis-categorized as breast.
- 295 images with overlaid annotations (calipers, doppler markers, text) intersecting the lesion ROI.
- ≥7 images contain a visible biopsy needle.

This mirror **does not modify** the source files (other than mask merging) — the
above issues are inherited as-is. Cite Pawłowska et al. when reporting cleanups.

## License

The original authors do not state an explicit license on the Cairo Univ. page or
in the *Data in Brief* paper. Treat as **research use, with required citation**.
Kaggle uploaders have tagged copies as CC BY 4.0 (uploader-asserted, not
author-confirmed). Do not assume redistribution rights beyond research.

## Citation

```bibtex
@article{Aldhabyani2020BUSI,
  title   = {Dataset of breast ultrasound images},
  author  = {Al-Dhabyani, Walid and Gomaa, Mohammed and Khaled, Hussein and Fahmy, Aly},
  journal = {Data in Brief},
  volume  = {28},
  pages   = {104863},
  year    = {2020},
  doi     = {10.1016/j.dib.2019.104863}
}
```