Datasets:
File size: 3,329 Bytes
2617f59 502d4a3 2617f59 502d4a3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | ---
license: cc0-1.0
task_categories:
- image-segmentation
tags:
- medical
- pathology
- h-and-e
- breast
- segmentation
- bcss
size_categories:
- n<1K
---
# BCSS — Breast Cancer Semantic Segmentation (Amgad et al. 2019)
Re-hosted mirror of the **Breast Cancer Semantic Segmentation** dataset
(Amgad et al., *Bioinformatics* 2019), originally distributed via the
[PathologyDataScience/BCSS](https://github.com/PathologyDataScience/BCSS)
GitHub repo and rebuilt here from the
[`nabil-m/bcss`](https://huggingface.co/datasets/nabil-m/bcss) HF mirror.
The data is **CC0 1.0** (public domain, no rights reserved); the upstream
codebase is MIT-licensed but covers software, not data. Redistribution is
unrestricted.
## Composition
| Split | ROIs |
|-------|-----:|
| train | 151 |
151 ROI patches extracted from TCGA breast cancer whole-slide images.
Patches are **color-normalized** RGB at the upstream MPP=0.25 µm/px
(40× equivalent), with native ROI resolution typically 2–4k px per
side. There is **no official train/val/test split** — group-shuffle by
`patient_id` downstream for honest evaluation.
## Schema
| Column | Type | Description |
|-------------|----------|-------------------------------------------------------|
| `image` | `Image` | RGB ROI (PNG, color-normalized, variable size) |
| `mask` | `Image` | Indexed 22-class mask (`L`, values 0..21) |
| `image_id` | `string` | Filename stem incl. xmin/ymin |
| `patient_id`| `string` | TCGA-XX-YYYY prefix |
| `xmin` | `int32` | ROI bbox xmin in WSI base-magnification pixels |
| `ymin` | `int32` | ROI bbox ymin in WSI base-magnification pixels |
## Mask labels
| Code | Class | | Code | Class |
|-----:|------------------------|-|-----:|----------------------|
| 0 | outside_roi (don't care) | | 11 | other_immune_infiltrate |
| 1 | tumor | | 12 | mucoid_material |
| 2 | stroma | | 13 | normal_acinus_or_duct |
| 3 | lymphocytic_infiltrate | | 14 | lymphatics |
| 4 | necrosis_or_debris | | 15 | undetermined |
| 5 | glandular_secretions | | 16 | nerve |
| 6 | blood | | 17 | skin_adnexa |
| 7 | exclude | | 18 | blood_vessel |
| 8 | metaplasia_NOS | | 19 | angioinvasion |
| 9 | fat | | 20 | dcis |
| 10 | plasma_cells | | 21 | other |
**Code 0 (`outside_roi`) is a "don't care" region** — the original paper
recommends excluding it from any loss. For binary tumor evaluation,
the canonical foreground is class 1.
## License
CC0 1.0 Universal — public domain. No rights reserved.
## Citation
```bibtex
@article{amgad2019structured,
title = {Structured crowdsourcing enables convolutional segmentation of histology images},
author = {Amgad, Mohamed and Elfandy, Habiba and Hussein, Hagar and others},
journal = {Bioinformatics},
volume = {35},
number = {18},
pages = {3461--3467},
year = {2019},
doi = {10.1093/bioinformatics/btz083}
}
```
|