README.md

---
dataset_info:
  features:
  - name: case_id
    dtype: string
  - name: cohort
    dtype: string
  - name: roi_index
    dtype: int32
  - name: split
    dtype: string
  - name: image
    dtype: image
  - name: nuclei_mask
    dtype: image
  - name: tissue_mask
    dtype: image
  - name: nuclei_count
    dtype: int32
  - name: nuclei_class_ids
    list: int32
  - name: tissue_class_ids
    list: int32
  - name: mpp
    dtype: float32
  - name: magnification
    dtype: int32
  splits:
  - name: train
    num_bytes: 381386223
    num_examples: 206
  download_size: 381439088
  dataset_size: 381386223
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
tags:
- medical
- histopathology
- melanoma
- segmentation
- h&e
license: cc0-1.0
pretty_name: PUMA — Melanoma Nuclei & Tissue Segmentation
---

# PUMA — Panoptic Segmentation of Nuclei and Tissue in Advanced Melanoma

H&E histopathology ROIs from advanced melanoma with expert nuclei **and** tissue
annotations. Mirror of the official Zenodo release for use in the MedOtter
benchmark suite.

> ⚠️ **This is the public training split only: 206 of the challenge's 310 ROIs.**
> The 104 test ROIs are embargoed until **2029-10-10** and are not public
> anywhere. Treat this as a single-split dataset.

## Contents

| | |
|---|---|
| ROIs | 206 (103 primary + 103 metastatic melanoma) |
| Image | 1024×1024 RGB, 40×, ~0.226 µm/px |
| Nuclei | 97,429 annotated, 10 classes |
| Tissue | 5 foreground classes + background |
| Splits | `train` only (see embargo note above) |
| Licence | **CC0 1.0** |

Source images are uncompressed RGBA TIFF with a constant-255 alpha channel;
the alpha is dropped here (verified constant across all 206 files) and pixels
are stored as lossless PNG.

## Label maps

**Tissue** (`tissue_mask`) — values follow the **challenge evaluation**
convention, i.e. what the PUMA leaderboard scores against:

| Value | Class |
|---|---|
| 0 | background (`tissue_white_background`, unpainted) |
| 1 | stroma |
| 2 | blood vessel |
| 3 | tumor |
| 4 | epidermis |
| 5 | necrosis |

⚠️ A **second official map exists** and disagrees: the organizers' QuPath
export script uses `tumor=1, stroma=2, epidermis=3, necrosis=4,
blood_vessel=5`. We use the evaluation convention because that is what the
leaderboard and published PUMA models use. We do, however, keep the QuPath
**paint order** (`tumor < stroma < epidermis < necrosis < blood_vessel`, later
wins), so the masks stay geometrically identical to the official exports.
In practice the order decides only ~0.001% of pixels.

**Nuclei** (`nuclei_mask`) — the single official map:

| Value | Class | | Value | Class |
|---|---|---|---|---|
| 1 | lymphocyte | | 6 | melanophage |
| 2 | tumor | | 7 | endothelium |
| 3 | stroma | | 8 | epithelium |
| 4 | plasma cell | | 9 | neutrophil |
| 5 | histiocyte | | 10 | apoptosis |

The challenge's Track 1 uses a 3-class collapse of this map
(`lymphocyte→1, tumor→2, all others→3`); Track 2 uses all 10.

## Fields

`case_id`, `cohort` (`primary`/`metastatic`), `roi_index`, `split`, `image`,
`nuclei_mask`, `tissue_mask`, `nuclei_count`, `nuclei_class_ids`,
`tissue_class_ids`, `mpp`, `magnification`.

`nuclei_class_ids` / `tissue_class_ids` list the classes actually present in
each ROI — useful because several classes are rare: **necrosis appears in only
9/206 ROIs**, epidermis in 28/206, neutrophil in 30/206. Sampling a handful of
ROIs at random will report zero for those classes.

## Instance-level annotation

Semantic masks merge touching nuclei of the same class. The original QuPath
GeoJSON files are therefore included verbatim at the repo root:

* `01_training_dataset_geojson_nuclei.zip`
* `01_training_dataset_geojson_tissue.zip`

Coordinates are in ROI pixel space (0–1024).

## Provenance & caveats

* Official source: [Zenodo record 15050523](https://zenodo.org/records/15050523)
  (v5, 2025-03-19). Earlier versions differ — **v1/v2 were CC BY 4.0**, v3+ are
  CC0; v3 dropped `metastatic_roi_103` (205 ROIs) and v5 restored a corrected
  copy. The paper cites the stale v3 DOI and the challenge page links v4.
* In the source archive, `training_set_metastatic_roi_103` uses a `.tiff`
  extension while all 205 others use `.tif` — `glob("*.tif")` silently drops it.
* The Zenodo description says "103 primary and 102 metastatic"; the archives
  actually contain **103 + 103** (verified).
* Ground truth is the expert tier: nuclei were initialised by a
  PanNuke-pretrained HoVer-Net, corrected by a medical expert, then reviewed and
  corrected by a dermatopathologist. Tissue was drawn manually throughout.
  Reported human ceiling: nuclei F1 0.857 (intra-observer) / 0.802 (inter);
  tissue Dice 0.90.
* Because the nuclei GT was *seeded* by a PanNuke-pretrained model, models
  pretrained on PanNuke may carry a mild prior advantage on boundary detail.
  This is model lineage, not data leakage.
* No patient/image overlap with PanNuke, MoNuSAC, NuCLS, CoNIC/Lizard, BCSS,
  Pan-Cancer-Nuclei-Seg, DSB2018 or TCGA-SKCM — PUMA is single-institution
  material and carries no cross-reference IDs.

## Citation

Schuiveling M, Liu H, Eek D, Breimer GE, Suijkerbuijk KPM, Blokx WAM, Veta M.
*A novel dataset for nuclei and tissue segmentation in melanoma with baseline
nuclei segmentation and tissue segmentation benchmarks.* GigaScience.
2025;14:giaf011. doi:10.1093/gigascience/giaf011

Dataset: doi:10.5281/zenodo.15050523 (CC0 1.0)