# Pan-Cancer-Nuclei-Seg manual subset: 1,356 H&E patches with hand-corrected nucleus masks across 14 TCGA cancer types (TCIA)

## Facts

- Provider: Le Hou, Joel Saltz, Tahsin Kurc et al. (Stony Brook University); mirrored by MedOtter
- License: CC-BY-3.0 (https://www.cancerimagingarchive.net/analysis-result/pan-cancer-nuclei-seg/)
- Size: 3 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/MedOtter/Pan-Cancer-Nuclei-Seg
- Access: public, readable without an account
- DOI: https://doi.org/10.7937/TCIA.2019.4A4DKP9U
- Category: Health & Biomedical
- Source: https://doi.org/10.1038/s41597-020-0528-1
- Page: https://scigantic.com/archive/2b106569-409a-53ae-8d6b-a4def91b3a0e
- Keywords: Health & Biomedical, histopathology, image segmentation, Parquet, Hugging Face

## About

The manually segmented portion of the Pan-Cancer-Nuclei-Seg resource from Stony Brook Biomedical Informatics, published through The Cancer Imaging Archive (TCIA). It has 1,356 H&E patches of 256x256 pixels (40x, about 0.25 microns per pixel) cropped from TCGA whole-slide images, each with a consensus per-nucleus instance mask corrected by hand from Mask R-CNN output. Fourteen cancer types are balanced at roughly 91 to 100 patches each: BLCA, BRCA, CESC, COAD, GBM, LUAD, LUSC, PAAD, PRAD, READ, SKCM, STAD, UCEC and UVM. Each row records the TCGA slide barcode and crop coordinates, so patches can be traced to their source slide and deduplicated against other TCGA nucleus sets; 27 patches also have per-annotator masks at the source and 32 contain no nuclei. The algorithm-generated whole-slide segmentations of about 5 billion nuclei are not included. Parquet under data/ with 16-bit instance masks, about 160 MB (measured 2026-11-21); read masks with numpy rather than converting to 8-bit. Use for pan-cancer nucleus instance segmentation benchmarks. TCIA lists the collection under CC BY 3.0 with DOI 10.7937/TCIA.2019.4A4DKP9U; cite Hou et al., Scientific Data 2020 (doi 10.1038/s41597-020-0528-1).

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("MedOtter/Pan-Cancer-Nuclei-Seg")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `data/train-00000-of-00001.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/2b106569-409a-53ae-8d6b-a4def91b3a0e
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/2b106569-409a-53ae-8d6b-a4def91b3a0e
- Over MCP: `get_schema_card` with id `2b106569-409a-53ae-8d6b-a4def91b3a0e` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Image segmentation datasets](https://scigantic.com/topics/image-segmentation)
- [Health & Biomedical datasets](https://scigantic.com/topics/health-and-biomedical)
- [Histopathology datasets](https://scigantic.com/topics/histopathology)
