---
license: cc-by-3.0
task_categories:
  - image-segmentation
tags:
  - medical
  - histopathology
  - h-and-e
  - nuclei
  - nucleus-segmentation
  - instance-segmentation
  - pan-cancer
  - tcga
pretty_name: Pan-Cancer-Nuclei-Seg (manual subset)
size_categories:
  - 1K<n<10K
---

# Pan-Cancer-Nuclei-Seg — Manual Subset

Gold-standard **manual** nucleus segmentation patches from the TCIA
*Pan-Cancer-Nuclei-Seg* resource (Hou et al., *Scientific Data* 2020): **1,356**
manually segmented **256×256 H&E** patches with per-nucleus instance masks,
spanning **14 TCGA cancer types**.

> **Scope — please read.** This repository contains **only the manual subset**,
> not the full Pan-Cancer-Nuclei-Seg resource. The headline "~5 billion nuclei"
> component (5,060 whole-slide images, ~666 GB of CSV polygon vertices) is
> **algorithm-generated** (a U-Net pipeline, *not* manual ground truth), is
> distributed **without source images** (the H&E WSIs live in TCGA/GDC), and is
> not hosted here. The 1,356 manual patches are the only part that is both
> gold-standard and image-paired.

- **Modality:** Histopathology — H&E brightfield, 256×256 patches at 40× (~0.25 µm/px)
- **Target:** nucleus segmentation (single foreground class)
- **Ground truth:** manually corrected Mask R-CNN masks agreed by annotators A, B
  and C collectively (per the dataset readme)
- **License:** CC BY 3.0
- **Source:** [TCIA Pan-Cancer-Nuclei-Seg](https://www.cancerimagingarchive.net/analysis-result/pan-cancer-nuclei-seg/)
  · Stony Brook BMI [Box `cnn-nuclear-segmentations-2019`](https://stonybrookmedicine.app.box.com/v/cnn-nuclear-segmentations-2019)

## Cancer types (14)
BLCA, BRCA, CESC, COAD, GBM, LUAD, LUSC, PAAD, PRAD, READ, SKCM, STAD, UCEC, UVM
(~91–100 patches each, balanced).

## Columns
| Column | Type | Notes |
|---|---|---|
| `patch_id` | int32 | Patch-ID from the dataset readme (non-contiguous, 1..1365 with gaps) |
| `image` | Image (RGB) | 256×256 H&E patch (`{id}_crop.png`), original PNG bytes |
| `instance_mask` | Image (16-bit, mode `I;16`) | Consensus manual mask (`{id}_labeled_mask_corrected.png`); 0 = background, 1..N = per-nucleus instance IDs. Original PNG bytes (lossless) |
| `cancer_type` | string | One of the 14 TCGA codes (lowercase) |
| `wsi_id` | string | **TCGA slide barcode** the patch was cropped from (e.g. `TCGA-2F-A9KR-01Z-00-DX1`) — use as the cross-dataset dedup key |
| `x`, `y` | int32 | Top-left coordinate of the crop in the source WSI |
| `size_original` | int32 | Crop side length in source-WSI pixels before resize |
| `size_in_40x` | int32 | Crop side length at 40× (400) |
| `num_instances` | int32 | Number of nuclei in the patch (`= instance_mask.max()`) |
| `has_multirater` | bool | True for the 27 patches that also have per-annotator masks at the source |

For binary nucleus-vs-background segmentation, treat `instance_mask > 0` as
foreground. **Read the mask via numpy** (`np.array(mask)`) — passing a 16-bit
`I;16` PNG through `PIL.Image.convert("L")` divides values by 256 and erases most
instance IDs.

**Empty masks:** 32 of the 1,356 patches contain no annotated nuclei
(`num_instances == 0`, background/stroma-only tissue) — these are faithful to the
source, not a packaging error. Filter on `num_instances > 0` if your task
requires at least one nucleus (instance counts range 0–141, mean ≈ 29).

## Provenance, naming and cross-dataset overlap
- **Provenance:** official, author-hosted (Saltz/Kurc, Stony Brook BMI). 1,356
  patches matches the paper.
- **Faithful naming:** this is the *manual subset only* (see scope note above).
- **Ground-truth tier:** the bulk Pan-Cancer-Nuclei-Seg masks are
  algorithm-generated; **only these 1,356 patches are manual**. 27 patches carry
  additional per-annotator masks at the source.
- **Overlap (leakage hazard):** every patch is **TCGA-derived**, so it shares
  source slides with other TCGA histopathology sets — notably **PanNuke**,
  **MoNuSeg**, **MoNuSAC**, **NuCLS** and **TIGER** (on BRCA/PRAD/BLCA/LUAD).
  Deduplicate against those by matching the **`wsi_id` TCGA barcode** before any
  joint benchmark.

## Citation
```bibtex
@article{hou2020pancancernucleiseg,
  title   = {Dataset of segmented nuclei in hematoxylin and eosin stained
             histopathology images of ten cancer types},
  author  = {Hou, Le and Gupta, Rajarsi and Van Arnam, John S. and Zhang, Yuwei
             and Sivalenka, Kaustubh and Samaras, Dimitris and Kurc, Tahsin M.
             and Saltz, Joel H.},
  journal = {Scientific Data},
  volume  = {7},
  number  = {1},
  pages   = {185},
  year    = {2020},
  doi     = {10.1038/s41597-020-0528-1}
}
```
TCIA data citation: Hou, L. et al. (2019). Dataset of Segmented Nuclei in
Hematoxylin and Eosin Stained Histopathology Images of Ten Cancer Types
(Pan-Cancer-Nuclei-Seg). The Cancer Imaging Archive.
doi:10.7937/TCIA.2019.4A4DKP9U