Graph-NuCLS: A Cell-Graph Dataset for Nucleus Classification from NuCLS

Graph-NuCLS is a node-level classification dataset derived from the NuCLS dataset with "main" labels. Each tissue patch is converted into a cell-graph where nodes represent detected cell nuclei and edges encode spatial proximity. The task is predicting the cell type of each nucleus across 7 classes. Note that node features describe cell morphology, texture, and color intensity whereas edge features are Euclidean distance in micrometers.

This dataset is part of the paper GrapHist: Graph Self-Supervised Learning for Histopathology.

⚠️ Edge Weight Note: While the architecture in GrapHist supports both positive and negative edge weights, by default edge features represent Euclidean distances—meaning farther nodes have larger, positive values. This can be counterintuitive for many graph neural network models. We recommend experimenting with edge weights, such as using their inverse (e.g., 1/distance) or negative distance (e.g., -distance), to better capture proximity and benefit learning.

Dataset Summary

Property	Value
# Patches	1,694
Avg # Nodes	30.76
Avg # Edges	78.62
# Classes	7

Node Classes

Label ID	Class Name
0	Lymphocyte
1	Macrophage
2	Non-TIL / Non-MQ Stromal
3	Other Nucleus
4	Plasma Cell
5	Tumor (Mitotic)
6	Tumor (Non-Mitotic)

Data Structure

graph-nucls/
├── README.md
├── animation.gif
├── data/
│   ├── {SLIDE_ID}_id-5ea4...left-11371_top-54469_bottom-54761_right-11671.pt
│   └── ...                                       
└── splits/
    ├── fold_1_train.csv
    ├── fold_1_test.csv
    ├── ...
    └── fold_5_test.csv

Each .pt file is a PyTorch Geometric Data object with the following attributes:

Attribute	Shape	Description
x	[num_nodes, 96]	Node feature matrix
edge_index	[2, num_edges]	Graph connectivity in COO format
edge_attr	[num_edges, 1]	Edge features
labels	[num_nodes]	Per-node class label
sample_id	str	Unique patch identifier

splits/fold_N_{train|test}.csv

Each fold CSV maps slides to their hospital and split assignment:

slide_name,hospital,type,fold
TCGA-E2-A14N-DX1,E2,train,1
TCGA-C8-A131-DX1,C8,train,1
...

To assign graph files to a fold split, match the SLIDE_ID prefix in the filename against the slide_name column.

---

Quick Start

import torch
from torch_geometric.data import Data

# Load a single graph
graph = torch.load("data/TCGA-A1-A0SP-DX1_id-5ea4095addda5f8398977ebc_left-11371_top-54469_bottom-54761_right-11671.pt", weights_only=False)

print(graph)
# Data(x=[26, 96], edge_index=[2, 63], edge_attr=[63, 1], sample_id='...', labels=[26])

print(f"Nodes: {graph.x.shape[0]}, Edges: {graph.edge_index.shape[1]}")
print(f"Node labels: {graph.labels}")

---

Citation

If you use this dataset, please cite both our work, and the original NuCLS dataset:

GrapHist (this dataset):

@misc{ogut2026graphist,
    title={GrapHist: Graph Self-Supervised Learning for Histopathology}, 
    author={Sevda Öğüt and Cédric Vincent-Cuaz and Natalia Dubljevic and Carlos Hurtado and Vaishnavi Subramanian and Pascal Frossard and Dorina Thanou},
    year={2026},
    eprint={2603.00143},
    url={https://arxiv.org/abs/2603.00143}, 
}

NuCLS (source annotations):

@article{amgad2022nucls,
  title={NuCLS: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer},
  author={Amgad, Mohamed and Atteya, Lamees A and Hussein, Hagar and Mohammed, Kareem Hosny and Hafiz, Ehab and Elsebaie, Maha AT and Alhusseiny, Ahmed M and AlMoslemany, Mohamed Atef and Elmatboly, Abdelmagid M and Pappalardo, Philip A and others},
  journal={GigaScience},
  volume={11},
  pages={giac037},
  year={2022},
  publisher={Oxford University Press}
}

---

License

This dataset is released under the CC BY-NC-SA 4.0 license.