MedOtter/CoNIC2022

CoNIC 2022 (Colon Nuclei Identification and Counting)

Public training release of the CoNIC 2022 Challenge: 4,981 non-overlapping 256x256 H&E histopathology patches (colon tissue, 20x, ~0.5 um/pixel) with nucleus instance segmentation, classification and counting annotations.

Faithful-naming note. This mirror contains the public CoNIC training split only (4,981 patches). The challenge test set (~103k nuclei) is permanently held out and is not part of this dataset. The patches are a re-tiled view of the Lizard dataset (whole-image .mat annotations cut into non-overlapping 256x256 tiles).

Composition

• 4,981 RGB patches of 256x256 px, 431,913 annotated nuclei
• 6 nucleus classes + background
• Single train split (the only public release; no official fold split)

Lizard source breakdown (source column)

source	patches	origin dataset
crag	2,304	CRAG
dpath	1,799	DigestPath
glas	702	GlaS
pannuke	112	PanNuke
consep	64	CoNSeP

Schema

Field	Type	Description
image	PIL RGB 256x256	H&E patch, uint8
inst_map	PIL grayscale uint16 256x256	Per-patch unique nucleus instance ID (0 = background)
class_map	PIL grayscale uint8 256x256	Semantic class per pixel (see table below)
patch_id	int	Row index 0..4980 (matches the original .npy ordering)
patch_info	str	Lizard source image token, e.g. consep_1-0000
source	str	Lizard source dataset (crag/dpath/glas/pannuke/consep)
count_*	int (x6)	Per-class nucleus count in the central 224x224 region

Semantic class encoding (class_map)

Value	Class
0	Background
1	Neutrophil
2	Epithelial
3	Lymphocyte
4	Plasma
5	Eosinophil
6	Connective

count_* columns (count_neutrophil ... count_connective) follow the same 1..6 ordering and are taken from the challenge counts.csv (counted within the central 224x224 region only, per the official protocol).

Cross-dataset overlap (leakage warning)

CoNIC is a re-tiled subset of Lizard, which is itself assembled from CRAG, DigestPath, GlaS, PanNuke and CoNSeP. In particular it shares source images with Angelou0516/PanNuke (112 patches), as well as GlaS (702), CRAG (2,304), DigestPath (1,799) and CoNSeP (64). Use the source / patch_info columns to perform source-image level deduplication before co-benchmarking against any of these datasets.

License

CC BY-NC-SA 4.0 (research / non-commercial use). Same license as the original CoNIC / Lizard release.

Citation

@article{graham2024conic,
  title={CoNIC Challenge: Pushing the frontiers of nuclear detection, segmentation, classification and counting},
  author={Graham, Simon and Vu, Quoc Dang and Jahanifar, Mostafa and Weigert, Martin and Schmidt, Uwe and Zhang, Wenhua and others},
  journal={Medical Image Analysis},
  volume={92},
  pages={103047},
  year={2024},
  publisher={Elsevier}
}

@article{graham2021conic,
  title={CoNIC: Colon Nuclei Identification and Counting Challenge 2022},
  author={Graham, Simon and Jahanifar, Mostafa and Vu, Quoc Dang and Hadjigeorghiou, Giorgos and Leech, Thomas and Snead, David and Raza, Shan E Ahmed and Minhas, Fayyaz and Rajpoot, Nasir},
  journal={arXiv preprint arXiv:2111.14485},
  year={2021}
}

@inproceedings{graham2021lizard,
  title={Lizard: A large-scale dataset for colonic nuclear instance segmentation and classification},
  author={Graham, Simon and Jahanifar, Mostafa and Azam, Ayesha and Nimir, Mohammed and Tsang, Yee-Wah and Dodd, Katherine and Hero, Emily and Sahota, Harvir and Tank, Atisha and Benes, Ksenija and others},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={684--693},
  year={2021}
}