tum-ai/HASSL-SingleCellBench

---

---

license: gpl-3.0 pretty_name: HASSL SingleCellBench task_categories:

• image-segmentation
• image-classification tags:
• single-cell
• nuclei-segmentation
• cell-segmentation
• instance-segmentation
• biomedical-imaging
• microscopy
• histology
• fluorescence-microscopy
• benchmark
• computer-vision size_categories:
• 1M<n<10M

---

HASSL SingleCellBench

HASSL SingleCellBench is a curated benchmark collection for single-cell, nuclei, and biomedical instance segmentation research. It aggregates selected public microscopy datasets into a unified folder structure for training, evaluation, and benchmarking across heterogeneous imaging modalities, cell types, tissue contexts, and annotation schemes.

This release contains the subset of datasets that could be redistributed in the uploaded benchmark package. Some datasets referenced in the original HASSL dataset overview were intentionally omitted because of licensing or redistribution constraints.

License

This repository is released under the GNU General Public License v3.0 (GPL-3.0).

Important note: this repository aggregates processed versions of multiple third-party benchmark datasets. The GPL-3.0 license applies to the repository packaging, preprocessing code, loaders, and provided metadata where applicable. The underlying datasets may remain subject to their original licenses, citation requirements, and redistribution terms. Users are responsible for complying with the terms of each original dataset.

Data Format

The dataset is stored as NumPy arrays.

For each sample directory:

• original/ contains .npy files with the extracted single-cell or nucleus image.
• mask/ contains .npy files with the corresponding segmentation mask.

A typical segmentation sample path looks like:

N_BCCD/train/original/example.npy
N_BCCD/train/mask/example.npy

For datasets with class-level folders, the same pattern is nested under the relevant class:

N_CoNIC/train/epithelial/original/example.npy
N_CoNIC/train/epithelial/mask/example.npy

Included Datasets

The following datasets are included in this uploaded release:

Folder	Dataset	Cells	Labels	Modality	Cell / Object Types	Original Ref.
N_BCCD	BCCD	90,813	U	Brightfield blood smear	Blood cells	[13]
N_CoNIC	CoNIC	7,696	Annotated	H&E histology	Colon epithelial, stromal, immune, neutrophils, eosinophils	[20]
N_cyto2	Cyto / Cyto2	71,783	U	Mixed microscopy	Mixed cultured cells	[44, 51]
N_databowl	Data Science Bowl 2018	14,902	U	Mixed IF / BF	Mixed species nuclei	[19]
N_DynamicNuclearNet	Dynamic Nuclear Net	347,572	U	Live-cell fluorescence	Cultured human nuclei	[54]
N_iPSC_Morpologies	iPSC morphology data	35,308	Annotated	Multichannel microscopy	Human induced pluripotent stem cells	[45]
N_iPSC_QCData	iPSC quality-control data	35,308	Annotated	Multichannel microscopy	Human induced pluripotent stem cells	[45]
N_lynsec13	LynSec	70,676	Annotated	H&E histology	Lymphoma cells, DLBCL	[29]
N_MoNuSAC	MoNuSAC	28,744	Annotated	H&E histology	Neoplastic, lymphocyte, macrophage, neutrophil	[55]
N_MoNuSeg	MoNuSeg	16,031	U	H&E histology	Mixed tumor and stromal nuclei	[35]
N_NuInsSeg	NuInsSeg	25,293	Annotated	H&E histology	31-organ nuclei	[39]
N_omnipose	Omnipose	37,038	Annotated	Phase contrast and fluorescence	Bacterial and other cells	[11]
N_PanNuke	PanNuke	104,594	Annotated	H&E histology	Tumor, immune, epithelial, stromal, dead	[17, 18]
N_Satorious	Sartorius Cell Instance Segmentation Challenge	34,621	Annotated	Phase contrast	Cortical neurons, astrocytes, SH-SY5Y	[27]
N_tissuenet	TissueNet	866,884	U	Multiplex immunofluorescence	Multiple tissue cell types	[53]

U indicates that the original dataset does not provide explicit annotated cell-type labels in the HASSL benchmark table. For those datasets, the dataset name or folder-level class is used as the label.

Datasets Referenced in the HASSL Overview but Not Included Here

The original HASSL dataset overview contains additional datasets that are not present in this uploaded release. These were omitted because of licensing, redistribution, or packaging constraints.

Not included in this release:

• CPM 15+17 and TNBC
• CryoNuSeg
• IHC TMA
• NeurIPS 2022 Cell-Seg
• Phenoplex
• YeaZ

Users who need these datasets should obtain them directly from the original sources and comply with their individual licenses and access requirements.

Repository Structure

The uploaded dataset follows a hierarchical dataset / split / class / original + mask layout. Depending on the source dataset, class folders may appear before or after the train/test split.

.
├── Dataloader_example.ipynb
├── N_BCCD
│   ├── test
│   │   ├── mask
│   │   └── original
│   └── train
│       ├── mask
│       └── original
├── N_CoNIC
│   ├── test
│   │   ├── connective
│   │   │   ├── mask
│   │   │   └── original
│   │   ├── epithelial
│   │   │   ├── mask
│   │   │   └── original
│   │   ├── esoinophil
│   │   │   ├── mask
│   │   │   └── original
│   │   ├── lymphocyte
│   │   │   ├── mask
│   │   │   └── original
│   │   ├── neutrophil
│   │   │   ├── mask
│   │   │   └── original
│   │   └── plasma
│   │       ├── mask
│   │       └── original
│   └── train
│       ├── connective
│       ├── epithelial
│       ├── esoinophil
│       ├── lymphocyte
│       ├── neutrophil
│       └── plasma
├── N_DynamicNuclearNet
│   ├── test
│   │   ├── mask
│   │   └── original
│   └── train
│       ├── mask
│       └── original
├── N_MoNuSAC
│   ├── Epithelial
│   │   ├── test
│   │   │   ├── mask
│   │   │   └── original
│   │   └── train
│   │       ├── mask
│   │       └── original
│   ├── Lymphocyte
│   ├── Macrophage
│   └── Neutrophil
├── N_MoNuSeg
│   ├── test
│   │   ├── mask
│   │   └── original
│   └── train
│       ├── mask
│       └── original
├── N_NuInsSeg
│   ├── human bladder
│   ├── human brain
│   ├── human cardia
│   ├── human cerebellum
│   ├── human epiglottis
│   ├── human jejunum
│   ├── human kidney
│   ├── human liver
│   ├── human lung
│   ├── human melanoma
│   ├── human muscle
│   ├── human oesophagus
│   ├── human pancreas
│   ├── human peritoneum
│   ├── human placenta
│   ├── human pylorus
│   ├── human rectum
│   ├── human salivory gland
│   ├── human spleen
│   ├── human testis
│   ├── human tongue
│   ├── human tonsile
│   ├── human umbilical cord
│   ├── mouse fat (white and brown)_subscapula
│   ├── mouse femur
│   ├── mouse heart
│   ├── mouse kidney
│   ├── mouse liver
│   ├── mouse muscle_tibia
│   ├── mouse spleen
│   └── mouse thymus
├── N_PanNuke
│   ├── Adrenal_gland
│   │   ├── Connective
│   │   │   ├── test
│   │   │   └── train
│   │   ├── Dead
│   │   ├── Epithelial
│   │   ├── Inflammatory
│   │   └── Neoplastic
│   ├── Bile-duct
│   ├── Bladder
│   ├── Breast
│   ├── Cervix
│   ├── Colon
│   ├── Esophagus
│   ├── HeadNeck
│   ├── Kidney
│   ├── Liver
│   ├── Lung
│   ├── Ovarian
│   ├── Pancreatic
│   ├── Prostate
│   ├── Skin
│   ├── Stomach
│   ├── Testis
│   ├── Thyroid
│   └── Uterus
├── N_Satorious
│   ├── test
│   │   ├── astrocytes
│   │   │   ├── mask
│   │   │   └── original
│   │   ├── neurons
│   │   │   ├── mask
│   │   │   └── original
│   │   └── shsy5y
│   │       ├── mask
│   │       └── original
│   └── train
│       ├── astrocytes
│       ├── neurons
│       └── shsy5y
├── N_cyto2
│   ├── test
│   │   ├── mask
│   │   └── original
│   └── train
│       ├── mask
│       └── original
├── N_databowl
│   ├── test
│   │   ├── mask
│   │   └── original
│   └── train
│       ├── mask
│       └── original
├── N_iPSC_Morpologies
│   ├── Big
│   ├── Long
│   ├── Mitotic
│   ├── RAR-treated
│   └── Round
├── N_iPSC_QCData
│   ├── Cell
│   ├── Debris
│   ├── DyingCell
│   └── MitoticCell
├── N_lynsec13
│   ├── negative
│   ├── non-tumor
│   ├── positive
│   └── tumor
├── N_omnipose
│   ├── bact_fluor_A22
│   ├── bact_fluor_bthai
│   ├── bact_fluor_cex
│   ├── bact_fluor_vibrio
│   ├── bact_fluor_wiggins
│   ├── bact_phase_5I_crop
│   ├── bact_phase_A22
│   ├── bact_phase_PAO1_staph
│   ├── bact_phase_PSVB
│   ├── bact_phase_bthai
│   ├── bact_phase_caulo
│   ├── bact_phase_cex
│   ├── bact_phase_dnaA
│   ├── bact_phase_ecoli_mut
│   ├── bact_phase_francisella
│   ├── bact_phase_ftsN
│   ├── bact_phase_hpylori
│   ├── bact_phase_murA
│   ├── bact_phase_serratia
│   ├── bact_phase_streptomyces
│   ├── bact_phase_vibrio
│   └── bact_phase_wiggins
└── N_tissuenet
    ├── test
    │   ├── mask
    │   └── original
    └── train
        ├── mask
        └── original

For compactness, repeated subtrees are abbreviated in the tree above. In general, all terminal train/test sample folders follow one of these patterns:

.../train/original/*.npy
.../train/mask/*.npy
.../test/original/*.npy
.../test/mask/*.npy

Download

Using the Hugging Face Hub Python library:

from huggingface_hub import snapshot_download

dataset_path = snapshot_download(
    repo_id="tum-ai/HASSL-SingleCellBench",
    repo_type="dataset"
)

print(dataset_path)

Using the Hugging Face CLI:

hf download tum-ai/HASSL-SingleCellBench --repo-type=dataset

If Downloaded as Zip Archives

If this repository is distributed as one .zip archive per top-level dataset folder, extract all archives with:

mkdir -p extracted

for f in *.zip; do
  unzip -q "$f" -d extracted/
done

Example Usage

A basic dataloader example is provided in:

Dataloader_example.ipynb

This notebook demonstrates how the processed benchmark folders can be loaded for downstream model training and evaluation.

Intended Use

This dataset is intended for:

• single-cell and nuclei segmentation benchmarking
• biomedical instance segmentation
• cross-dataset generalization experiments
• microscopy representation learning
• robust evaluation across heterogeneous cell and tissue domains
• development of unified dataloaders for biomedical segmentation datasets

Limitations

The dataset combines benchmarks with different image modalities, preprocessing conventions, annotation schemes, label definitions, and split structures. Users should inspect each folder before training or evaluation and verify that the image-mask pairing, labels, and splits match their experimental setup.

The included release is not identical to the full dataset list described in the original HASSL overview. Some referenced datasets were excluded from this upload because of licensing or redistribution restrictions.

Citation

If you use this benchmark package, please cite this repository and the original datasets used in your experiments.

@dataset{hassl_singlecellbench,
  title = {HASSL SingleCellBench},
  author = {TUM.ai},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/tum-ai/HASSL-SingleCellBench},
  license = {GPL-3.0}
}

Original dataset references from the HASSL benchmark table:

[11] Omnipose: Cutler et al., Nature Methods, 2022.
[13] BCCD: Depto et al., Tissue and Cell, 2021.
[17] PanNuke: Gamper et al., European Congress on Digital Pathology, 2019.
[18] PanNuke extension: Gamper et al., arXiv, 2020.
[19] Data Science Bowl 2018, Kaggle, 2018.
[20] CoNIC: Graham et al., arXiv, 2021.
[27] Sartorius Cell Instance Segmentation Challenge, Kaggle, 2021.
[29] LynSec: Hussein et al., Zenodo, 2023.
[35] MoNuSeg: Kumar et al., IEEE Transactions on Medical Imaging, 2020.
[39] NuInsSeg: Mahbod et al., arXiv, 2023.
[44] Cellpose-SAM: Pachitariu et al., bioRxiv, 2025.
[45] iPSC morphology dataset: Pfaendler, ETH Zurich Research Collection, 2022.
[51] Cellpose: Stringer et al., bioRxiv, 2020.
[53] TissueNet: Van Valen Lab, DeepCell datasets, 2022.
[54] Dynamic Nuclear Net: Van Valen Lab, DeepCell datasets, 2023.
[55] MoNuSAC: Verma et al., IEEE Transactions on Medical Imaging, 2021.

Maintainers

This dataset is maintained by TUM.ai and Marr Lab @ Helmholtz Munich as part of the HASSL single-cell benchmarking workflow.