What it is
Most public nuclei datasets come from a handful of organs or from fluorescence images. NuInsSeg was built from H&E-stained tissue sections across 31 organs, 23 human and 8 mouse, imaged at 40x at the Medical University of Vienna. Each patch was cut from a larger 2048x2048 field of view, and every nucleus in it was outlined by hand.
What sets it apart is the second kind of mask. Wherever the tissue was too complex for anyone to outline nuclei reliably, the authors drew an ambiguous area mask. That is honest about where the ground truth stops, and it lets a model be trained and scored without being punished for regions no person could label either.
| Fact | Value |
|---|---|
| Record | Zenodo 10518968, version 1.0, published 2024-01-16 |
| Paper | Mahbod et al., Scientific Data 11, 295 (2024) |
| Images | 665 patches, 512x512 pixels, 8-bit color, lossless PNG |
| Magnification | 40x (TissueFAXS on a Zeiss Axio Imager Z1) |
| Nuclei | 30,698 (human 23,247 in 472 patches, mouse 7,451 in 193 patches) |
| Organs | 23 human, 8 mouse |
| Masks in the source | labeled and binary masks, border-removed masks, distance maps, weighted masks, ambiguous area masks, and the original ImageJ outlines |
| Size | one zip, about 1.6 GB |
| License | CC BY 4.0 |
Human organs: cerebellum, cerebrum, colon, epiglottis, jejunum, kidney, liver, lung, melanoma, muscle, oesophagus, palatine tonsil, pancreas, peritoneum, placenta, salivary gland, spleen, stomach (cardia), stomach (pylorus), testis, tongue, umbilical cord and urinary bladder. Mouse organs: bone (femur), fat (subscapularis), heart, kidney, liver, muscle (tibialis anterior), spleen and thymus.
All counts in this table are the paper's and the record's own.
Browse it in DataTorch
Dataset · datatorch/nuclei-segmentation-nuinsseg · 666 files
Nuclei Segmentation (NuInsSeg)
Nuclei instance segmentation dataset for H&E-stained histology (NuInsSeg, Mahbod et al. 2024, Medical University of Vienna). 665 image patches of 512x512 pixels at 40x magnification from 31 human and mouse organs, with 30,698 manually outlined nuclei as polygon annotations converted from the official labeled masks, plus the dataset's ambiguous-area masks as a separate class. License: CC BY 4.0. Source: https://zenodo.org/records/10518968. Please cite: Mahbod, A. et al. NuInsSeg: A fully annotated dataset for nuclei instance segmentation in H&E-stained histological images. Scientific Data 11, 295 (2024).
Open →
The hosted copy holds the 665 patches with the 30,698 nuclei as polygons, converted from the official labeled masks, and the ambiguous area masks as a class of their own. So you can see at a glance which parts of a patch the authors left unlabeled on purpose.
Open any patch to zoom into a dense region, check an outline against the stain, and correct or add nuclei with your team. Exports come out as COCO.
What people use it for
Training and benchmarking nuclei instance segmentation on H&E, and testing whether a model holds up across organs it has not seen. Nuclei density varies widely between organs, from about 6 nuclei per patch in mouse muscle to over 230 in mouse spleen, so the organ mix is a test in itself.
Things to know before you rely on it, from the paper: the tissue is mostly healthy (melanoma is the one exception), it comes from a single institution, the sections were prepared in two ways (celloidin and paraffin) that look different, and one senior cell biologist chose the fields of view.
Source and citation
Source: https://zenodo.org/records/10518968 (read 2026-11-12). Paper: https://pmc.ncbi.nlm.nih.gov/articles/PMC10940572/. The code to generate the masks is at github.com/masih4/NuInsSeg. Cite: Mahbod, A. et al. NuInsSeg: A fully annotated dataset for nuclei instance segmentation in H&E-stained histological images. Scientific Data 11, 295 (2024).
Outline your own nuclei
Label histology images with your team, review each other's work and export COCO, in the browser.
Frequently asked questions
How many images and nuclei are in NuInsSeg?
665 image patches of 512x512 pixels with 30,698 segmented nuclei. 472 patches (23,247 nuclei) are human tissue and 193 patches (7,451 nuclei) are mouse tissue.
What are the ambiguous area masks?
Masks drawn over regions where accurate and reliable manual annotation was not possible, because the tissue there is too complex to outline nuclei with confidence. They cover the whole dataset, so you can leave those regions out of training or scoring instead of treating them as background.
How were the nuclei annotated?
By hand, as freehand outlines in ImageJ, with no semi-automatic tool. Three annotators with a cell biology background drew them, and a senior cell biologist checked and corrected the masks.
Is there an official train and test split?
The paper evaluates with five-fold cross-validation, a random split with a fixed random state, and reports Dice, aggregated Jaccard index (AJI) and panoptic quality (PQ). The hosted copy does not add a split of its own.
What is the license?
Creative Commons Attribution 4.0 International (CC BY 4.0), as stated on the Zenodo record. Cite Mahbod et al., Scientific Data 11, 295 (2024).
Open the hosted copy on DataTorch.
Browse the files in your browser, then download the dataset with a free account.
