Nuclei Segmentation (NuInsSeg)

NuInsSeg is a fully annotated nuclei instance segmentation dataset for H&E-stained histology from the Medical University of Vienna (Mahbod et al., Scientific Data 2024), mirrored from Zenodo record 10518968 (v1.0) with every nucleus converted to a browsable polygon annotation. It is the first patch-level pathology dataset on the hub and the histology counterpart to the fluorescence nuclei sets (BBBC039, DSB2018). The release also marks "vague" areas where even experts could not draw a reliable nucleus outline; those are included here as a second class so you can see exactly where the ground truth is uncertain.

Contents

Images	665 patches of 512×512 px, 8-bit RGB PNG, from 31 organs (23 human, 8 mouse; 472 human and 193 mouse patches)
Annotations	32,988 polygons: 30,694 nucleus (from 30,698 labeled nuclei; see Conversion notes) and 2,294 vague-area regions
Classes	2 (nucleus, vague-area)
Dataset	nuclei-segmentation-nuinsseg-v1
Split	None in the release; the authors publish a 5-fold split (133 images per fold) in their GitHub repository
Imaging	Brightfield H&E, TissueFAXS slide scanner (Zeiss Axio Imager Z1, 40×/0.75 air objective, Baumer HXG40c colour camera); each patch is the central 512×512 crop of a 2048×2048 field of view

The organ and species of every patch are encoded in its file name (human_kidney_05.png, mouse_thymus_01.png) and stored as file metadata (species, organ). All tissues are healthy except human melanoma. Nuclei density varies from a handful per patch (umbilical cord, muscle) to several hundred (thymus, spleen, bone marrow), which is what makes this set a useful stress test for H&E nuclei models.

Using it

Browse any image to see the nucleus polygons and the purple vague-area regions. To download everything (free account required):

• ZIP: use the Download button on this project.
• Python SDK: pip install datatorch, then:

datatorch login
datatorch pull datatorch/nuclei-segmentation-nuinsseg

Annotations export in COCO format through the project's export schema.

Conversion notes

• Source of the polygons. The release ships two instance-label folders per organ. label masks is the raw rasterisation of the ImageJ ROIs (35,138 IDs, where one nucleus is often split across several IDs); label masks modify is the authors' corrected version with one 16-bit ID per nucleus. The corrected masks sum to exactly 30,698 nuclei across 665 images and match the paper's per-organ Table 2 for all 31 organs, so they are the ground truth used here. The two folders have identical foreground pixels.
• Method. Each instance ID was traced with OpenCV findContours (external contours, simple chain approximation). Every connected component with at least 3 points and non-zero area was kept, so 65 nuclei whose mask is split into separate pieces are stored as multi-part polygons. 4 of 30,698 nuclei (in human_lung_11, human_salivory_03, human_spleen_12, human_umbilical_cord_07) are one-pixel-wide shapes with zero contour area and could not be expressed as polygons; they were dropped, giving 30,694 nucleus polygons.
• Vague areas. vague areas/mask binary (8-bit, 0/255) was traced the same way: 2,334 connected regions, of which 40 zero-area specks were dropped, leaving 2,294 vague-area polygons on 664 images (mouse_muscle_tibia_12 has none). About 12% of nucleus pixels fall inside a vague area; the overlap is intentional and the two classes are independent.
• ImageJ ROIs. The raw freehand ROI files were decoded by hand to confirm the coordinate convention (x = column, y = row, all within the 512×512 frame) and all sampled ROIs land on a labeled nucleus. They were not used as the polygon source because three organs (human jejunum, human kidney, mouse femur) store ROIs under scanner field names rather than image names, 6 images have one more ROI than labeled nucleus, and the corrected masks are the authors' final word.
• Naming. Folder names follow the release, which differ slightly from the paper's table: human rectum is the paper's "Colon", human brain is "Cerebrum", human tonsile is "Palatine tonsil", human cardia/human pylorus are "Stomach (cardia)/(pylorus)", mouse fat (white and brown)_subscapula is "Fat (subscapularis)" and mouse muscle_tibia is "Muscle (tibialis anterior)". File stems are the authors' own.
• Project file count. The project lists 666 files: the 665 patches in the dataset plus the COCO JSON used for the import, which the DataTorch CLI keeps under imports/ for provenance. It is not part of the dataset.
• Images are byte-identical to the release. One patch (human_liver_31.png) is RGBA with a fully opaque alpha channel; the other 664 are RGB. Not included: the derived binary, border-eroded, distance-map and weighted-map PNGs, the .mat stacked masks and the overlay previews, all of which can be regenerated from the labels with the authors' code.

License and citation

CC BY 4.0. You may share and adapt the data, including commercially, with attribution. If you use this data, please cite:

Mahbod, A., Polak, C., Feldmann, K., Khan, R., Gelles, K., Dorffner, G., Woitek, R., Hatamikia, S. & Ellinger, I. NuInsSeg: A fully annotated dataset for nuclei instance segmentation in H&E-stained histological images. Scientific Data 11, 295 (2024). https://doi.org/10.1038/s41597-024-03117-2

Data: Mahbod, A. et al. NuInsSeg (v1.0). Zenodo. https://doi.org/10.5281/zenodo.10518968

Related

• Nuclei Segmentation (DSB2018) and Nuclei Segmentation (BBBC039), the fluorescence counterparts
• More microscopy datasets on DataTorch
• All public datasets