NIS3D — A Completely Annotated Benchmark for Dense 3D Nuclei Image Segmentation

NIS3D is a dense 3D nuclei segmentation benchmark for fluorescence microscopy of embryonic tissue. It provides 6 large-volume 3D images (Zebrafish, Drosophila, Mus Musculus) with 22,000+ manually annotated cells vetted by three independent annotators, accompanied by per-cell confidence scores derived from inter-annotator agreement.

Published at NeurIPS 2023 (Datasets and Benchmarks Track).

Dataset Summary

Field	Details
Modality	3D fluorescence microscopy (volumetric TIFF stacks)
Target	Cell nuclei (instance segmentation)
Species	Zebrafish, Drosophila, Mus Musculus (mouse)
Samples	6 (Zebrafish_1, Zebrafish_2, Drosophila_1, Drosophila_2, MusMusculus_1, MusMusculus_2)
Annotated cells	22,000+
File format	Multi-page TIFF (.tif), opens with Fiji/ImageJ or tifffile
License	CC-BY-4.0
Source	Zenodo: https://zenodo.org/records/11456029 (DOI 10.5281/zenodo.11456029)
GitHub	https://github.com/yu-lab-vt/NIS3D

Per-Sample Files

Each sample folder under NIS3D/ contains:

File	Role
data.tif	Raw 3D fluorescence volume
GroundTruth.tif (or gt.tif for MusMusculus_2)	Recommended ground-truth instance label map (different integers = different cells, 0 = background)
ConfidenceScore.tif (or scoreOfConfidence.tif for MusMusculus_2)	Per-cell reliability map: 0 = background, 1 = "undefined mask" (boundaries indeterminate, ignored at evaluation), 2 = 1/3 annotator agreement, 3 = 2/3, 4 = 3/3 (full consensus)
visulize.tif	Composite visualization (raw + GT + confidence) — display only
Info.txt	Acquisition details (microscope, resolution, pre-processing, etc.)

The single GroundTruth.tif is the consensus across three annotators; per-annotator disagreement is encoded into ConfidenceScore.tif rather than rival GT files. At evaluation, predictions falling on confidence-1 voxels should be ignored (neither TP nor FP) because boundaries there are too weak for any annotator to delineate.

Splits

The Zenodo archive ships two suggestive split schemes under suggestive splitting/, in addition to the canonical full-size 6 samples in NIS3D/:

Split scheme	Description
cross-image/	Leave-one-per-species: train on *_2, test on *_1 (full-size volumes)
in-image/	Each of the 6 samples is cropped into disjoint train/test halves

Both schemes are advisory — researchers may use the canonical NIS3D/ folder and define their own splits.

Per-Sample Stats

Sample	data.tif size	GroundTruth.tif size	Cells (approx.)
Zebrafish_1	1.25 GB	68 MB	(high density)
Zebrafish_2	45 MB	6.5 MB
Drosophila_1	382 MB	28 MB
Drosophila_2	386 MB	17 MB
MusMusculus_1	554 MB	45 MB
MusMusculus_2	273 MB	47 MB

Per-sample acquisition details (microscope, resolution, pre-processing) are in each Info.txt.

Citation

@inproceedings{zheng2023nis3d,
  title     = {NIS3D: A Completely Annotated Benchmark for Dense 3D Nuclei Image Segmentation},
  author    = {Zheng, Wei and Peng, James Cheng and Hou, Zeyuan and Lyu, Boyu and Wang, Mengfan and Mi, Xuelong and Qiao, Shuoxuan and Wan, Yinan and Yu, Guoqiang},
  booktitle = {Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2023},
  url       = {https://openreview.net/forum?id=Icxwnu9hcO}
}

@dataset{zheng2024nis3dzenodo,
  title     = {NIS3D},
  author    = {Zheng, Wei and Peng, James Cheng and Hou, Zeyuan and Lyu, Boyu and Wang, Mengfan and Mi, Xuelong and Qiao, Shuoxuan and Wan, Yinan and Yu, Guoqiang},
  year      = {2024},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.11456029},
  url       = {https://doi.org/10.5281/zenodo.11456029}
}