NucMM-Z Dataset

Overview

NucMM-Z (Neuronal Nuclei from Zebrafish) is a 3D electron microscopy (EM) dataset for nuclei instance segmentation from zebrafish brain tissue.

Property	Value
Modality	Electron Microscopy (EM)
Task	Nuclei instance segmentation
Anatomy	Zebrafish brain
Volume Size	64 × 64 × 64 voxels per patch
Train Volumes	27
Val Volumes	27
Total Size	~1.09 GB

Dataset Structure

NucMM-Z/
├── image.tif           # Full raw volume (~1 GB)
├── mask.h5             # Full annotation volume
├── README.txt          # Original readme
├── Image/
│   ├── train/          # 27 training patches (.h5)
│   └── val/            # 27 validation patches (.h5)
└── Label/
    ├── train/          # 27 training labels (.h5)
    └── val/            # 27 validation labels (.h5)

Label Format

• Instance Segmentation: Each nucleus has a unique integer ID
• Background: 0
• Typical density: 50-300 nuclei per 64×64×64 volume

Usage with EasyMedSeg

from dataloader import NucMMZImageDataset, NucMMZVideoDataset

# Image mode (2D slices) - Recommended
dataset = NucMMZImageDataset(split='train')
sample = dataset[0]  # Returns dict with 'image' and 'mask'

# Video mode (3D volumes as frame sequences)
dataset = NucMMZVideoDataset(split='train')
video = dataset[0]  # Returns dict with 'frames' and 'masks'

Benchmark Results (SAM2)

Mode	Model	Mean Dice	Mean IoU
Image	sam2_hiera_large	0.3438	0.2566
Video	sam2_video_hiera_large	0.0631	0.0425

Recommendation: Use image mode for this dataset.

Source

• Original: PyTorch Connectomics NucMM
• Paper: Wei et al., MICCAI 2020

License

MIT

Redistribution and commercial use are permitted under the terms below.

Source of the terms: https://huggingface.co/datasets/pytc/NucMM

Audit note (verbatim from the MedOtter dataset card):

2026-08-27 re-audit (closes the 2026-08-17 gap): The official NucMM challenge page (nucmm.grand-challenge.org, which the project page's [Dataset] link points at) links the public release to a Google Drive folder ('NucMM-Release', subfolders 'Mouse (NucMM-M)' and 'Zebrafish (NucMM-Z)', no license file at top level) and, for complete annotations, to a data request form; that form, run by the Harvard VCG group and headed by an image of both halves, states verbatim: 'The dataset is licensed under the MIT License. Please refer to the PyTorch Connectomics (https://github.com/zudi-lin/pytorch_connectomics) codebase for more details.' (that codebase's LICENSE is MIT, 'Copyright (c) 2019-2022 PyTorch Connectomics Contributors'). The authors' own HuggingFace release pytc/NucMM independently carries 'License: mit' and contains both halves (NucMM-M.zip 395 MB, NucMM-Z.zip 662 MB). Per half the verdicts now AGREE: NucMM-M (mouse visual-cortex micro-CT, collected in-house per the MICCAI 2021 paper) = MIT, and NucMM-Z (zebrafish EM, from co-author Petkova's Harvard data) = MIT — both are covered by the form's dataset-wide MIT statement and by the MIT-tagged pytc/NucMM release, which supersedes the 2026-08-17 audit (NucMM-M 'no data license / all rights reserved', NucMM-Z 'conditional') that had looked only at the license-silent challenge/Drive channel. MIT permits redistribution including commercial use with retention of the license notice, and the request form itself lists 'Commercial' as an accepted intended use. Caveat: the complete (test) annotations remain gated behind the request form; a public mirror should carry the public release (raw images + training/validation annotations), which is exactly what the MIT-tagged pytc/NucMM repo distributes.

Please cite:

Lin, Z., Wei, D., Petkova, M. D., Wu, Y., Ahmed, Z., K, K. S., Zou, S., Wendt, N., Boulanger-Weill, J., Wang, X., Dhanyasi, N., Arganda-Carreras, I., Engert, F., Lichtman, J., & Pfister, H. (2021). NucMM Dataset: 3D Neuronal Nuclei Instance Segmentation at Sub-Cubic Millimeter Scale. *MICCAI 2021*. https://doi.org/10.48550/arXiv.2107.05840