SELMA3D (annotated patches)

SELMA3D — Self-supervised Learning for 3D light-sheet Microscopy image segmentation (MICCAI 2024 challenge; Chen et al., arXiv:2501.03880). 3D light-sheet fluorescence microscopy (LSFM) of cleared brain tissue (DISCO-family clearing), covering four structures with different morphologies: blood vessels, c-Fos⁺ active neurons, cell nuclei, and amyloid-β plaques.

This mirror = the public ANNOTATED-PATCHES subset only (89 patches with expert ground-truth masks). Two parts of the full challenge are not here:

1. The 3.89 TB unannotated self-supervised pretraining corpus (BioStudies S-BIAD1197) — no masks, not included (irrelevant to a supervised segmentation benchmark, and far too large).
2. The 124 held-out test patches (including the unseen microglia structure used for generalization) — withheld by the organizers on Grand Challenge; not publicly downloadable.

The 89 patches reconcile exactly with the paper's Table 2 training rows (24 vessel + 19 c-Fos + 12 nucleus + 34 plaque). The paper abstract's "315" and the MedSAM list's "92" do not match the verifiable public count.

Dataset Details

Field	Value
Modality	3D light-sheet fluorescence microscopy (LSFM), cleared tissue
Body part	Brain (mouse + human donor)
Task	3D semantic segmentation (binary, per structure)
Patches (samples)	89 (1 mask each)
Format	NIfTI .nii.gz (16-bit signed / float, LPS+); nnU-Net file naming
Split	train only (test set held out by organizers)
License	CC BY-NC 4.0 (see note)
Source	EBI BioStudies S-BIAD1196, DOI 10.6019/S-BIAD1196 (official)
Challenge	https://selma3d.grand-challenge.org/ (MICCAI 2024)

License note (a real conflict — read before commercial use)

The BioStudies deposit metadata legally attaches CC BY 4.0 to these files, but the Grand Challenge data page states the data is CC BY-NC. Both allow redistribution with attribution; they differ only on commercial use. This mirror is tagged with the more restrictive CC BY-NC 4.0 out of caution. Either way, use must cite the SELMA3D challenge and the four source-dataset papers below.

Structures / subsets

Folder	Structure	Specimen	Patches	Image channels	Source dataset
VessAP_vessel	Blood vessels	mouse	24	2 (_0000 WGA microvessels, _0001 EB major vessels)	VesSAP (Todorov 2020)
cFos-Active_Neurons	c-Fos⁺ active neurons	mouse	19	1	Kaltenecker 2024
shannel_cells	Cell nuclei	human	12	1	SHANEL (Zhao 2020)
AD_plaques	Amyloid-β plaques	mouse	34	1	DISCO-MS (Bhatia 2022)

Ground truth

One expert-consensus mask per patch, produced by a hierarchical VR-based 3D annotation process: initial semantic segmentation by 4 LSFM-expert annotators → verified/refined by a senior expert → approved by 2 organizing-team leads. There are no competing rater/auto tiers — this single mask is the gold standard.

Loader notes (data is stored verbatim — these are read-side caveats)

• Vessel = 2 channels. Stack patchvolume_NNN_0000.nii.gz (WGA microvessels) and patchvolume_NNN_0001.nii.gz (EB major vessels); the other 3 structures are single-channel.
• Binarize masks with mask > 0, never mask == 1. The vessel GT encoding is heterogeneous across patches: 19 use {0,1}, 4 use {0,255}, 1 uses {0,95,255}. mask > 0 yields the correct binary vessel foreground for all; the other 3 structures are already clean {0,1}.
• Squeeze + cast raw. Some vessel raw volumes are 4D (X,Y,Z,1) and dtypes vary (int16 / float32 / big-endian int16); squeeze the trailing singleton and cast to float before intensity normalization. c-Fos/nucleus/plaque raw are float64 3D.
• Patch spatial sizes differ by structure: vessel 500×500×50, c-Fos 100³, nucleus 200³, plaque 300³.

Relationship to other datasets (lineage / leakage)

SELMA3D is assembled from the Ertürk lab's prior published LSFM datasets, which are also its mandatory citations: vessels ← VesSAP (Todorov 2020), nuclei ← SHANEL (Zhao 2020), plaques ← DISCO-MS (Bhatia 2022), c-Fos ← Kaltenecker 2024. The subset folder names preserve this provenance. There is no per-sample cross-reference ID; overlap risk exists only if you separately benchmark VesSAP/SHANEL/DISCO-MS. No relation to CT/MRI suites or to other microscopy/EM sets (NucMM, UroCell, CREMI, AxonEM, LungVis10).

Repository structure

SELMA3D_training_annotated/
  VessAP_vessel/{raw,gt}/        # raw: patchvolume_NNN_{0000,0001}.nii.gz ; gt: patchvolume_NNN.nii.gz
  cFos-Active_Neurons/{raw,gt}/  # raw: patchvolume_NNN_0000.nii.gz       ; gt: patchvolume_NNN.nii.gz
  shannel_cells/{raw,gt}/
  AD_plaques/{raw,gt}/
train.jsonl                      # canonical per-sample index (image channels + mask + metadata)
dataset_metadata.json           # dataset-level provenance & structure description

train.jsonl record schema

{"patch_id": "vessel_000", "subset": "VessAP_vessel", "structure": "vessel",
 "specimen": "mouse", "split": "train",
 "image": ["SELMA3D_training_annotated/VessAP_vessel/raw/patchvolume_000_0000.nii.gz",
           "SELMA3D_training_annotated/VessAP_vessel/raw/patchvolume_000_0001.nii.gz"],
 "mask": "SELMA3D_training_annotated/VessAP_vessel/gt/patchvolume_000.nii.gz",
 "num_channels": 2, "channel_names": ["WGA_microvessels", "EB_major_vessels"],
 "shape": [500, 500, 50], "gt_label_values": [0, 1]}

Source & citation

• BioStudies (annotated patches): https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1196 — DOI 10.6019/S-BIAD1196
• BioStudies (unannotated pretraining, NOT mirrored here): S-BIAD1197 — DOI 10.6019/S-BIAD1197
• Challenge: https://selma3d.grand-challenge.org/ (MICCAI 2024); proposal Zenodo 10.5281/zenodo.10991463

@article{chen2025selma3d,
  author  = {Chen, Ying and Al-Maskari, Rami and Horvath, Izabela and Ali, Mayar and
             H{"o}her, Luciano and Yang, Kaiyuan and Lin, Zengming and Zhai, Zhiwei and
             Shen, Mengzhe and Xun, Daniel and Wang, Yan and Xu, Tingying and
             Goubran, Maged and Wu, Yu and Mori, Kensaku and Paetzold, Johannes C. and
             Ert{"u}rk, Ali},
  title   = {{SELMA3D} challenge: Self-supervised learning for 3D light-sheet
             microscopy image segmentation},
  journal = {arXiv preprint arXiv:2501.03880},
  year    = {2025},
  doi     = {10.48550/arXiv.2501.03880}
}

@article{todorov2020vessap,
  author  = {Todorov, Mihail Ivilinov and Paetzold, Johannes C. and Schoppe, Oliver and others},
  title   = {Machine learning analysis of whole mouse brain vasculature},
  journal = {Nature Methods}, volume = {17}, pages = {442--449}, year = {2020},
  doi     = {10.1038/s41592-020-0792-1}
}

@article{zhao2020shanel,
  author  = {Zhao, Shan and Todorov, Mihail Ivilinov and Cai, Ruiyao and others},
  title   = {Cellular and Molecular Probing of Intact Human Organs},
  journal = {Cell}, volume = {180}, number = {4}, pages = {796--812}, year = {2020},
  doi     = {10.1016/j.cell.2020.01.030}
}

@article{bhatia2022discoms,
  author  = {Bhatia, Harsharan Singh and Brunner, Andreas and Rong, Zhouyi and others},
  title   = {Spatial proteomics in three-dimensional intact specimens},
  journal = {Cell}, volume = {185}, number = {26}, pages = {5040--5058}, year = {2022},
  doi     = {10.1016/j.cell.2022.11.021}
}

@article{kaltenecker2024cfos,
  author  = {Kaltenecker, Doris and Al-Maskari, Rami and Negwer, Moritz and others},
  title   = {Virtual reality-empowered deep-learning analysis of brain cells},
  journal = {Nature Methods}, volume = {21}, pages = {1306--1315}, year = {2024},
  doi     = {10.1038/s41592-024-02245-2}
}