Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
cannot reshape array of size 31250000 into shape (20,1024,1024)
Error code:   UnexpectedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

BrainEM

BrainEM: A Large-Scale and Diverse Benchmark for EM Neuron Segmentation in Connectomics
NeurIPS 2026 Evaluations and Datasets Track.

Authors: Zhenghua Li, Lanyue Zhang, Zekang Yang, Kai Li, Yuwei Peng, Song-Hai Shi, Xiaolin Hu.
Contact: li-zh24@mails.tsinghua.edu.cn; xlhu@tsinghua.edu.cn.
Code and reproduction instructions: https://github.com/kwinderic/BrainEM

Overview and intended use

BrainEM brings together 16 EM datasets, including the newly acquired and manually annotated MC93 mouse dataset. The benchmark evaluates dense 3D neuron instance segmentation in two settings: Same-Source (CREMI training, FlyWire test) and Cross-Source (multi-source training, held-out FIB25, AC3, Wafer4 and MC93 targets). Follow the paper and code configurations for training/validation/test assignments; not every released file is a training file. Raw EM images and matching instance-label volumes are paired in the metadata. The data support research in connectomics, not clinical diagnosis.

Files and loading

HDF5 files (.h5) and multi-page TIFF files (.tif) contain volumes in (z, y, x) array order. Images and labels are separate files; integer label IDs identify instances within each volume, with 0 as background. Labels from different volumes do not share an instance-ID namespace. MC93 has shape (20, 3960, 3990). PNG files are previews, not replacement training/test volumes.

hf download kwinderic/BrainEM --repo-type dataset --local-dir datasets/EM

Use h5py to inspect and read the volume dataset within each HDF5 file, and tifffile.imread to read the complete TIFF stack. Preserve all sections; loading only the first TIFF page does not load a full volume. The code may use shorter local aliases (for example pyw93 for MC93, or flywire_crop_* for the coordinate-named FlyWire files); update local paths/configuration to match the files actually downloaded.

Croissant metadata

BrainEM.croissant.json describes the collection; subsets/ contains one Croissant JSON for each subset. The recordSet entries provide a machine-readable manifest of raw/label pairs with pinned download URLs, SHA-256 hashes, byte sizes and file formats. They do not decode HDF5/TIFF voxel tensors automatically. file-manifest.csv and volume-pairs.csv provide the same file inventory and pairing in tabular form. Previews are excluded from the segmentation-data manifest.

MC93 acquisition and annotation

MC93 was acquired by the authors from C57BL/6 mouse primary visual cortex (V1). Preparation used transcardial perfusion fixation and rOTO heavy-metal staining. ATUM sections were imaged on a SEMNova-1000 at 2 kV, with 4 nm XY sampling and 50 nm section spacing. Tiles of 12000 x 12000 pixels, with 12.5% overlap, formed a 23 x 11 grid. The released volume has array shape (z, y, x) = (20, 3960, 3990). Acquisition took place from September to December 2022 under IACUC protocol 22-SSH1.

A single annotator with PhD-level neuroscience training performed dense tracing and instance segmentation using TrakEM2 in Fiji/ImageJ, following an internal protocol. Annotation took approximately four weeks. A senior expert proofread approximately 5-10% of the volume for major errors. No independent double annotation, adjudication or formal inter-annotator agreement measurement was performed.

Limitations and responsible use

Source datasets differ in species, imaging conditions, resolution and annotation quality. The collection is not representative of all nervous systems, and reference segmentations can contain errors. Preserve source-level exclusions when evaluating generalization. MC93 represents one animal, one brain region and one imaging session, without pathological conditions. Its 20-section axial extent limits spatial context. Partial proofreading cannot exclude residual annotation errors.

H01 contains human-derived tissue; consult the source documentation for provenance and consent. This release does not include direct personal identifiers. No synthetic data are included. Improved segmentation can support connectomics, but errors may distort inferred connectivity and results should not be generalized beyond the tested settings.

Licenses and attribution

The repository carries a CC BY-NC 4.0 collection-level declaration. This does not relicense all underlying data. MC93 is released under CC BY 4.0. Reused datasets retain their source-specific terms and attribution requirements. The source links and inherited license statements are listed below and in the per-subset metadata. Observe the applicable source terms when downloading, reusing or redistributing data.

Subset Source Source license / terms
AC3-AC4 https://www.cell.com/cell/fulltext/S0092-8674(15)00824-7 Research use only; see Kasthuri et al. 2015 (Cell) and Open Connectome Project release terms.
H01-AxonEM https://h01-release.storage.googleapis.com/landing.html https://creativecommons.org/licenses/by/4.0/
pinkey-AxonEM https://elifesciences.org/articles/76120 https://creativecommons.org/licenses/by/4.0/
Basil https://www.microns-explorer.org/ https://creativecommons.org/licenses/by/4.0/
FAFB https://temca2data.org/ Research use only; see https://temca2data.org/ for FAFB release terms.
FIB25 https://elifesciences.org/articles/26975 https://creativecommons.org/licenses/by/4.0/
Flywire https://flywire.ai/ https://creativecommons.org/licenses/by-nc/4.0/
Harris https://www.nature.com/articles/sdata201546 https://creativecommons.org/licenses/by/4.0/
Hemi-brain-fib https://elifesciences.org/articles/57443 https://creativecommons.org/licenses/by/4.0/
MC93 https://huggingface.co/datasets/kwinderic/BrainEM https://creativecommons.org/licenses/by/4.0/
MiRA-ADWT https://link.springer.com/chapter/10.1007/978-3-031-72111-3_55 Research use only; see SegNeuron release terms.
MiRA-PIB https://link.springer.com/chapter/10.1007/978-3-031-72111-3_55 Research use only; see SegNeuron release terms.
MiRA-SCN https://link.springer.com/chapter/10.1007/978-3-031-72111-3_55 Research use only; see SegNeuron release terms.
MiRA-ZF https://link.springer.com/chapter/10.1007/978-3-031-72111-3_55 Research use only; see SegNeuron release terms.
cremi https://cremi.org/ Research use only; see https://cremi.org/data/ for challenge release terms.
wafer4 https://openaccess.thecvf.com/content/CVPR2024/html/Liu_Cross-Dimension_Affinity_Distillation_for_3D_EM_Neuron_Segmentation_CVPR_2024_paper.html Research use only; see CAD release terms.

Citation

@inproceedings{li2026brainem,
  title = {BrainEM: A Large-Scale and Diverse Benchmark for EM Neuron Segmentation in Connectomics},
  author = {Li, Zhenghua and Zhang, Lanyue and Yang, Zekang and Li, Kai and Peng, Yuwei and Shi, Song-Hai and Hu, Xiaolin},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026}
}

Please also acknowledge the original data sources when using their subsets. Data corrections are tracked through repository revisions; report issues through the linked code repository.

Downloads last month
159