Hugging Face's logo

Datasets:
MedOtter
/
MitoEM

Tasks:
Image Segmentation
Modalities:
Image
Text
Formats:
parquet
Size:
< 1K
Tags:
medical
biomedical
electron-microscopy
ssem
connectomics
mitochondria
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
MitoEM / README.md
Parth1503's picture
Parth1503
Upload dataset
6387a07 verified
preview code
|
Raw
History Blame Contribute Delete
10.4 kB
metadata 
license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - medical
  - biomedical
  - electron-microscopy
  - ssem
  - connectomics
  - mitochondria
  - instance-segmentation
  - 3d-segmentation
  - brain
  - cortex
  - miccai
  - mitoem
pretty_name: MitoEM - 3D mitochondria instance segmentation in ssSEM (labeled half)
size_categories:
  - 1K<n<10K
dataset_info:
  config_name: preview
  features:
    - name: sample_id
      dtype: string
    - name: subset
      dtype: string
    - name: organism
      dtype: string
    - name: split
      dtype: string
    - name: z
      dtype: int32
    - name: n_instances
      dtype: int32
    - name: fg_fraction
      dtype: float32
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
  splits:
    - name: train
      num_bytes: 268684200
      num_examples: 80
    - name: val
      num_bytes: 67196014
      num_examples: 20
  download_size: 335902375
  dataset_size: 335880214
configs:
  - config_name: preview
    data_files:
      - split: train
        path: preview/train-*
      - split: val
        path: preview/val-*

MitoEM (publicly-labeled half)

MitoEM — A Large-scale 3D Mitochondria Instance Segmentation Dataset from Electron Microscopy (Wei et al., MICCAI 2020). Two (30 µm)³ tissue blocks imaged by serial-section multi-beam SEM (ssSEM) at 8 × 8 × 30 nm, one from rat and one from human cortex, densely annotated for mitochondria instances.

This mirror = the 500 publicly-labeled slices per subset (z 0–499). Each source volume is 1000 × 4096 × 4096 voxels, but only the first half carries public ground truth. Slices z 500–999 form the challenge test half and their labels are withheld by the organizers (server-side evaluation at mitoem.grand-challenge.org). They are omitted here because they cannot be evaluated offline and would double the download for no benchmark value.

Dataset Details

Field Value
Modality Serial-section multi-beam SEM (ssSEM)
Resolution 8 × 8 × 30 nm (x, y, z)
Task Mitochondria instance segmentation (binary semantic = label > 0)
Subsets MitoEM-R (rat), MitoEM-H (human)
Slices 500 per subset — train z 0–399, val z 400–499
Slice size 4096 × 4096
Images 8-bit grayscale PNG
Masks uint16 TIFF (deflate), 0 = background, non-zero = instance ID
Instances ≈10.5 k (human) / ≈5.4 k (rat) in the labeled range; ~350–390 per slice
License CC BY 4.0 (layered — see below; upstream tags the annotations MIT)
Source pytc/EM30 (images) + pytc/MitoEM (labels) — the authors' own org

License — CC BY 4.0, not MIT

Both upstream repos declare MIT, but that covers the challenge's annotations, not the underlying imagery. MitoEM-H is the EM30-H human cortex volume, which is understood to derive from the H01 human cortex release (Shapson-Coe et al.) under CC BY 4.0 — the attribution applied by MedOtter/AxonEM, which is served from the same pytc/EM30 archive. This mirror is therefore tagged with the most restrictive governing layer, cc-by-4.0, for consistency.

Provenance of that call: the H01 lineage is an upstream-attribution finding carried over from AxonEM, not something re-derived here; the MICCAI 2020 paper itself says only "human frontal lobe, Layer II". Tagging CC BY 4.0 is safe in either case — if the imagery were governed solely by the upstream MIT tag, CC BY 4.0 merely imposes a stricter attribution duty than required.

Both licences are permissive and redistribution-friendly; only the attribution obligation differs. MitoEM-R's rat volume has no separately adjudicated upstream, so CC BY 4.0 is applied uniformly as the conservative choice.

Attribution required: cite the two MitoEM papers below, and credit the H01 human cortex release for MitoEM-H.

Repository structure 

MitoEM-H/
  im/im0000.png … im0499.png              # 4096×4096 uint8
  mito-train-v2/seg0000.tif … seg0399.tif # 4096×4096 uint16 instance IDs
  mito-val-v2/seg0400.tif … seg0499.tif
MitoEM-R/
  im/…  mito-train-v2/…  mito-val-v2/…
train.jsonl                               # canonical per-slice index
val.jsonl
dataset_metadata.json

Image im{N}.png pairs with mask seg{N}.tif for the same N — both subsets are already spatially aligned at 4096 × 4096, so no cropping or offsetting is needed on read.

train.jsonl record schema 

{"sample_id": "MitoEM-H_z0000", "subset": "MitoEM-H", "organism": "human",
 "split": "train", "z": 0,
 "image": "MitoEM-H/im/im0000.png",
 "mask": "MitoEM-H/mito-train-v2/seg0000.tif",
 "shape": [4096, 4096], "n_instances": 386, "fg_fraction": 0.0508}

⚠️ Geometry note — the upstream padding convention is not what it looks like

The upstream human images ship as EM30-H-im-pad.zip: 1040 slices of 5120 × 5120, while its labels are 500 slices of 4096 × 4096. The convention implied by AxonEM's directory name (pad-20-512-512) suggests the label frame sits at [z+20, 512:4608, 512:4608]. It does not. A 2D offset search scoring image/mask intensity contrast against the known-aligned rat subset peaks sharply at (0, 0):

candidate contrast
offset (0, 0) +38.80 ← peak (decays to +25.7 by just 16 px)
offset (512, 512) −1.32
shifted negative control −1.65

The z offset is 0 as well. The padding is appended at the far edges (1000→1040 in z, 4096→5120 in y/x), so the padded volume's origin coincides with the nominal origin — which reconciles with AxonEM's crop origins (y/x ∈ {0, 1792, 3584} + 1536 = 5120; z 950 + 90 = 1040). pad-20-512-512 describes each AxonEM crop's internal margin, not a base-volume offset.

This mirror stores both subsets already cropped and aligned, so readers never encounter the issue. It is documented only so the mirror can be reconciled against upstream.

Two further upstream quirks handled during preparation, noted for anyone going back to the source: EM30-R-im.zip ships 1001 __MACOSX AppleDouble entries that a naive glob("*.png") double-counts to 2000 slices; and the label TIFFs are internally uncompressed (33.5 MB each) with the zip's deflate doing all the work, so extracting them verbatim yields 33.5 GB of masks. They are re-encoded here as deflate TIFF (losslessly identical, ~120 KB each).

Ground truth

v2 instance labels — the corrected release used by the IEEE TMI 2023 challenge report. 0 = background; every non-zero value is a mitochondrion instance ID. There are no competing rater or auto-generated tiers, so no gold-standard tier selection is required.

  • IDs are volume-global and sparse — they are not contiguous within a slice (e.g. a human slice with 386 instances has IDs up to 19245). Do not assume max(id) == n_instances.
  • For binary semantic mitochondria segmentation use mask > 0.
  • Annotated instances have a minimum size of 2000 voxels.
  • "Mitochondria-on-a-string" (MOAS) and the small/medium/large size bins from the paper are evaluation strata for error analysis, not stored label classes.

⚠️ Relationship to AxonEM (benchmark non-independence)

MitoEM-H and AxonEM-Human are the same image volume (EM30-H) — both challenges serve the human images from the identical upstream archive. AxonEM is already mirrored at MedOtter/AxonEM, where the human subset is 9 crops of that volume (EM30-H-train-9vol-pad-20-512-512, files im_{z}-{y}-{x}_pad.h5).

Using AxonEM's crop origins and the (0, 0) alignment established above, 5 of its 9 human crops intersect MitoEM's public labeled range: the four at z=0 and the one at z=475; the four at z=950 fall in MitoEM's withheld test half.

The annotation targets differ (axons vs mitochondria), so this is not label leakage — but the two are not statistically independent benchmarks, and a model tuned on one has seen the other's pixels. There is no cross-reference ID column; the join key is the EM30-H voxel origin embedded in AxonEM's filenames.

MitoEM-R is unaffected. AxonEM's other volume is EM30-M (mouse, 40 × 8 × 8 nm, 750 valid slices) — a different acquisition.

No overlap with the other EM/microscopy sets in this suite (NucMM, UroCell, CREMI, 3D-Platelet-EM, SELMA3D), nor with Lucchi/EPFL or Kasthuri++ — the latter is mouse somatosensory cortex at 12 × 12 × 30 nm, a different species, region and resolution from MitoEM-R's rat V1.

Source & citation

The organizers ask that both papers be cited:

@inproceedings{wei2020mitoem,
  author    = {Wei, Donglai and Lin, Zudi and Franco-Barranco, Daniel and
               Wendt, Nils and Liu, Xingyu and Yin, Wenjie and Huang, Xin and
               Gupta, Aarush and Jang, Won-Dong and Wang, Xueying and
               Arganda-Carreras, Ignacio and Lichtman, Jeff W. and Pfister, Hanspeter},
  title     = {{MitoEM} Dataset: Large-Scale 3D Mitochondria Instance
               Segmentation from {EM} Images},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  series    = {LNCS}, volume = {12265}, pages = {66--76}, year = {2020},
  doi       = {10.1007/978-3-030-59722-1_7}
}

@article{shapsoncoe2024h01,
  author  = {Shapson-Coe, Alexander and Januszewski, Micha{\l} and Berger, Daniel R. and others},
  title   = {A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution},
  journal = {Science},
  volume  = {384}, number = {6696}, pages = {eadk4858}, year = {2024},
  doi     = {10.1126/science.adk4858}
}

@article{francobarranco2023mitoem,
  author  = {Franco-Barranco, Daniel and Lin, Zudi and Jang, Won-Dong and others},
  title   = {Current Progress and Challenges in Large-Scale 3D Mitochondria
             Instance Segmentation},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {42}, number = {12}, pages = {3956--3971}, year = {2023},
  doi     = {10.1109/TMI.2023.3320497}
}