eminorhan/openorganelle-2d

OpenOrganelle 2D

This dataset contains a large collection of 2D slices from the EM volumes on HHMI Janelia's OpenOrganelle data repository. The dataset contains a total of ~2.67 million x, y, and z slices obtained from 79 different 3D EM volumes on OpenOrganelle.

Notes

1. The following volumes on OpenOrganelle are missing from the current repository because they were too large to process and store here: jrc_fly-larva-1, jrc_fly-mb-z0419-20, jrc_mus-guard-hair-follicle, jrc_mus-liver-zon-1, jrc_mus-liver-zon-2, jrc_mus-meissner-corpuscle-2, jrc_mus-pacinian-corpuscle, jrc_zf-cardiac-1, jrc_dauer-larva

2. Again due to HF storage constraints, we only processed and stored every 20th slice along each axis (i.e. 20x subsampling of the full data). The full data take up over 100 TB on disk. This subsampled version only takes up ~6.35 TB of disk space.

3. We also divided large slices into smaller equal-sized pieces of no more than 2048 pixels along any given axis: e.g. a 8192x8192 slice would be broken up into 16 parts of size 2048x2048 pixels each and each part would be given a unique part id i_j (e.g. 0_0, 0_1, 0_2, 0_3, 1_0, 1_1,..., 3_3) identifying its "part coordinates" within the larger slice.

4. We then removed small slices where any edge was shorter than 64 pixels and applied per-slice normalization to the remaining slices.

5. More specifically, the slices were prepared with this preprocessing script and then pushed to the HF datasets Hub using this script. We used the highest resolution data (stored in s0) from all volumes.

6. The dataset rows are pre-shuffled to make the data shards roughly uniform in size.

Usage

Non-streaming mode: We recommend caching the dataset on local disk if you have enough disk space (~6.35 TB). You can then load the dataset as follows:

>>> ds = load_dataset("eminorhan/openorganelle-2d", split='train')

and inspect e.g. the first data row:

>>> print(ds[0])
>>> {
'image': <PIL.PngImagePlugin.PngImageFile image mode=L size=3535x3565 at 0xFFF93CFA52D0>,
'volume_name': 'jrc_mus-choroid-plexus-3',
'axis': 'y',
'slice': 3960,
'part_id': '3_1'
}

where:

• image contains the actual 2D slice encoded as a PIL.Image object.
• volume_name is an identifier string indicating the EM volume the slice comes from.
• axis indicates the axis along which the slice was taken (x, y, or z).
• slice is the slice index along the axis.
• part_id is an identifier string i_j indicating the part coordinates i and j of the slice if the slice was obtained by dividing a larger slice into smaller equal-sized pieces (see above). If the slice was not obtained by dividing a larger slice, part_id will be 0_0.

Streaming mode: Alternatively, if you don't have enough disk space or if you don't want to download the full dataset to your local disk, you can load it in streaming mode instead and then inspect e.g. the first data row as follows:

>>> ds = load_dataset("eminorhan/openorganelle-2d", split="train", streaming=True)
>>> fr = next(iter(ds))
>>> print(fr)

License: The data originally come from HHMI Janelia's OpenOrganelle data portal released under the CC-BY-4.0 license.

Citation: If you use these data, please cite the following paper:

@article{heinrich2021whole,
  title={Whole-cell organelle segmentation in volume electron microscopy},
  author={Heinrich, Larissa and Bennett, Davis and Ackerman, David and Park, Woohyun and Bogovic, John and Eckstein, Nils and Petruncio, Alyson and Clements, Jody and Pang, Song and Xu, C Shan and others},
  journal={Nature},
  volume={599},
  number={7883},
  pages={141--146},
  year={2021},
  publisher={Nature Publishing Group UK London}
}

Paper link