CREMI - Circuit Reconstruction from Electron Microscopy Images

Dataset Description

CREMI is a benchmark dataset for evaluating algorithms for automatic reconstruction of neurons and neuronal connectivity from serial section electron microscopy data.

Modality

Serial section Transmission Electron Microscopy (ssTEM)

Anatomy

Adult Drosophila melanogaster brain

Segmentation Targets

Neuron segmentation (volumes/labels/neuron_ids)
Synaptic cleft segmentation (volumes/labels/clefts)

Volume Size

125 slices × 1250 × 1250 pixels per sample
Resolution: 4×4×40 nm (x,y,z)

Dataset Structure

CREMI/
└── train/
    ├── sample_A.hdf    # 167 MB
    ├── sample_B.hdf    # 160 MB
    └── sample_C.hdf    # 165 MB

Note: Test files (A+, B+, C+) are not included as they have no labels.

HDF5 Structure

Each training file contains:

volumes/raw: (125, 1250, 1250) uint8 - grayscale EM images
volumes/labels/neuron_ids: (125, 1250, 1250) uint64 - neuron instance segmentation
volumes/labels/clefts: (125, 1250, 1250) uint64 - synaptic cleft segmentation
annotations/ - synaptic partner annotations

Usage

import h5py
from huggingface_hub import hf_hub_download

# Download a training volume
path = hf_hub_download(
    repo_id="Angelou0516/CREMI",
    filename="train/sample_A.hdf",
    repo_type="dataset"
)

# Load data
with h5py.File(path, 'r') as f:
    raw = f['volumes/raw'][:]  # (125, 1250, 1250)
    neuron_ids = f['volumes/labels/neuron_ids'][:]  # (125, 1250, 1250)
    clefts = f['volumes/labels/clefts'][:]  # (125, 1250, 1250)

Citation

@misc{cremi2016,
    title={CREMI: MICCAI Challenge on Circuit Reconstruction from Electron Microscopy Images},
    author={Funke, Jan and Saalfeld, Stephan and Bock, Davi and Turaga, Srini and Perlman, Eric},
    year={2016},
    url={https://cremi.org/}
}

License

Please refer to the original dataset website: https://cremi.org/

Datasets:

Angelou0516
/

CREMI