README.md

---
license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - medical-imaging
  - electron-microscopy
  - neuroscience
  - axon-segmentation
  - 3d-segmentation
  - connectomics
size_categories:
  - 1K<n<10K
---

# AxonEM Dataset

Large-scale 3D Axon Instance Segmentation of Brain Cortical Regions from serial section Electron Microscopy (sEM).

## Dataset Description

AxonEM contains high-resolution electron microscopy volumes of mouse and human brain cortex tissue for axon instance segmentation.

### Subsets

| Subset | Species | Volumes | Resolution | Original Size |
|--------|---------|---------|------------|---------------|
| Human | Homo sapiens | 9 | 30×8×8 nm | 1000×4096×4096 |
| Mouse | Mus musculus | 9 | 40×8×8 nm | 750×4096×4096 |

### Volume Information

Each training sub-volume has shape **(90, 1536, 1536)** voxels with padding:
- Padding: 20 slices in Z, 512 pixels in Y/X (on each side)
- Annotated region: (50, 512, 512) after removing padding

### File Structure

```
AxonEM/
├── EM30-H-train-9vol-pad-20-512-512/    # Human subset
│   ├── im_0-0-0_pad.h5                   # Image volume
│   ├── seg_0-0-0_pad.h5                  # Segmentation (instance labels)
│   └── ...
├── EM30-M-train-9vol-pad-20-512-512/    # Mouse subset
│   ├── im_0-0-0_pad.h5
│   ├── seg_0-0-0_pad.h5
│   ├── valid_mask.h5                     # Valid annotation mask
│   └── ...
└── README.md
```

### HDF5 Format

Each `.h5` file contains a single dataset with key `'main'`:
- **Image files** (`im_*.h5`): uint8 grayscale EM images, shape (90, 1536, 1536)
- **Segmentation files** (`seg_*.h5`): uint8 instance labels, shape (90, 1536, 1536)
  - 0 = background
  - 1-N = axon instance IDs

### Loading Example

```python
import h5py
import numpy as np

# Load a volume
with h5py.File('EM30-H-train-9vol-pad-20-512-512/im_0-0-0_pad.h5', 'r') as f:
    image = f['main'][:]  # (90, 1536, 1536) uint8

with h5py.File('EM30-H-train-9vol-pad-20-512-512/seg_0-0-0_pad.h5', 'r') as f:
    labels = f['main'][:]  # (90, 1536, 1536) uint8

# Convert to binary mask (axon vs background)
binary_mask = (labels > 0).astype(np.uint8)

# Remove padding to get annotated region
z_pad, y_pad, x_pad = 20, 512, 512
image_cropped = image[z_pad:-z_pad, y_pad:-y_pad, x_pad:-x_pad]  # (50, 512, 512)
```

### Using with EasyMedSeg

```python
from dataloader.axonem import AxonEMImageDataset, AxonEMVideoDataset

# Image mode (2D slices)
dataset = AxonEMImageDataset(
    hf_repo_id="Angelou0516/AxonEM",
    subset="human",  # or "mouse"
)

# Video mode (3D volumes)
dataset = AxonEMVideoDataset(
    hf_repo_id="Angelou0516/AxonEM",
    subset="human",
)
```

## Citation

```bibtex
@inproceedings{wei2021miccai,
  title={AxonEM Dataset: 3D Axon Instance Segmentation of Brain Cortical Regions},
  author={Wei, Donglai and Xu, Kisuk and Liao, Ran and Pfister, Hanspeter and 
          Haehn, Daniel and Bhanu, Shubham and Bhattacharyya, Chandrajit},
  booktitle={International Conference on Medical Image Computing and 
             Computer-Assisted Intervention (MICCAI)},
  year={2021}
}
```

## Links

- [Grand Challenge](https://axonem.grand-challenge.org/)
- [arXiv Paper](https://arxiv.org/abs/2107.05451)
- [PyTorch Connectomics](https://github.com/zudi-lin/pytorch_connectomics)

## License

This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).