README.md

---
license: cc-by-4.0
task_categories:
  - image-segmentation
  - image-to-text
tags:
  - neuroscience
  - neuron-reconstruction
  - neuron-tracing
  - brain-wide-reconstruction
  - light-microscopy
  - morphology
  - swc
size_categories:
  - 100G<n<1T
---
CORAL is the dataset for **CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy**.

The benchmark contains data for two neuron reconstruction settings:

- **Block-level reconstruction**: reconstruct local neuron morphology from small 3D image blocks.
- **Brain-wide reconstruction**: reconstruct complete neurons across whole-mouse-brain 2D light microscopy slices.


## Download

Install the Hugging Face Hub client:

```bash
pip install -U huggingface_hub
```

Download the complete dataset:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="yangzekang2000/CORAL",
    repo_type="dataset",
)
print(local_dir)
```

Download only the block-level reconstruction data:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="yangzekang2000/CORAL",
    repo_type="dataset",
    allow_patterns=[
        "cubes1937/**",
    ],
)
```

Download only the brain-wide reconstruction data:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="yangzekang2000/CORAL",
    repo_type="dataset",
    allow_patterns=[
        "slices/**",
        "swcs/**",
    ],
)
```

Download selected metadata files:

```python
from huggingface_hub import hf_hub_download

annos_path = hf_hub_download(
    repo_id="yangzekang2000/CORAL",
    repo_type="dataset",
    filename="cubes1937/annos.json",
)

hemisphere_path = hf_hub_download(
    repo_id="yangzekang2000/CORAL",
    repo_type="dataset",
    filename="swcs/brain_hemisphere.json",
)
```

## Repository Structure

```text
CORAL/
├── cubes1937/
│   ├── cubes/                 # 1937 raw 3D image blocks (.tif)
│   ├── swcs/                  # 1937 block-level SWC
│   ├── mask-r1/               # 1937 block-level mask
│   ├── annos.json             # Per-block metadata
│   └── C2-cubes1937_tvt/      # Train/validation/test split
├── slices/                    # Whole-brain 2D slice images (.tif)
├── swcs/
│   ├── single-neurons-step2/               # Single-neuron SWC annotations
│   ├── single-neurons-step2-somatree100/   # The initial tree near the neuron soma.
│   ├── all-neurons-step2.swc               # Combined SWC annotation for all 32 neurons
│   └── brain_hemisphere.json               # Hemisphere plane and neuron hemisphere labels
```

## Block-level Reconstruction: `cubes1937`

`cubes1937` provides the block-level reconstruction dataset. It contains **1937** local 3D image blocks, paired with SWC morphology annotations and mask annotations.

Each block is named by its spatial location in the full brain volume:

```text
cube300_x{X}_y{Y}_z{Z}.tif
cube300_x{X}_y{Y}_z{Z}.swc
cube300_x{X}_y{Y}_z{Z}_mask.tif
```

Main contents:

- `cubes1937/cubes/`: raw 3D image blocks in TIFF format.
- `cubes1937/swcs/`: block-level neuron morphology annotations in SWC format.
- `cubes1937/mask-r1/`: block-level mask annotations in TIFF format.
- `cubes1937/annos.json`: detailed metadata for the 1937 blocks, including coordinates, SWC filename, intensity statistics, neuron IDs, neuron length, node/edge counts, fiber and branch point counts, hemisphere label, and density.
- `cubes1937/C2-cubes1937_tvt/`: train/validation/test split files:
  - `train.txt`
  - `val.txt`
  - `test.txt`

Split statistics:

| Split | Number of samples |
|---|---:|
| Train | 767 |
| Validation | 140 |
| Test | 1022 |
| unused | 8 |
| Total | 1937 |

Each line in a split file contains comma-separated relative paths:

```text
cubes/<name>.tif,swcs/<name>.swc,mask-r1/<name>_mask.tif
```

Example loading code:

```python
import json
from pathlib import Path

root = Path("path/to/CORAL/cubes1937")

with open(root / "annos.json", "r") as f:
    annos = json.load(f)

with open(root / "C2-cubes1937_tvt" / "train.txt", "r") as f:
    train_samples = [line.strip().split(",") for line in f if line.strip()]

cube_rel, swc_rel, mask_rel = train_samples[0]
cube_path = root / cube_rel
swc_path = root / swc_rel
mask_path = root / mask_rel
```

## Brain-wide Reconstruction: `slices` and `swcs`

The brain-wide reconstruction data includes whole-mouse-brain 2D slice images and neuron-level SWC annotations.

Main contents:

- `slices/slices/`: 2D slice images of the mouse whole brain in TIFF format.
- `swcs/single-neurons-step2/`: SWC annotations for each individual neuron.
- `swcs/all-neurons-step2.swc`: a merged SWC annotation containing all **32** neurons.
- `swcs/brain_hemisphere.json`: an arbitrary 3D plane used to divide the left and right hemispheres, together with hemisphere annotations for the 32 neurons.

`brain_hemisphere.json` contains:

- `plane`: plane type, azimuth/elevation, normal vector, offset, and a point on the plane.
- `neurons`: per-neuron soma location, signed distance to the plane, left/right neurite length, total length, and hemisphere label.

Hemisphere statistics:

| Neuron hemisphere category | Number of neurons |
|---|---:|
| Entirely in the left hemisphere | 12 |
| Entirely in the right hemisphere | 14 |
| Crossing both hemispheres | 6 |
| Total | 32 |

Example loading code:

```python
import json
from pathlib import Path

root = Path("path/to/CORAL")

slice_dir = root / "slices" / "slices"
single_neuron_dir = root / "swcs" / "single-neurons-step2"
all_neurons_swc = root / "swcs" / "all-neurons-step2.swc"

slice_paths = sorted(slice_dir.glob("*.tif"))
single_neuron_swcs = sorted(single_neuron_dir.glob("*.swc"))

with open(root / "swcs" / "brain_hemisphere.json", "r") as f:
    hemisphere_info = json.load(f)
```

<!-- ## Citation

If you use this dataset, please cite:

```bibtex
@misc{coral_neuron_reconstruction,
  title  = {CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy},
  author = {CORAL Contributors},
  year   = {2026}
}
``` -->

<!-- ## License

This dataset is released under the CC BY 4.0 license. -->