---
license: mit
---

# Dataset Card for DenSpine

## Volumetric Files
The dataset is comprised of dendrites from 3 brain samples: `seg_den` (also known as `M50`), `mouse` (`M10`), and `human` (`H10`).

Every species has 3 volumetric `.h5` files:
- `{species}_raw.h5`: instance segmentation of entire dendrites in volume (labelled `1-50` or `1-10`), where trunks and spines share the same label
- `{species}_spine.h5`: "binary" segmentation, where trunks are labelled `0` and spines are labelled their `raw` dendrite label
- `{species}_seg.h5`: spine instance segmentation (labelled `51-...` or `11-...`), where every spine in the volume is labelled uniquely

## Point Cloud Files
In addition, we provide preprocessed point clouds sampled along a dendrite's centerline skeletons for ease of use in evaluating point-cloud based methods.
```python
data=np.load(f"{species}_1000000_10000/{idx}.npz", allow_pickle=True)
trunk_id, pc, trunk_pc, label = data["trunk_id"], data["pc"], data["trunk_pc"], data["label"]
```
- `trunk_id` is an integer which corresponds to the dendrite's `raw` label
- `pc` is a shape `[1000000,3]` isotropic point cloud
- `trunk_pc` is a shape `[skeleton_length, 3]` (ordered) array, which represents the centerline of the trunk of `pc`
- `label` is a shape `[1000000]` array with values corresponding to the `seg` labels of each point in the point cloud 

We provide a comprehensive example of how to instantiate a PyTorch dataloader using our dataset in `dataloader.py` (potentially using the FFD transform with `frenet=True`).

## Training splits for `seg_den`
The folds used for training/evaluating the `seg_den` dataset, based on `raw` labels are defined as follows:
```python
seg_den_folds = [
        [3, 5, 11, 12, 23, 28, 29, 32, 39, 42],
        [8, 15, 19, 27, 30, 34, 35, 36, 46, 49],
        [9, 14, 16, 17, 21, 26, 31, 33, 43, 44],
        [2, 6, 7, 13, 18, 24, 25, 38, 41, 50],
        [1, 4, 10, 20, 22, 37, 40, 45, 47, 48],
]
```