README.md

---
license: cc-by-nc-nd-4.0
task_categories:
- image-segmentation
modality:
- MRI
language: []
tags:
- medical-imaging
- neonatal-brain
- hypoxic-ischemic-encephalopathy
- HIE
- lesion-segmentation
- diffusion-MRI
- ADC
- 3D
pretty_name: BONBID-HIE
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
dataset_info:
  features:
  - name: subject_id
    dtype: string
  - name: split
    dtype: string
  - name: num_slices
    dtype: int32
  - name: lesion_voxels
    dtype: int32
  - name: preview_slice_idx
    dtype: int32
  - name: adc_slice
    dtype: image
  - name: zadc_slice
    dtype: image
  - name: lesion_mask
    dtype: image
  - name: overlay
    dtype: image
  splits:
  - name: train
    num_bytes: 1370200
    num_examples: 85
  - name: val
    num_bytes: 42629
    num_examples: 4
  download_size: 1449540
  dataset_size: 1412829
---

# BONBID-HIE

**BONBID-HIE** (BOston Neonatal Brain Injury Dataset for Hypoxic Ischemic Encephalopathy) is a curated MRI dataset for neonatal HIE lesion segmentation, released as part of the MICCAI 2023 BONBID-HIE challenge.

## Dataset Summary

| Field | Details |
|---|---|
| Modality | Diffusion MRI (ADC-derived maps) |
| Body Part | Neonatal brain (term/late-preterm with HIE) |
| Subjects (Train) | 85 |
| Subjects (Val)   | 4 (Docker sanity-check split) |
| Subjects (Test)  | 44 (password-encrypted, not redistributed here) |
| Format | MetaImage `.mha` (3D volumes) |
| Total Size | ~1.3 GB |
| Scanners | GE 1.5T Signa, Siemens 3T Trio |
| License | CC BY-NC-ND 4.0 |

## Data Structure

Each split contains three parallel directories:

- `1ADC_ss/` — skull-stripped Apparent Diffusion Coefficient map (model **input**)
- `2Z_ADC/` — Z-score normalized ADC map (additional input/aid; **NOT ground truth**)
- `3LABEL/` — manual expert lesion annotation (**recommended ground truth**, train/val only)

File naming:
- `1ADC_ss/MGHNICU_{ID}-VISIT_01-ADC_ss.mha`
- `2Z_ADC/Zmap_MGHNICU_{ID}-VISIT_01-ADC_smooth2mm_clipped10.mha`
- `3LABEL/MGHNICU_{ID}-VISIT_01_lesion.mha`

Plus `BONBID2023_clinicaldata_val.xlsx` (clinical metadata for the val split).

## Ground Truth

The recommended ground truth is the **manual expert lesion annotation** in `3LABEL/`, drawn by a trained physician using MRICroN. For 27 uncertain cases, consensus was reached among three pediatric neuroradiologists. The `2Z_ADC/` map is provided as an algorithm-development aid and is NOT a ground-truth annotation.

## Notes

- Test split is omitted: it was distributed only to MICCAI 2023 challenge participants and is password-encrypted on Zenodo. Training samples (n=85) plus validation (n=4) are reproduced here.
- Val split is small (n=4) — intended as a Docker sanity-check, not a statistical validation set. Cross-validation on the train split is the typical evaluation strategy.

## Citation

```bibtex
@article{bao2024bonbid,
  title   = {{BOston Neonatal Brain Injury Data for Hypoxic Ischemic Encephalopathy (BONBID-HIE): I. MRI and Lesion Labeling}},
  author  = {Bao, Rina and Song, Ya'nan and Bates, Sara V. and others},
  journal = {Scientific Data},
  publisher = {Nature},
  year    = {2024},
  doi     = {10.1038/s41597-024-03986-7},
  url     = {https://www.nature.com/articles/s41597-024-03986-7}
}
```

## Source

Original release: [Zenodo record 10602767](https://zenodo.org/records/10602767) (V3, paper-cited)
Challenge portal: [bonbid-hie2023.grand-challenge.org](https://bonbid-hie2023.grand-challenge.org/)