Dataset Card for Fed-BraTS

Fed-BraTS is a Hugging Face dataset export prepared from the BraTS-GLI portion of Spirit-26/BraTS-2024-Complete. It contains multi-parametric brain MRI cases for tumor segmentation, with four MRI modalities, a segmentation mask, site metadata, and a PNG preview image for convenient inspection in the Hugging Face Dataset Viewer.

The dataset is designed to support medical image segmentation experiments and federated learning workflows where cases can be partitioned by site or sampled across sites.

Dataset Details

Dataset Description

Each row represents one BraTS subject. The volumetric MRI scans and segmentation mask are stored in the repository as NIfTI files (.nii.gz). The dataset table stores repo-relative paths to those files, rather than decoding the NIfTI volumes directly in the Dataset Viewer. This keeps the Hub preview stable while preserving access to the original medical imaging files.

The dataset includes the following imaging channels:

• t1n: native T1-weighted MRI
• t1c: post-contrast T1-weighted MRI
• t2w: T2-weighted MRI
• t2f: T2-FLAIR MRI
• Mask: segmentation mask
• preview_t1c: embedded middle-slice PNG preview generated from the t1c volume

Dataset Sources

• Source dataset: Spirit-26/BraTS-2024-Complete
• Derived subset: BraTS-GLI training cases selected from metadata entries marked as train or train-additional
• Export repository: flwrlabs/Fed-BraTS

Uses

Direct Use

This dataset can be used for training, validating, and benchmarking brain tumor segmentation models on multi-parametric MRI data. It is especially suitable for workflows that require explicit access to each modality and segmentation mask as NIfTI files.

Direct Use in Federated Learning

The Site column can be used to create site-based partitions for federated learning experiments. For example, each institution/site can be mapped to a client, or cases can be partitioned using IID or non-IID strategies depending on the experimental setup.

Example with Flower Datasets:

from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import NaturalIdPartitioner

fds = FederatedDataset(
    dataset="flwrlabs/Fed-BraTS",
    partitioners={"train": NaturalIdPartitioner(partition_by="Site")},
)
partition = fds.load_partition(partition_id=0)

Out-of-Scope Use

This dataset is intended for research and model development. It is not intended for clinical diagnosis, treatment planning, or any use where predictions could directly affect patient care without appropriate validation, regulatory review, and clinical oversight.

Dataset Structure

The dataset has one configuration, default, with two splits:

• train: training split generated from the source training pool
• test: held-out internal validation split generated per site from the source training pool

The split generation is stratified by Site: cases are grouped by site, shuffled with a fixed seed, and approximately 20% per site is assigned to the held-out test split when possible.

Data Instances

A typical row has the following structure:

{
    "BraTS Subject ID": "BraTS-GLI-00000-000",
    "Site": "ExampleSite",
    "t1n": "data/train/BraTS-GLI-00000-000/BraTS-GLI-00000-000-t1n.nii.gz",
    "t1c": "data/train/BraTS-GLI-00000-000/BraTS-GLI-00000-000-t1c.nii.gz",
    "t2w": "data/train/BraTS-GLI-00000-000/BraTS-GLI-00000-000-t2w.nii.gz",
    "t2f": "data/train/BraTS-GLI-00000-000/BraTS-GLI-00000-000-t2f.nii.gz",
    "Mask": "data/train/BraTS-GLI-00000-000/BraTS-GLI-00000-000-seg.nii.gz",
    "preview_t1c": "<embedded PNG image>"
}

Data Fields

• BraTS Subject ID (string): BraTS subject identifier.
• Site (string): acquisition or contributing site identifier from the source metadata.
• t1n (string): repo-relative path to the native T1-weighted MRI NIfTI file.
• t1c (string): repo-relative path to the post-contrast T1-weighted MRI NIfTI file.
• t2w (string): repo-relative path to the T2-weighted MRI NIfTI file.
• t2f (string): repo-relative path to the T2-FLAIR MRI NIfTI file.
• Mask (string): repo-relative path to the segmentation mask NIfTI file.
• preview_t1c (image): embedded PNG preview of the middle axial slice from the t1c volume.

Data Splits

The export script creates the following split files:

Split	File
train	metadata/train.parquet
test	metadata/test.parquet

The NIfTI files referenced by the Parquet rows are stored under:

data/train/<BraTS Subject ID>/
data/test/<BraTS Subject ID>/

Preview PNGs are stored under:

previews/train/
previews/test/

Dataset Creation

Curation Rationale

The dataset was exported to make the BraTS-GLI imaging data easier to use from the Hugging Face Hub while avoiding Dataset Viewer issues associated with decoding NIfTI files directly. NIfTI volumes are therefore preserved as files in the repository, and the tabular dataset contains string paths plus an image preview column.

Processing Steps

1. Read and normalize the source metadata columns.
2. Select eligible training cases.
3. Verify that each case has t1n, t1c, t2w, t2f, and a segmentation mask.
4. Split cases into train and test pools using a per-site validation ratio.
5. Copy NIfTI files into split-specific directories.
6. Generate a middle-slice PNG preview from each t1c image.
7. Write split metadata as Parquet files with Hugging Face datasets features.

Loading the Dataset

from datasets import load_dataset

_ds = load_dataset("flwrlabs/Fed-BraTS")
print(_ds)
print(_ds["train"][0])

The modality and mask columns contain repo-relative paths. To load the NIfTI files after downloading or cloning the repository, use nibabel:

from pathlib import Path
import nibabel as nib
from datasets import load_dataset

repo_root = Path("/path/to/Fed-BraTS")
ds = load_dataset("flwrlabs/Fed-BraTS")
row = ds["train"][0]

t1c_path = repo_root / row["t1c"]
t1c = nib.load(str(t1c_path)).get_fdata()

Limitations

• The dataset table stores NIfTI paths as strings; users must download the repository files to load the full 3D volumes.
• preview_t1c is only a 2D visualization preview and should not be used as a substitute for the full MRI volume.
• The test split is an internally generated held-out split from the source training pool, not necessarily the official BraTS challenge validation/test split.
• Site metadata may reflect acquisition or contributing-site information from the source metadata, but users should verify exact semantics against the original BraTS documentation.

License and Terms of Use

This dataset is derived from Spirit-26/BraTS-2024-Complete, which aggregates BraTS 2024 collections with collection-specific terms. Users are responsible for complying with the original BraTS/Synapse terms and citation requirements. The BraTS-GLI component is described by the source dataset as research-only and requiring citation.

Citation

If you use this dataset, cite the original BraTS paper. If you're using this dataset with Flower Datasets, you can cite Flower.

@article{menze2014multimodal,
  title={The multimodal brain tumor image segmentation benchmark (BRATS)},
  author={Menze, Bjoern H and Jakab, Andras and Bauer, Stefan and Kalpathy-Cramer, Jayashree and Farahani, Keyvan and Kirby, Justin and Burren, Yuliya and Porz, Nicole and Slotboom, Johannes and Wiest, Roland and others},
  journal={IEEE transactions on medical imaging},
  volume={34},
  number={10},
  pages={1993--2024},
  year={2014},
  publisher={IEEE}
}

@article{DBLP:journals/corr/abs-2007-14390,
  author       = {Daniel J. Beutel and
                  Taner Topal and
                  Akhil Mathur and
                  Xinchi Qiu and
                  Titouan Parcollet and
                  Nicholas D. Lane},
  title        = {Flower: {A} Friendly Federated Learning Research Framework},
  journal      = {CoRR},
  volume       = {abs/2007.14390},
  year         = {2020},
  url          = {https://arxiv.org/abs/2007.14390},
  eprinttype    = {arXiv},
  eprint       = {2007.14390},
  timestamp    = {Mon, 03 Aug 2020 14:32:13 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2007-14390.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Dataset Card Contact

For questions about this Hugging Face export, contact the maintainers of flwrlabs/Fed-BraTS. For questions about the original data, licensing, or challenge rules, refer to the official BraTS resources and the source dataset documentation.