mylyu/M4Raw_brain

M4Raw Brain v1.6

M4Raw Brain is a multi-contrast, multi-repetition, four-channel k-space dataset acquired on a 0.3-T whole-body MRI system. The release contains T1w, T2w, FLAIR, and GRE brain acquisitions from healthy volunteers, including explicit motion subsets and an expanded repeated-acquisition test cohort.

Companion dataset: M4Raw-Abdomen is a separate low-field abdominal MRI k-space and segmentation dataset and will be made public soon. Until then, the linked private repository is accessible only to authorized users.

This Hugging Face distribution exposes the scientific H5 files from Zenodo record 8056074 individually, so one acquisition or examination can be accessed without downloading a multi-gigabyte ZIP archive. All 2,230 scientific H5 payloads are byte-identical to the corresponding files extracted from the Zenodo archives: they have not been recompressed, rewritten, or numerically changed. The Hub repository as a whole is not a byte-identical copy of the original distribution because ZIP containers were removed, paths were reorganized, and Hub-specific documentation and metadata were added.

One participant at the same stored central slice. The top row shows acquisition 01; the bottom row shows the repeat average (three acquisitions for T1w/T2w and two for FLAIR/GRE). Every stored RSS slice is displayed directly, following the original M4Raw tutorial, without transpose, rotation, or anatomical flip. Each contrast is independently windowed from the 0.5th to 99.5th nonzero percentiles for display.

Release inventory

Content	Participants	H5 files
Training: T1w/T2w/FLAIR	128	1,024
Validation: T1w/T2w/FLAIR	30	240
Motion subsets: T1w/T2w/FLAIR	25	200
Expanded test: T1w/T2w/FLAIR	25	400
GRE for the original cohort	183	366
Total	208 unique	2,230

The released motion subsets contain 25 participants: eight in the inter-scan subset and 17 in the intra-scan subset, totaling 200 H5 files. One sentence in the paper's subset-partition Methods paragraph reports 26 participants, while the paper's cohort summaries and the released files correspond to 25. For operational use, metadata/subjects.tsv and the sequence manifests are authoritative.

T1w and T2w contain three repetitions in the training, validation, and motion subsets and six repetitions in the test subset. FLAIR contains two and four repetitions, respectively. GRE contains two repetitions for each of the 183 original-cohort participants.

Layout

multicoil_train/                    1,024 H5
multicoil_val/                        240 H5
multicoil_test/                       400 H5
multicoil_motion/inter_scan/           64 H5
multicoil_motion/intra_scan/          136 H5
multicoil_gre/                        366 H5
metadata/                         manifests and provenance
code/                             portable readers and validator
assets/                           reproducible homepage montage

The original scientific filenames are retained: <scan_id>_<sequence><repetition>.h5, for example 2022061003_T201.h5.

Load and download with Hugging Face

Install the official clients and HDF5 reader:

pip install datasets huggingface_hub h5py

Use load_dataset to load the complete file inventory or a sequence manifest. These tables let you discover paths and select examinations without downloading the scientific payloads:

from datasets import load_dataset

repo_id = "mylyu/M4Raw_brain"
revision = "v1.6.0"

all_files = load_dataset(
    repo_id, "file_inventory", split="train", revision=revision
)
t2_examinations = load_dataset(
    repo_id, "t2_examinations", split="train", revision=revision
)

case = t2_examinations.filter(lambda row: row["scan_id"] == 2022061003)[0]
h5_path = case["h5_rep01"]
print(len(all_files), h5_path)

Available configurations are t1_examinations, t2_examinations, flair_examinations, gre_examinations, acquisitions, subjects, and file_inventory.

Download the selected H5 path from the manifest:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="mylyu/M4Raw_brain",
    repo_type="dataset",
    revision="v1.6.0",
    filename=h5_path,
)
print(path)

Download every repository file with snapshot_download:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="mylyu/M4Raw_brain",
    repo_type="dataset",
    revision="v1.6.0",
    local_dir="M4Raw_Brain",
)

Or download one examination by pattern:

snapshot_download(
    repo_id="mylyu/M4Raw_brain",
    repo_type="dataset",
    revision="v1.6.0",
    local_dir="M4Raw_case_2022061003",
    allow_patterns=["**/2022061003_*.h5"],
)

For random-access reading of one remote H5, HfFileSystem works directly with h5py and uses HTTP range requests:

import h5py
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
remote_path = f"datasets/{repo_id}@{revision}/{h5_path}"
with fs.open(remote_path, "rb") as stream, h5py.File(stream, "r") as h5:
    print(h5["kspace"].shape)              # (18, 4, 256, 256)
    print(h5["reconstruction_rss"].shape)  # (18, 256, 256)

The generic load_dataset("hdf5", ...) loader is not used for the H5 payloads because it expects a tabular HDF5 layout whose datasets have the same first-dimension length. M4Raw instead stores k-space, an RSS target, and an XML header with intentionally different shapes. See the Hugging Face loading guide and H5_FORMAT.md.

H5 layout

kspace             [slice, coil, frequency_encode, phase_encode] complex64
reconstruction_rss [slice, frequency_encode, phase_encode]       float32
ismrmrd_header     scalar UTF-8 XML

Brain k-space already follows fastMRI's practical masking convention: phase encoding is the final axis. No in-memory transpose is needed. This differs from M4Raw-Abdomen, whose packaged adapter transposes native PE/FE axes. Both repositories expose the same read_h5, read_slice, apply_pe_mask, and download_case.py interfaces; see H5_FORMAT.md.

The original Brain H5 files do not store an acquired_pe_mask dataset. The common reader derives it from nonzero k-space support and returns it through the same API used by M4Raw-Abdomen.

Every stored k-space and RSS matrix has a nominal 256 x 256 reconstruction grid, but the number of acquired phase-encoding lines is sequence dependent:

Sequence	H5 files	Acquired PE lines	Stored PE size
T1w	699	195	256
T2w	699	195	256
FLAIR	466	198	256
GRE	366	256	256

Thus, a 256 x 256 array does not by itself imply complete PE sampling. Unavailable PE rows are zero in the stored k-space, and reconstruction_rss uses all available source lines. Per-file support is recorded in metadata/acquisitions.tsv as acquired_pe_lines.

Two legacy custom XML extensions require care: kspaceDimension labels PE/FE opposite to the verified stored layout for T1w, T2w, and FLAIR (the GRE labels match), and repetitionInformation contains template references that do not reliably identify the accompanying repetitions. Stock fastMRI ignores both fields. Use the documented axis contract and packaged sequence manifests instead; see metadata/legacy_header_findings.tsv.

Splits and references

• multicoil_train and multicoil_val are the original motion-screened development split from the Scientific Data paper.
• multicoil_motion intentionally retains examinations classified as inter-scan or intra-scan motion.
• multicoil_test is the 25-participant v1.6 test cohort with twice the usual repetition count, supporting higher-SNR repeat averages.
• multicoil_gre covers the 183 original-cohort participants and is not available for the 25 added test participants.

Participants are disjoint across training, validation, motion, and test. Use metadata/subjects.tsv and the sequence manifests rather than inferring membership from folder names alone.

Validation

python -m pip install -r code/requirements.txt
sha256sum -c SHA256SUMS
python code/validate_release.py --release-root . --mode full --workers 8 \
  --skip-checksums

The validator is read-only. Full mode opens every H5, checks XML and array contracts, derives PE support, checks finite values, and independently recomputes every RSS target.

Processing and limitations

The scanner-exported k-space was corrected for image off-centering by phase modulation before public release; reconstruction_rss was computed from the corrected multicoil k-space. No fully sampled, noise-free ground truth is provided. Repeat averages are higher-SNR references rather than independent anatomical ground truth.

GRE uses an anterior-posterior phase-encoding direction, unlike the other contrasts, and GRE was not included in the original motion screen. The dataset contains healthy volunteers from one scanner and one four-channel head coil, so performance may not generalize to other systems or clinical populations. See DATA_CARD.md for the full scope and limitations.

Citation

Please cite:

Lyu M, Mei L, Huang S, et al. M4Raw: A multi-contrast, multi-repetition, multi-channel MRI k-space dataset for low-field MRI research. Scientific Data. 2023;10:264. https://doi.org/10.1038/s41597-023-02181-4

@article{lyu_m4raw_2023,
  title = {M4Raw: A multi-contrast, multi-repetition, multi-channel MRI
           k-space dataset for low-field MRI research},
  author = {Lyu, Mengye and Mei, Lifeng and Huang, Shoujin and others},
  journal = {Scientific Data},
  volume = {10},
  pages = {264},
  year = {2023},
  doi = {10.1038/s41597-023-02181-4}
}

The dataset is licensed under Creative Commons Attribution 4.0 International. See LICENSE.