mylyu/M4Raw_abdomen

M4Raw-Abdomen v1.0

Private release-candidate snapshot. The scientific files are frozen, but the final citation and ethics/consent wording must be confirmed before this repository is made public. The immutable public tag will be v1.0.0; this private snapshot is tagged v1.0.0-rc1.

M4Raw-Abdomen pairs repeated multicoil k-space acquired at 0.3 T with reconstructed NIfTI images and multi-organ segmentation masks. It contains 65 T1-weighted examinations and 66 T2-weighted examinations. Each released examination has three acquisitions. The T2 cohort additionally includes three matched free-breathing acquisitions without segmentation masks.

All H5 and NIfTI files are stored individually. A user can download one file or one examination without downloading an archive or the complete dataset.

One matched examination shown at upper-abdominal and kidney levels. Each single acquisition (acquisition 02) is paired with the arithmetic mean of all three prealigned RSS acquisitions. The two slice levels together show the liver, both kidneys, spleen, and stomach where annotated. Images are independently windowed from the 0.5th to 99.5th nonzero percentiles for display. Contours show released masks; free-breathing T2 has no mask.

Release inventory

Content	Count
T1-weighted examinations	65
Respiratory-triggered T2-weighted examinations	66
Free-breathing T2-weighted examinations	66
H5 files	591
Single-acquisition NIfTI images	131
Three-acquisition NIfTI images	131
Segmentation masks	131
Registration transform JSON files	460

The H5 folders reproduce the sequence-specific study cohorts:

• multicoil_train: T1 and respiratory-triggered T2 from the 50 development examinations per contrast, with all three acquisitions.
• multicoil_test: T1 from 15 and respiratory-triggered T2 from 16 held-out examinations, with all three acquisitions.
• multicoil_t2fb_train and multicoil_t2fb_test: the matched free-breathing T2 development and held-out acquisitions.
• nifti/T1 and nifti/T2: 1rep, 3rep, and masks.
• transforms: frozen repetition-to-reference translation metadata.
• metadata: manifests, split tables, quality findings, QA, provenance, and checksums.
• code: portable readers, conversion, download, and validation utilities.

Important PE/FE axis warning

fastMRI and VarNet users: M4Raw-Abdomen stores native k-space as [slice, coil, PE, FE]. This differs from the fastMRI Brain and M4Raw Brain convention [slice, coil, FE, PE], where PE is the final axis. Do not pass native M4Raw-Abdomen arrays directly to a stock fastMRI data transform or one-dimensional mask function: without adaptation, it will treat FE as PE.

Collection	Native in-plane order	Anatomical row/column	PE axis	Before fastMRI masking
M4Raw-Abdomen	[PE,FE]	[AP,LR] because PE=AP and FE=LR	second-to-last	swap the final two axes
fastMRI Brain / M4Raw Brain	[FE,PE]	generally [AP,LR] because FE=AP and PE=LR	last	no axis swap

Adapt each k-space slice and its RSS target in the data loader before calling fastMRI masking or VarNet code:

import h5py
import numpy as np

with h5py.File(path, "r") as h5:
    kspace_native = h5["kspace"][slice_index]             # [coil,PE,FE]
    target_native = h5["reconstruction_rss"][slice_index] # [PE,FE]
    acquired_pe_mask = h5["acquired_pe_mask"][:].astype(bool)  # [PE]

kspace_fastmri = np.ascontiguousarray(
    np.swapaxes(kspace_native, -2, -1)
)  # [coil,FE,PE]
target_fastmri = np.ascontiguousarray(
    np.swapaxes(target_native, -2, -1)
)  # [FE,PE]

assert kspace_fastmri.shape[-1] == acquired_pe_mask.size == 205

The one-dimensional acquired_pe_mask is not transposed: after the array swap, it aligns with the final PE axis. A retrospective mask must be a subset of this available-line support. For fastMRI tensor broadcasting, reshape the selected mask to [1,1,PE,1]. The packaged code/fastmri_adapter.py performs the same axis conversion without modifying the H5 file.

Load and download with Hugging Face

For a private repository, first run hf auth login. Authentication is not required after the dataset becomes public.

Install the official clients and the scientific file readers you need:

pip install datasets huggingface_hub h5py nibabel

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_abdomen"
revision = "v1.0.0-rc1"

all_files = load_dataset(
    repo_id, "file_inventory", split="train", revision=revision
)
t1_examinations = load_dataset(
    repo_id, "t1_examinations", split="train", revision=revision
)

case = t1_examinations.filter(lambda row: row["scan_id"] == 2023033001)[0]
h5_path = case["h5_rep02"]
print(len(all_files), h5_path)

Available configurations are t1_examinations, t2_examinations, t2_free_breathing, 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_abdomen",
    repo_type="dataset",
    revision="v1.0.0-rc1",
    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_abdomen",
    repo_type="dataset",
    revision="v1.0.0-rc1",
    local_dir="M4Raw_Abdomen",
)

Or download one examination and its derived files by pattern:

snapshot_download(
    repo_id="mylyu/M4Raw_abdomen",
    repo_type="dataset",
    revision="v1.0.0-rc1",
    local_dir="M4Raw_case_2023033001",
    allow_patterns=[
        "**/2023033001_*.h5",
        "nifti/**/2023033001_*.nii.gz",
        "transforms/**/2023033001_*.json",
    ],
)

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)              # (28, 4, 205, 256)
    print(h5["reconstruction_rss"].shape)  # (28, 205, 256)
    print(h5["acquired_pe_mask"][:].sum()) # 144 for T1

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, a PE support mask, and an XML header with intentionally different shapes. See the Hugging Face loading guide and H5_FORMAT.md.

Load a released NIfTI image and its geometrically matched mask with nibabel:

import nibabel as nib
import numpy as np

image = nib.load("nifti/T1/1rep/2023033001_T1.nii.gz")
mask = nib.load("nifti/T1/masks/2023033001_T1.nii.gz")
assert image.shape == mask.shape
assert np.allclose(image.affine, mask.affine)
image_array = image.get_fdata(dtype=np.float32)
mask_array = np.asarray(mask.dataobj, dtype=np.uint8)

H5 layout

kspace             [slice, coil, phase_encode, frequency_encode] complex64
reconstruction_rss [slice, phase_encode, frequency_encode]       float32
acquired_pe_mask   [phase_encode]                                 uint8
ismrmrd_header     scalar UTF-8 XML

PE and FE are the primary native array labels. Every released Cartesian abdominal sequence has 205 PE rows by 256 FE columns at 1.40625-mm spacing. For these acquisitions, PE is anterior-posterior (AP) and FE is left-right (LR), so native [PE,FE] is anatomically [AP,LR].

This order is intentional. In typical axial brain MRI, including the fastMRI Brain convention and the main M4Raw Brain contrasts, PE is generally LR and FE is AP; storing those data as [FE,PE] also produces anatomical [AP,LR] rows and columns. The encoding orders differ, but direct image display is consistent:

import h5py
import matplotlib.pyplot as plt

with h5py.File(path, "r") as h5:
    image = h5["reconstruction_rss"][slice_index]  # [PE,FE] = [AP,LR]
plt.imshow(image, cmap="gray", origin="upper")

No matrix transpose, rotation, or flip is needed for direct H5 display. The AP/LR mapping explains the display orientation but does not replace PE/FE in Cartesian reconstruction code. Encoding directions remain sequence-specific, and non-Cartesian trajectories such as radial or spiral imaging do not have a single global PE/FE pair; readers should therefore use each dataset's stated contract rather than infer axes from anatomy alone.

T1 contains 144 available PE rows at indices 30-173. T2 contains 143 rows at indices 31-173. The RSS target uses every available line; it is not a fully sampled 256-line reference. See H5_FORMAT.md for the complete contract.

fastMRI masking order

The native slice layout is [coil,PE,FE]. A stock fastMRI one-dimensional mask acts on the final axis, so the adapter transposes in memory to [coil,FE,PE]:

python code/fastmri_adapter.py multicoil_train/<file>.h5 --slice 0

The adapter never modifies the H5 file. This transpose serves the fastMRI PE-last masking convention; it is not an image-orientation correction.

Images and masks

• 1rep is the designated second acquisition.
• 3rep is the arithmetic mean of the three prealigned RSS targets.
• Masks use the same coordinates as 1rep.

Label	T1-weighted	T2-weighted
0	background	background
1	liver	liver
2	combined kidneys	combined kidneys
3	spleen	spleen
4	stomach	not used

Free-breathing T2 is reconstruction-only and has no packaged NIfTI or mask. Its H5 files are stored in respiratory-triggered acquisition-02 coordinates.

The three acquisition series are not pixel-matched, even when they share a scan_id:

Series	Slices	Slice thickness	Gap	Center spacing
Breath-hold T1w	28	4 mm	0 mm	4 mm
Respiratory-triggered T2w	20	8 mm	1 mm	9 mm
Free-breathing T2w	20	8 mm	1 mm	9 mm

The corresponding NIfTI shapes are 256 x 256 x 28 for T1w and 256 x 256 x 20 for respiratory-triggered T2w. None of the 61 shared T1/T2 IDs has an identical NIfTI shape and affine. Free-breathing T2 was globally translated into the respiratory-triggered T2 reference grid, but respiratory deformation and through-plane differences remain. Therefore, the same scan_id denotes participant association, not voxelwise correspondence.

Prealignment

The released complex k-space is translation corrected rather than an untouched scanner export. Frozen in-plane translations were applied once by a unit-magnitude Fourier phase ramp. The operation preserves k-space shape, acquired-line support, noise magnitude, and spatial resolution. Acquisition 02 is unchanged for T1 and respiratory-triggered T2. Forward and inverse transforms and their provenance are stored in every H5 and under transforms/.

No rotation, scaling, through-plane registration, deformable registration, or synthesized k-space samples were used.

Splits

The study trained separate T1 and T2 models. A participant may therefore have different physical train/test or fold roles between contrasts. Do not pool the physical folders to define a multicontrast split.

• Use metadata/sequence_specific_splits.tsv to reproduce the paper's contrast-specific 40/10 development split and held-out cohorts.
• Use metadata/paired_T1_T2_subject_disjoint_split.tsv for paired T1/T2 models: 30 training, 14 validation, and 17 test participants.
• Use metadata/optional_joint_subject_disjoint_split.tsv when pooling all available contrasts: 34 training, 15 validation, and 21 test participants.

Validation

Install the lightweight dependencies and validate the downloaded snapshot:

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 checks every H5, recomputes every RSS target, regenerates every image NIfTI, validates mask geometry and labels, and audits manifests and splits.

Intended use and limitations

M4Raw-Abdomen supports low-field abdominal MRI reconstruction, denoising, multirepetition processing, and downstream multi-organ segmentation. It is a healthy-volunteer, single-scanner research cohort and does not establish clinical generalization. The available-line RSS target is not a fully sampled ground truth, and global translation correction does not remove respiratory deformation or ghosting.

See DATA_CARD.md for annotation details, known quality findings, and the full limitations statement.

Citation and license

The final dataset and accompanying-paper citation will be added before the repository becomes public. The data are licensed under Creative Commons Attribution 4.0 International. See LICENSE.