MRIxFields2026

Welcome to MRIxFields2026, a multi-field structural brain MRI resource spanning five magnetic field strengths: 0.1T, 1.5T, 3T, 5T, and 7T.

MRIxFields2026 contains 2539 MRI acquisitions from 1095 participants, covering T1-weighted (T1W), T2-weighted (T2W), and T2-weighted fluid-attenuated inversion recovery (T2-FLAIR) imaging. The dataset combines a fully paired cohort of 40 healthy volunteers (600 acquisitions) with five field-specific unpaired cohorts comprising 1055 healthy volunteers and participants with neurologic conditions (1939 acquisitions).

Motivation

Public brain MRI datasets spanning a broad range of magnetic field strengths remain limited, particularly for subject-matched acquisitions across multiple fields. MRIxFields2026 addresses this gap by providing multi-contrast brain MRI from 0.1T to 7T, combining paired and unpaired cohorts to support the development and systematic evaluation of cross-field MRI synthesis, enhancement, and harmonization methods.

Dataset

MRIxFields2026 consists of two complementary components:

• Paired cohort: 40 healthy volunteers scanned at all five field strengths with T1w, T2w, and T2FLAIR imaging, resulting in 600 subject-matched acquisitions.
• Unpaired cohort: 1055 participants (1939 acquisitions), with varying imaging contrast availability across field strengths.

Research Tasks

MRIxFields2026 supports a range of cross-field neuroimaging applications, including:

• Cross-field MRI synthesis and translation
• Ultra-low-field MRI enhancement
• Cross-field MRI harmonization
• Foundation model pretraining and domain generalization
• Anatomical fidelity and quantitative morphometry evaluation

The paired cohort enables direct assessment of anatomical fidelity across field strengths, while the unpaired cohorts provide a broader setting for evaluating model robustness and generalizability.

Contents

Repository Structure

The MRIxFields2026 dataset is organized into two main collections: Radiology-AI-BIDs and Radiology-AI-orig-BIDs. Each collection is distributed as a multi-volume 7-Zip archive to facilitate the transfer of large imaging datasets.

The archive volumes are stored in the respective data/ directories. Two JSON manifest files are provided for each collection: one documenting the original dataset files and the other listing the corresponding 7-Zip archive volumes.

All volumes belonging to the same archive must be downloaded before extraction. The original directory structure and file organization are restored after successful extraction.

MRIxFields2026/
|
|-- README.md
|
|-- participant_metadata.xlsx
|
|-- Radiology-AI-BIDs/
|   |
|   |-- data/
|   |   |-- Radiology-AI-BIDs.7z.001
|   |   |-- Radiology-AI-BIDs.7z.002
|   |   |-- ...
|   |   `-- Radiology-AI-BIDs.7z.021
|   |
|   |-- manifest-For-Radiology-AI-BIDs-files.json
|   `-- manifest-Radiology-AI-BIDs-7z-parts.json
|
`-- Radiology-AI-orig-BIDs/
    |
    |-- data/
    |   |-- Radiology-AI-orig-BIDs.7z.001
    |   |-- Radiology-AI-orig-BIDs.7z.002
    |   |-- ...
    |   `-- Radiology-AI-orig-BIDs.7z.011
    |
    |-- manifest-For-Radiology-AI-orig-BIDs-files.json
    `-- manifest-Radiology-AI-orig-BIDs-7z-parts.json

Download and Extraction

For detailed instructions on downloading datasets from Hugging Face, please refer to the official Hugging Face download guide.

The dataset is distributed as multi-volume 7-Zip archives. To extract the data, all volumes belonging to the same archive must first be downloaded and placed in the same directory. The original filenames and file extensions should be preserved.

Extraction using 7-Zip

The archives can be extracted using 7-Zip on Windows, Linux, or macOS.

For Linux, run the following commands in the directory containing the downloaded archive volumes:

Preprocessed images:

7z x Radiology-AI-BIDs.7z.001

Original reconstructed images:

7z x Radiology-AI-orig-BIDs.7z.001

For Windows, right-click the first volume (.7z.001), select 7-Zip → Extract Here or Extract to..., and enter the extraction password when prompted.

Only the first volume (.7z.001) needs to be selected to initiate extraction. 7-Zip will automatically read the subsequent volumes, provided that all volumes are present in the same directory.

The extraction password is provided separately to users with approved access, and in Data Availability Statement in our manuscript. Upon successful extraction, the original BIDS-style directory structure will be restored.

Extracted Data Structure

MRIxFields2026 provides two versions of the MRI data, organized in BIDS-style directories:

• For-Radiology-AI-orig-BIDs/ (Original Images): Contains original reconstructed MRI images without preprocessing or spatial registration, preserving the original image characteristics and spatial information.
• For-Radiology-AI-BIDs/ (Preprocessed Images): Contains preprocessed and spatially registered MRI images, facilitating cross-field comparisons, image synthesis, harmonization, and quantitative neuroimaging analysis.

All images are provided in compressed NIfTI format (.nii.gz) with identifying metadata removed, accompanied by corresponding JSON metadata files (.json).

MRIxFields2026/
|
|-- For-Radiology-AI-BIDs/
|   |
|   |-- sub-P0001/
|   |   |-- ses-0p1T/
|   |   |   `-- anat/
|   |   |       |-- sub-P0001_ses-0p1T_T1w.nii.gz
|   |   |       |-- sub-P0001_ses-0p1T_T1w.json
|   |   |       |-- sub-P0001_ses-0p1T_T2w.nii.gz
|   |   |       |-- sub-P0001_ses-0p1T_T2w.json
|   |   |       |-- sub-P0001_ses-0p1T_FLAIR.nii.gz
|   |   |       `-- sub-P0001_ses-0p1T_FLAIR.json
|   |   |
|   |   |-- ses-1p5T/anat/
|   |   |-- ses-3T/anat/
|   |   |-- ses-5T/anat/
|   |   `-- ses-7T/anat/
|   |
|   |-- sub-P0002/ ... sub-P0040/
|   |
|   |-- sub-R0001/
|   |   `-- ses-0p1T/
|   |       `-- anat/
|   |           |-- sub-R0001_ses-0p1T_T1w.nii.gz
|   |           |-- sub-R0001_ses-0p1T_T1w.json
|   |           `-- ... (Other available contrasts)
|   |
|   `-- sub-R0002/ ... sub-R1055/
|
`-- For-Radiology-AI-orig-BIDs/
    |
    |-- sub-P0001/
    |   |-- ses-0p1T/anat/
    |   |-- ses-1p5T/anat/
    |   |-- ses-3T/anat/
    |   |-- ses-5T/anat/
    |   `-- ses-7T/anat/
    |
    |-- sub-P0002/ ... sub-P0040/
    |
    |-- sub-R0001/
    |   `-- ses-0p1T/anat/
    |
    `-- sub-R0002/ ... sub-R1055/

For the preprocessed dataset, the following procedures were performed:

Paired data: T2w and FLAIR images were rigidly registered to the corresponding T1w images at each field strength. Following skull stripping with HD-BET, 3T T1w images were rigidly registered to the 0.5-mm MNI152 template, while T1w images from the other field strengths were registered to the corresponding 3T images. These transformations were applied to T2w and FLAIR images, with ANTs SyN used to correct residual misalignment when necessary.

Unpaired data: Images were skull stripped using HD-BET and rigidly registered to the MNI152 template.

Intensity preprocessing: N4 bias field correction and min-max intensity normalization to the range of 0-1 were applied.

Quality control: Registration quality was visually assessed, and cases that failed this assessment were excluded.

The original dataset retains the reconstructed images without these preprocessing steps.

Note: The two directories contain different versions of the same MRI acquisitions and should not be counted as independent datasets. The original images are reconstructed MRI images rather than raw k-space data. Detailed preprocessing methods are described in the accompanying publication.

Participant Metadata

A participant-level metadata file (participant_metadata.xlsx) is provided in the root directory of the repository. It contains demographic, imaging-availability, and clinical information for participants included in MRIxFields2026.

The spreadsheet includes the following information:

• Participant ID (ID): An anonymized identifier used to link participant-level metadata with the imaging data.
• Field-strength information (Fields): Indicates All for paired participants or the corresponding magnetic field strength for unpaired participants.
• Imaging contrast availability (T1W, T2W, and T2FLAIR): Indicates the availability of each imaging contrast, with check marks denoting available acquisitions.
• Demographic information (Age and Sex): Age is recorded in years, and sex is represented as M or F.
• Clinical information: Includes white matter hyperintensities, infarction / encephalomalacia, cerebral atrophy, ventriculomegaly, and cavum septum pellucidum. Check marks indicate recorded clinical findings.

Participant identifiers in the spreadsheet follow the same BIDS-style subject labeling convention as the imaging directories, ranging from sub-P0001 to sub-P0040 for the paired cohort and sub-R0001 to sub-R1055 for the unpaired cohort. This consistent naming convention allows direct matching of participant-level metadata with the corresponding MRI images.

Access to the participant metadata is subject to the same Data Use Agreement (DUA) and access approval requirements as the imaging data.

File Naming Convention

All MRI images are stored in compressed NIfTI format (.nii.gz), accompanied by corresponding JSON metadata files (.json). Both original and preprocessed images follow the same BIDS-style naming convention and are organized by subject, session, and imaging contrast.

Files follow the naming convention:

sub-<SubjectID>_ses-<SessionID>_<Contrast>.nii.gz

The corresponding JSON metadata files use the same filename stem with a .json extension.

Paired Data

Subject identifiers range from P0001 to P0040, with five sessions corresponding to magnetic field strengths of 0.1T, 1.5T, 3T, 5T, and 7T.

For example:

• sub-P0001_ses-0p1T_T1w.nii.gz
• sub-P0001_ses-1p5T_FLAIR.nii.gz
• sub-P0001_ses-3T_T2w.nii.gz

The same subject identifier is maintained across all five field strengths and three imaging contrasts, allowing direct identification of subject-matched acquisitions.

Unpaired Data

Subject identifiers range from R0001 to R1055, with each participant typically having one session corresponding to the acquired magnetic field strength.

For example:

• sub-R0001_ses-0p1T_T1w.nii.gz
• sub-R0001_ses-0p1T_T2w.nii.gz
• sub-R0101_ses-1p5T_T1w.nii.gz

The number of available imaging contrasts varies across participants and field strengths.

Naming Components

• Subject ID: P0001-P0040 for paired subjects and R0001-R1055 for unpaired subjects.
• Session ID: 0p1T, 1p5T, 3T, 5T, or 7T, indicating the magnetic field strength.
• Contrast: T1w, T2w, or FLAIR.
• File format: .nii.gz for MRI images and .json for corresponding metadata.

Both For-Radiology-AI-BIDs/ and For-Radiology-AI-orig-BIDs/ follow this naming convention. The two directories contain preprocessed and original reconstructed images, respectively.

MRIxFields2026 Challenge

The MRIxFields2026 Challenge was organized in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2026. The challenge focused on three tasks: ultrahigh-field MRI synthesis, ultra-low-field MRI enhancement, and unified cross-field translation. It provides a benchmarking platform with baseline models, tutorials, and evaluation tools to facilitate method development and enable fair performance comparisons. Further details and resources are available in the MRIxFields2026 Challenge GitHub repository.

MRIxFields Consortium Publication Requirements

For studies using MRIxFields series data, prior to manuscript submission, users must notify the MRIxFields Consortium (mrixfields@outlook.com), providing:

A description of how the dataset was used in the study.

A copy of the manuscript intended for submission.

Furthermore, all resulting publications must include “MRIxFields Consortium” as a group author in the author list, typically listed at the end of the author list. Group authorship does not diminish the importance of the corresponding author. In addition, all resulting publications must cite the MRIxFields article (specific details will be announced upon publication) as a reference.

For the detailed information of MRIxFields Consortium, please download the document here.