Data

Data sources

MRIxFields2026 provides a large-scale multi-field MRI dataset spanning 0.1T to 7T, integrating unpaired multi-center cohorts with prospectively acquired travelling-volunteer paired data.

💾 Data Overview: Scale & Diversity

• Total Volume:

  • Total Volume: The dataset comprises approximately 2500+ retrospective unpaired cases (5 fields; 3 modalities ) and 600 paired cases (40 paired volunteers×5 fields×3 modalities).

• ○Multi-field structural brain MRI data featuring:5 Field Strengths: 0.1T, 1.5T, 3T, 5T, 7T systems.

  • 3 Modalities: 3D T1-Weighted (T1W), 2D/3D T2-Weighted (T2W), 2D/3D T2 FLAIR.(For 2D data, we utilize a pre-trained super-resolution network to reconstruct 3D volumes from 2D inputs.)
  • 3 Manufacturers: Siemens Healthineers, United Imaging Healthcare, Point Imaging.

Field Strength	Prospective T1w (paired)	Prospective T2w (paired)	Prospective T2 FLAIR (paired)	Retrospective T1w (unpaired)	Retrospective T2w (unpaired)	Retrospective T2 FLAIR (unpaired)	Scanners	Manufacturer
0.1T	40 (3D)	40 (3D)	40 (2D)	100 (3D)	100 (3D)	99 (2D)	piMR-820H	Point Imaging
1.5T	40 (3D)	40 (3D)	40 (2D)	104 (3D)	221 (2D)	221 (2D)	uMR 670	United Imaging Healthcare
3T	40 (3D)	40 (3D)	40 (2D)	143 (3D)	142 (3D)	143 (2D)	MAGNETOM Prisma	Siemens Healthineers
5T	40 (3D)	40 (3D)	40 (2D)	121 (3D)	43 (3D)	102 (2D)	uMR Jupiter	United Imaging Healthcare
7T	40 (3D)	40 (3D)	40 (2D)	235 (3D)	81 (3D)	84 (2D)	MAGNETOM Terra	Siemens Healthineers

📊 Dataset seperation: Training, Validation, and Testing

Overall Split

• Training set: 1945+ cases (T1W, T2W, T2 FLAIR), consisting of the 1900+ retrospective unpaired cases together with 45 paired cases (3 paired volunteers× 5 fields×3 modalities)
• Validation set: 255 cases, corresponding to 17 volunteers × 5 fields × 3 modalities
• Test set: 300 cases, corresponding to 20 held-out volunteers × 5 fields × 3 modalities

The precise role of each split varies across tasks, depending on whether the task is driven primarily by unpaired learning, paired target-domain supervision, or controllable conditional generation.

Task 1: Ultra-High-Field MRI Synthesis from Arbitrary Input Field Strengths

• Training Set (1945+ cases)

  • The training set consists 1900+ retrospective unpaired cases, supplemented with 45 paired cases (3 paired volunteers× 5 fields×3 modalities).

• Validation Set (255 cases):

  • The validation set consists of 17 travelling volunteers scanned across five field strengths and 3 modalities (17 × 5 × 3 =255 cases).

  These paired data enable controlled within-case validation against corresponding 7T targets.

• Test Set (300 cases):

  • The test set includes 20 held-out travelling volunteers scanned across five field strengths and 3 modalities (20 × 5 × 3 = 300 cases).

  Paired 7T references are withheld and used exclusively for final evaluation.

Task 2: Higher-Field MRI Generation from Ultra-Low-Field MRI

• Training Set (1945+ cases)

  • The training set consists 1900+ retrospective unpaired cases, supplemented with 45 paired cases (3 paired volunteers× 5 fields×3 modalities).

• Validation Set (255 cases):

  • The validation set consists of 17 travelling volunteers (17 × 5 × 3 =255 cases), including paired 0.1T and higher-field scans with 3 modalities for within-case evaluation.

• Test Set (300 cases):

  • The test set consists of 20 held-out travelling volunteers (20 × 5 × 3 = 300 cases).

  The 0.1T images serve as inputs, while higher-field references are withheld for evaluation.

Task 3: Controllable Field-to-Field MRI Synthesis with a Unified Conditional Model

• Training Set (1945+ cases)

  • The training set consists 1900+ retrospective unpaired cases, supplemented with 45 paired cases (3 paired volunteers× 5 fields×3 modalities).

• Validation Set (255 cases):

  • The validation set consists of 17 travelling volunteers (N = 255 cases), enabling evaluation across all possible field-to-field translation directions.

• Test Set (300 cases):

  • The test set consists of 20 held-out travelling volunteers (N = 300 cases).

All target-domain references are withheld and used only for final ranking.

Pre-processing

To ensure anatomical consistency across field strengths while preserving realistic cross-field variability, we designed a unified yet cohort-specific preprocessing pipeline. The pipeline treats the prospective travelling-volunteer paired cohort and the retrospective unpaired multi-field cohort differently, reflecting their distinct roles in training and evaluation.

1. Spatial Registration

1.1 Prospective Travelling-Volunteer Cohort (Paired Data)

For the paired dataset, we adopted a case-specific, 3T-anchored registration strategy to preserve within-case anatomical correspondence across field strengths while minimizing artificial deformation.

Step 0: Intra-case multi-contrast alignment

For each case, T2W and T2 FLAIR images were first rigidly registered to the corresponding T1W image to ensure within-case cross-contrast alignment.

Step 1: Skull stripping

Skull stripping was then performed using HD-BET on the T1W images, and the resulting brain masks were propagated to T2W and T2 FLAIR images.

Step 2: Reference definition

The skull-stripped 3T T1W image was rigidly registered to the MNI152 (0.5 mm, skull-off) template, establishing a case-specific reference in standard space.

Step 3: Cross-field alignment

T1W images from all other field strengths (0.1T, 1.5T, 5T, and 7T) were rigidly registered to the case-specific 3T T1W image in MNI space, ensuring consistent global alignment across fields.

Step 4: Nonlinear refinement (optional)

For cases with residual misalignment, SyN nonlinear registration (ANTs) was applied as a refinement step to improve local anatomical correspondence.

Multi-contrast propagation

The estimated transformations from T1W images were propagated to the corresponding T2W and T2 FLAIR images. When necessary, additional SyN-based refinement was applied to these modalities to correct residual mismatch.

1.2 Retrospective Multi-Field Cohort (Unpaired Data)

For the retrospective dataset, 0.1T images were first skull-stripped using HD-BET, after which all modalities were registered to the MNI152 template (0.5 mm, skull-off).

2. Bias Field Correction and Intensity Normalization

After spatial registration, N4 bias field correction was applied to mitigate intensity inhomogeneity.

All images were then normalized to the range of (0, 1) using min-max scaling.

3. 2D-to-3D Super-Resolution Reconstruction

For acquisitions with inherently anisotropic resolution (i.e., 2D scans), we employed a learning-based super-resolution strategy to reconstruct volumetric 3D data. Specifically, a pre-trained super-resolution network was used to perform slice-to-volume reconstruction, transforming 2D inputs into isotropic 3D volumes. This approach enables recovery of through-plane structural information and improves cross-plane consistency, which is critical for subsequent cross-field analysis and synthesis tasks.

4. Quality Control

All data underwent expert-guided quality control, including:

• Removal of scans with severe motion artifacts
• Exclusion of images with incomplete brain coverage
• Verification of registration accuracy across field strengths and modalities

This process ensures high data quality and reliable anatomical alignment across the dataset.

5. Data Organization

All processed data are organized according to the BIDS (Brain Imaging Data Structure) standard, with consistent metadata across volunteers, field strengths, and acquisition sites.

Reference Information

• ANTs registration: GitHub
• Super-resolution: GitHub
• HD-BET: GitHub

Post-processing

To enable anatomically informed downstream analysis, we incorporated a post-processing step based on automated brain segmentation.

1. Automated Segmentation

All processed MRI volumes were segmented using SynthSeg, a deep learning-based whole-brain segmentation framework. SynthSeg was selected due to its robustness to variations in image contrast, resolution, and acquisition conditions, making it particularly suitable for multi-field MRI data.

Segmentation was applied uniformly to all images after preprocessing, without additional manual intervention.

2. Structural Feature Extraction

Based on the segmentation outputs, regional structural measures were derived for each case, including:

• Regional brain volumes for deep gray matter structures
• Bilateral measurements across anatomical regions

The primary structures include:

• Thalamus
• Caudate
• Putamen
• Pallidum
• Hippocampus
• Amygdala
• Accumbens

These regions are consistently identifiable across field strengths and are suitable for downstream quantitative analysis.

3. Quality Control of Segmentation

To ensure robustness of the segmentation outputs:

• SynthSeg was applied under a unified configuration across all datasets
• Outputs were visually inspected for gross failures or anatomical inconsistencies
• Cases with severe segmentation artifacts were flagged during the quality control stage

4. Role in the Pipeline

This post-processing step provides anatomically meaningful representations of the MRI data, enabling:

• Standardized structural characterization across field strengths
• Region-wise quantitative analysis
• Compatibility with downstream evaluation and analysis pipelines

Reference Information

• SynthSeg: GitHub Repository
• Paper: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining

---

Data Directory Structure

ChallengeData
├── Training_retrospective/          # Training (unpaired, different volunteers)
│   ├── T1W/
│   │   ├── 0.1T/*.nii.gz
│   │   ├── 1.5T/*.nii.gz
│   │   ├── 3T/*.nii.gz
│   │   ├── 5T/*.nii.gz
│   │   └── 7T/*.nii.gz
│   ├── T2W/
│   │   ├── 0.1T/*.nii.gz
│   │   ├── 1.5T/*.nii.gz
│   │   ├── 3T/*.nii.gz
│   │   ├── 5T/*.nii.gz
│   │   └── 7T/*.nii.gz
│   └── T2FLAIR/
│       ├── 0.1T/*.nii.gz
│       ├── 1.5T/*.nii.gz
│       ├── 3T/*.nii.gz
│       ├── 5T/*.nii.gz
│       └── 7T/*.nii.gz
├── Training_prospective/            # Training (paired, travelling volunteers)
│   ├── T1W/ {0.1T, 1.5T, 3T, 5T, 7T}
│   ├── T2W/{...}/
│   └── T2FLAIR/{...}/
├── Validating_prospective/          # Validation (paired)
│   ├── T1W/{0.1T, 1.5T, 3T, 5T, 7T}
│   ├── T2W/{...}/
│   └── T2FLAIR/{...}/
└── Testing_prospective/             # Test (paired)
    ├── T1W/{0.1T, 1.5T, 3T, 5T, 7T}
    ├── T2W/{...}/
    └── T2FLAIR/{...}/

Our naming convention is: "Type_contrast_fieldStrength_ID.nii.gz", where Type is either P (prospective) or R (retrospective). For example, P_T1W_0.1T_0006.nii.gz. The same ID corresponds to the same volunteer.

Retrospective data uses global sequential IDs from 0001 to 1056. Each field strength occupies a distinct ID range: 0.1T corresponds to 0001–0100 (100 volunteers), 1.5T to 0101–0328 (228 volunteers), 3T to 0329–0471 (143 volunteers), 5T to 0472–0737 (266 volunteers), and 7T to 0738–1056 (319 volunteers). Since each volunteer was scanned at only one field strength, data across different field strengths are unpaired.

Prospective data uses IDs from 0001 to 0040. The same ID across all field strengths corresponds to the same volunteer. For example, P_T1W_0.1T_0006.nii.gz and P_T1W_7T_0006.nii.gz are from the same person scanned at 0.1T and 7T respectively. This enables paired training and evaluation across field strengths.

For retrospective data (from different volunteers), we only performed rigid registration. If you are not satisfied with the cross-contrast alignment, you may perform your own registration. For prospective data (from the same volunteer), we applied deformable registration to achieve better spatial alignment across field strengths and contrasts. However, some degree of mismatch is unavoidable. If you identify any severe misalignment issues in the data, please contact us at: mrixfields@outlook.com.

---