Data

Data sources

CMRx4DFlow2026 provides an extensive multi-center, multi-vendor k-space 4D Flow MRI dataset.

💾 Data Overview: Scale & Diversity

• Total Volume:

  • Over 400 cases (volunteers and patients), including more than 300 aortic 4D Flow MRI cases.

• Multi-channel k-space data from:

  • 10+ medical centers
  • 4+ vendors: GE, Philips, Siemens, United Imaging
  • 3 field strengths: 1.5T, 3T, 5T
  • 6+ anatomical regions: Aorta, heart, brain, liver, kidney, carotid arteries

📊 Acquisition Parameters

• Field of View (FOV): 200 × 200 × 40 mm³ to 450 × 450 × 90 mm³
• Spatial Resolution: 0.9 to 3.0 mm isotropic/anisotropic voxels
• Temporal Resolution: 23 to 120 ms across cardiac phases
• Velocity Encoding (VENC):
  • Liver/Kidney/Brain: 40–100 cm/s
  • Aorta/Carotid: 100–200 cm/s
• Sequence Type: ECG-gated 3D Cartesian 4D Flow MRI
  • Protocols include both retrospective gating and prospective gating

📊 Diseases

• Aortic stenosis
• Pulmonary atresia
• Tricuspid regurgitation
• Ventricular septal defect
• Pulmonary hypertension
• Chronic kidney disease
• Acute kidney injury
• ...

📊 Dataset Separation: Training, Validation, and Testing
The dataset is carefully divided to support a robust development and evaluation cycle:

Regular Task 1 & 2

• Training Set (138 Cases):

  • Fully sampled raw k-space data provided for each case. Serves as the "ground truth" input for deep learning models, allowing training from undersampled k-space to high-fidelity 4D Flow images.

• Validation Set (32 Cases):

  • Undersampled raw k-space data (simulated with acceleration factors from 10x to 50x) derived from fully sampled data, without corresponding ground truth.

• Test Set (43 Cases):

  • Final undisclosed dataset for challenge ranking. Consists of undersampled raw k-space data (with varying acceleration factors) and fully sampled references.

Special Task 1: Generalizability across new sites and diseases

• Validation Set (40 Cases):

  • Undersampled raw k-space data (simulated with acceleration factors from 10x to 50x) acquired from different centers than the training set and from patients with different diseases. No corresponding ground truth is provided.

• Test Set (60 Cases):

  • Final undisclosed dataset for special task ranking. Undersampled raw k-space data (with acceleration factors from 10x to 50x) from different centers and patients with different diseases; fully sampled references are withheld.

Special Task 2: Generalizability across different anatomical regions

• Validation Set (40 Cases):

  • Undersampled raw k-space data (simulated with acceleration factors from 10x to 50x) from multiple anatomical regions:
    • Cerebrovascular (10 cases)
    • Portal Vein (10 cases)
    • Renal Artery (10 cases)
    • Carotid (10 cases)
  • No corresponding ground truth is provided.

• Test Set (80 Cases):

  • Final undisclosed dataset for special task ranking. Undersampled raw k-space data (with acceleration factors from 10x to 50x) from the same four anatomical regions:
    • Cerebrovascular (20 cases)
    • Portal Vein (20 cases)
    • Renal Artery (20 cases)
    • Carotid (20 cases)
  • Fully sampled references are withheld for evaluation.

---

Pre-processing

The raw k-space data exported from the scanner will be processed and transformed to '.mat' format using the script provided by each vendor. A readme file will be provided to describe the content and usage of the data.

Details of data description and different undersampling masks:

Filename	Dimension	Description
kdata_full.mat	(Nv, Nt, Nc, SPE, PE, FE)	Retrospectively fully sampled multi-channel k-space data
segmask.mat	(SPE, PE, FE)	Segmentation mask for the region of interest
coilmap.mat	(Nc, SPE, PE, FE)	Coil sensitivity maps
usmask_ktGaussian{R}.mat	(1, Nt, 1, SPE, PE, 1)	Gaussian k-t undersampling mask for $R\times$ acceleration
params.csv	N/A	Acquisition and reconstruction metadata

Note: The Cartesian trajectory is used for k-space data sampling.
The readout direction (FE) is fully sampled, and undersampling occurs in the two phase-encoding directions (PE and SPE).
# represents varied acceleration factors (10x, 20x, 30x, 40x, 50x). ACS lines are not included for calculations.
If the data has no temporal dimension, the corresponding mask dimension becomes (PE, SPE).

Dimension Reference: Both fully sampled and undersampled data share the standardized dimension order (Nv, Nt, Nc, SPE, PE, FE)

Dimension definitions:

• FE (Frequency Encoding): Image-space X axis (readout/frequency-encoding direction) sample count; affects FOV and resolution along X.
• PE (Phase Encoding): Image-space Y axis (phase-encoding direction) sample/step count; affects FOV and resolution along Y.
• SPE (Slice Phase Encoding): Image-space Z axis (slice phase-encoding / second phase-encoding direction); affects volumetric coverage and resolution along Z.
• Nc (number of coils): Number of receiver coil channels; each channel provides independent complex-valued measurements for SNR improvement and parallel imaging.
• Nt (number of cardiac phases): Number of time frames in the cardiac cycle; each frame corresponds to a specific cardiac phase.
• Nv (number of velocity encodings): Number of velocity-encoding conditions; used to compute velocity-induced phase and estimate velocity vectors.

Further background on 4D flow MRI:

• 4D Flow MRI Overview
• CMRxRecon 4D Flow Data Overview

---

Data Directory Structure

ChallengeData
├─ Task
│  ├─ TrainSet
│  │  └─ Aorta
│  │     └─ Center001
│  │        └─ UIH_30T_uMR890
│  │           └─ P007
│  │              ├─ kdata_full.mat
│  │              ├─ coilmap.mat
│  │              ├─ segmask.mat
│  │              └─ params.csv
│  └─ ValidationSet
│     └─ Aorta
│        └─ Center002
│           └─ Siemens_30T_VIDA
│              └─ P007
│                 ├─ kdata_ktGaussian10.mat
│                 ├─ usmask_ktGaussian10.mat
│                 ├─ coilmap.mat
│                 ├─ segmask.mat
│                 └─ params.csv

---