T2*-MOVE: An Open k-Space Dataset to Address Motion in T2* Quantification from Gradient Echo MRI

Dataset to reproduce "Motion-Robust T2* Quantification from Low-Resolution Gradient Echo Brain MRI with Physics-Informed Deep Learning", published in Magnetic Resonance in Medicine (https://doi.org/10.1002/mrm.70050) [1]

Brain magnetic resonance imaging (MRI) is inherently sensitive to patient motion, which can degrade image quality and hinder clinical diagnoses, making motion correction (MoCo) essential for reliable measurements [2].  Paired datasets that include scans both with and without motion artifacts are relevant for the development and validation of new MoCo algorithms [3]. To the best of our knowledge, only one publicly available brain MRI dataset exists so far that contains paired motion-free and motion-corrupted images and includes the corresponding raw k-space data [4].

Our dataset, T2*-MOVE, addresses this gap by providing an open k-space dataset with paired motion acquisitions, specifically designed to develop MoCo methods for T2* quantification from gradient echo MRI, which is especially vulnerable to patient motion due to its sensitivity to motion-related changes in B₀ inhomogeneities [5].

Participants

22 healthy volunteers (10 females, 12 males) aged 21 to 34 years (mean: 26.6 ± 2.7 years) were recruited to participate in this study between May 2023 and October 2024. The study was approved by the local ethics committee (approval numbers 440/18 S-AS and 2023-386-S-SB) and written informed consent was obtained from all participants before the scanning.

To enable additional simulation of motion artefacts in motion-free data with real motion curves, we extracted motion data from in-house fMRI time series. These fMRI data originate from two ongoing, unpublished studies (approval numbers 472/16 S and 15/20 S), involving independent cohorts from the above imaging cohorts.

Data acquisition

All data were acquired on a 3T Philips Elition X MR scanner (Philips Healthcare, Best, The Netherlands) with a 32-channel head coil. The imaging protocol consisted of a structural 3D MPRAGE sequence and several repeated acquisitions of an interleaved multi-slice 2D GRE sequence, which acquires even and odd slices in two packages. We additionally acquired half- and quarter-resolution GRE data which are needed for the state-of-the-art MoCo method [6]. The relevant sequence parameters are summarized in the following table:

*: Values are identical for full, half and quarter-resolution acquisitions.

Instructions for repeated GRE acquisitions: 13 of the 22 volunteers performed repeated scans under two motion conditions, with the instruction to (1) remain still, and (2) to move randomly throughout the acquisition (e.g. imitating sneezing or coughing). As a reference point for positioning their head inside the scanner bore, we attached an earplug to the head coil using tape. Two subsets of volunteers received different instructions:

• Random motion timing: 11 volunteers were instructed as follows: ”For the next six minutes, randomly move from time to time in a controlled, natural way - as if you were sneezing, coughing or swallowing - without exaggerating the movement.”
• Instructed motion timing: 8 volunteers received precise timings on when to move and when to lie still during the motion scan. They were instructed as follows: ”For the next scan, between the START and STOP commands, move in a controlled, natural way - as if you were sneezing, coughing or swallowing - without exaggerating the movement.”

To explore different motion types, one of the instructed motion timing subjects (sub-10) was asked to perform four motion patterns with three different amplitudes each, resulting in twelve different motion-corrupted acquisitions.

Exaggerated motion: Three of the “instructed motion timing” subjects were excluded from the main analysis in [1], due to exaggerated motion in the k-space center (ten seconds, corresponding to approximately nine central k-space lines). Their data is provided here in a separate subset (test_extreme_moco).

Data Processing

• MRI data: The raw k-space data were exported from the scanner with the Pack’n Go tool (GyroTools LLC, Zurich, Switzerland) and converted to “.mat” format using MRecon (GyroTools LLC, Zurich, Switzerland) in MATLAB. The “.mat” files were subsequently converted to a fastMRI-like format (“.h5”, see description below) in Python. Coil-combined images were reconstructed in Python (example code available, link below) and saved as separate “.h5” files. The corresponding anatomical T1-weighted MPRAGE images were segmented and registered to the T2*-weighted data with SPM12 in MATLAB. The motion timing instructions for the “instructed motion timing” subjects were converted to reference motion timing k-space masks using the known acquisition scheme.
Anonymization: The T1-weighted image data were anonymized by defacing with ‘spm_deface’ (SPM12) in MATLAB. The axial, 2D-encoded raw data (T2*-weighted acquisitions) were anonymized by replacing all slices in which the eyeball is visible (and inferior slices) with zeros, similar to how the de-identification was handled for the fastMRI dataset [7]. The defacing was checked manually for all subjects by reconstruction and subsequent visual inspection the de-faced raw data. The headers were anonymized by removing all sensitive information during data export from the scanner, which was manually checked.

• Motion data: 94 motion curves were extracted from 62 functional MRI (fMRI) time series. The motion curves contain six degrees of freedom (three translational and three rotational parameters in mm and °) for 150 time points, corresponding to a length 223.5s in steps of 1.5s. In the publication, principal component analysis (PCA) was performed to generate a larger set of training curves (see description in [1]). The PCA components as well as the generated curves are provided in the subfolder “PCA_Train_Curves”.

Data Format

• MRI data – for each subject in each subset (“train_recon”, “val_recon”, “val_moco”, “test_moco”, “test_extreme_moco”):

The raw k-space data is available as .h5 file with the following structure:

└── ‘mrecon_header’                           (Parameters extracted from Mrecon)

└── ‘ismrmrd _header’                        (XML header in ISMRM-RD format)

└── ‘kspace’                                              (multi-coil data, reformatted into   
                                                                                         'complex64’)

└── ‘sens_maps’                                     (sensitivity maps, reformatted into
                                                                                         'complex64’)

• Motion data:

Code Availability

Demo code how to load the data is available at: https://github.com/compai-lab/2025-nefeli-eichhorn.

Code to reproduce the results of the paper "Motion-Robust T2* Quantification from Low-Resolution Gradient Echo Brain MRI with Physics-Informed Deep Learning" is available here: https://github.com/compai-lab/2025-mrm-eichhorn. Note that this code is not adapted to the data format of this data record. Data loading functions need to be substituted.

[1] Eichhorn, H. et al. (2025). Motion-Robust T2* Quantification from Low-Resolution Gradient Echo Brain MRI with Physics-Informed Deep Learning. Magnetic Resonance in Medicine.
[2] Godenschweger, F. et al. (2016). Motion correction in MRI of the brain. Physics in Medicine and Biology, 61(5).
[3] Spieker, V. and Eichhorn, H. et al. (2024). Deep Learning for Retrospective Motion Correction in MRI: A Comprehensive Review. IEEE Transactions on Medical
Imaging, 43(2). 
[4] Skak Madsen, K. et al. (2025). Markerless Prospective Motion Correction Data [dataset]. In PublicNeuro Datasets. Neurobiology Research Unit, Copenhagen University Hospital Rigshospitalet.
[5] Magerkurth, J. et al. (2011). Quantitative T2* -mapping based on multi-slice multiple gradient echo flash imaging: Retrospective correction for subject motion effects: Movement Correction in T2* Mapping. Magnetic Resonance in  Medicine, 66(4).
[6] Nöth, U. et al. (2014). An improved method for retrospective motion correction in quantitative T2* mapping. NeuroImage, 92.
[7] Zbontar, J. et al. (2018). FastMRI: An open dataset and benchmarks for accelerated MRI. arXiv:1811.08839.