Data - Second Round¶
General Remarks¶
The data used in this challenge was acquired in healthy control subjects that underwent MRI examinations and that were instructed first to lie still and then to degrade the scan quality by following predescribed motion patterns. That way, a ground truth (still) as well as motion degraded image was acquired for each subject. Motion estimates were measured with the markerless tracking system Tracoline (TracInnovations, Ballerup, Denmark).
The following parameters describe the acquisition:
- Task 1: Magnetic resonance imaging, axial T1-weighted Short-TI Inversion Recovery, no contrast. Flip Angle 150°. TR/TE/TI 3500/8.6/1387 ms. Voxel size 0.9x0.9x5.0 mm. GRAPPA (accel. Factor PE: 2, 27 ref. lines, equidistant subsampling). Partial Fourier Off.
- Task 2: Magnetic resonance imaging, axial T2-weighted Turbo Spin Echo, no contrast. Flip Angle 150°. TR/TE 4400/117 ms. Voxel size 0.4x0.4x5.0 mm. GRAPPA (accel. Factor PE: 2, 32 ref. lines, equidistant subsampling). Partial Fourier Off. Two averages (Acq 0: even lines, Acq 1: odd lines).
Exemplary images (without motion): 
Training Data¶
We encourage the participants to use the publicly available training data from the fastMRI challenge to pre-train their models. The fastMRI challenge provides 409 axial T1-weighted scans as well as 2524 axial T2-weighted scans at 3T. Between the opening of registration and release of the optimization / validation data, the participants have opportunity to apply for access at https://fastmri.med.nyu.edu.
Additional, publicly available data can be used for training the models. This should be indicated in the submission.
Optimization / Validation Data¶
10 validation subjects will be provided for optimization and validation of the pre-trained models. These consist of anonymized, degraded multi coil k-space data in HDF5 format as well as anonymized ground truth k-space data in HDF5 format and ground truth images in BIDS MR format. Motion estimates will be provided as a text file. See information regarding access below.
Test Data¶
For final evaluation 20 test subjects (without ground truth image) will be provided. This dataset only consists of anonymized, degraded k-space data in HDF5 format without ground truth images. The teams are expected to reconstruct the degraded raw data with their trained and optimized models and submit their estimates of the ground truth images. Only participants, who applied for access to the data by filling out the DUA below, will have access to the test data. For more details regarding Submission and Evaluation, please consult the corresponding pages.
Data access¶
All data pertaining to the RealNoiseMRI challenge is provided and hosted by the Neurobiology Research Unit, Rigshospitalet, Denmark. Interested scientists may apply for access to the RealNoiseMRI challenge dataset for the purposes of internal research or education only. Access is contingent on adherence to the RealNoiseMRI challenge dataset usage agreement shown below, which also outlines policies for publication and citation. Note, while the data are fully anonymized in order to fullfill GDPR guidelines for sharing, we are following the guidance provided by the Open Brain Consent and hence are additionally requiring you to download, fill out and return the RealNoiseMRI challenge dataset usage agreement to confirm your intent with and usage of the challenge data.
Please download the RealNoiseMRI challenge dataset usage agreement and return it in signed form to RealNoiseMRI@nru.dk. You will upon this receive a secure file transfer with the optimization / validation data.
Data details¶
The anonymization will follow exactly the same steps as the fastMRI challenge. In order to provide anonymized k-space data it will be provided in the vendor neutral HDF5 format. Furthermore, in order to remove facial features, parts of k-space will need to be removed. All processed k-spaces will then be reconstructed to images in DICOM format, loaded into a picture archival communication system (PACS) and visually checked by certified MR technologists to confirm exclusion of identifying facial features.
For loading HDF5 data e.g. the first part of the following code can be used: https://github.com/ismrmrd/ismrmrd-python-tools/blob/master/recon_ismrmrd_dataset.py.
The ground truth images will be provided in BIDS MR format. This is a completely standard procedure for any public MRI dataset [1] (e.g. all MRI datasets on OpenNeuro have been treated like this).
[1] Gorgolewski KJ et. al. The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Sci Data. 2016;3: 160044. PMID: 27326542.