K2S Challenge Datasets

Access Training Data

WE ARE NOT SHARING ANY DATA AFTER THE CHALLENGE HAS BEEN CONCLUDED

Training and test case characteristics

• Training cases have undersampled knee MRI k-space data of 3D CUBE sequences, corresponding segmentation of cartilage and bone, as well as fully-sampled images as auxiliary data. Test cases will have only the undersampled knee MRI k-space data of 3D CUBE sequences. No additional data will be provided at test time.

• 300 training and validation cases (participants are encouraged to create their own validation splits from the training data).
  
  
• 50 test cases.

Data sources

All data was acquired from clinical scans on a GE Signa MR750 3T MRI scanner using an 18-channel knee T/R coil array. Acquisition parameters were as follows: FOV=15 x 15 cm2; 0.6 mm slice thickness, in-plane acquisition matrix=256×256; 200 slices, ±62.5kHz readout bandwidth; TR=1002ms; TE=29ms; acceleration by a factor of 4 (ky/kz=2/2); 2 ARC-reconstructed multi-coil full k-space data k-space and coil combined images were obtained and are provided from these acquisitions.

Data pre-processing methods

All data is processed with GE Orchestra Python SDK and used to apply ARC reconstruction to reconstruct from under-sampled k-space into full k-space data. This ARC-reconstructed k-space is subsequently undersampled by a factor of 8 to yield 8x under-sampled acquisition rather than additionally under-sampling the true k-space (not reconstructed by ARC) 2x to yield 8x under-sampling.  Our custom 8x center-weighted Poisson undersampling pattern in which the central 5% square in ky/kz is fully sampled (50x50 pixels) to allow for sensitivity-map estimation and not bias submissions towards or against any particular approach.

Ethics approval

Data acquired as per the institutional review board (IRB) was anonymized with no patient identifiers.

Data usage agreement

 Creative Commons Attribution-NonCommercial-NoDerivatives (CC BY NC ND).