Data Description and Format

Data Description and Format

Click here to download

Cardiac function data were collected from a CINE 2D breath-held fully sampled fully balanced steady state free precession (TrueFisp/bSSFP) sequence. These data were subsequently subsampled. All data will be on a Cartesian grid and include a small fully sampled region in the center of k-space. However, both 1D and 2D undersampling patterns should be expected. Coil sensitivities are not available.

ISMRM Raw Data Format (ISMRMRD)

Data is in the ISMRM Raw data format [link], with a small caveat. Data is considered Non-Cartesian even though the points are Cartesian. Thus the trajectory field is populated in the acquisition header and contains the matrix coordinates.

To help with the use of this format, we are providing several useful tools:
  1. Matlab: example reconstruction, ESPIRIT reconstruction, import tool, and output code for submission [example_recon_m.zip]
  2. C/C++: example reconstruction [example_recon_c.zip]
  3. Virtual machine: we have created a standard Virtual machine which includes the example reconstruction with all libraries setup. To use this machine you need VirtualBox [VirtualBox.org] and to import the machine an appliance. The root password is ismrm$ and instructions are on the Desktop. [ChallengeMachine.ova]

Raw Data Format:

While we highly encourage the use of the ISMRM raw data format, a RAW format (*.raw) will also be provided. The k-space data format is float32 with alternating real/imag, then Kx x Ky x COILS x TIME FRAMES. Matrix is full size, with missing k-space data replaced with zero-value.

Submission Format

Data must be coil combined before submission with intensity weighting equal to root sum-of-squares. That is, you can combine data with root sum of squares or derive coil sensitivities and correct for in-homogeneities and scaling by multiplication by the sum-of-squares of the sensitivity maps themselves. All data should be submitted in float32 row-format magnitude data. For example, a Nx x Ny x Nframes dataset should be stored: data[ frame*(Nx*Ny) + y*(Nx)+x] with a total of Nx*Ny*Nframes elements.

Reconstructions Guidelines

Any reconstruction is allowed that does not require input parameters beyond those detailing the input data. For example, your code should not require subject specific tuning of regularization parameters. Your data will be compared against a root sum of squares reconstruction performed via FFT of the truth data in Matlab. This implies that no scaling by matrix size within in the FFT. Example data (case 0), shows the expected output of your reconstruction. To enable comparison, your reconstruction must match the scaling and orientation of the reference reconstruction. In the first phase of the challenge, you are responsible for all reconstructions and use all available tools to reconstruct the data. However, if you succeed in making it to the second round, your reconstruction will be run by the organizing team. To enable this there will be several rules for phase 2:
  1. No input parameters allowed beyond those detailing the input data. You will not be allowed to specify regularization/tuning parameters on a per case basis.
  2. Non-Matlab and mexed matlab must run in a virtual machine. You will be responsible for setting up this machine so that we can run it. To make this easier, we are providing a basic virtual machine which has the example reconstruction [ChallengeMachine.ova].
  3. Matlab reconstructions calling only standard or self-contained libraries will be allowed. If you use a community library you will need to include it with submission as a sub-folder and setup linking within your reconstruction.
  4. Must run in <90minutes per dataset and use less than 8GB of ram. All reconstructions will be performed on relatively powerful machine (8 thread i7, 16 GB ram). Multi-thread programing is allowed for Matlab and virtual machines.