Challenge

The goal of Sparse Reconstruction Challenge for Diffusion MRI (SPARC dMRI), is to test the ability of various algorithms to faithfully recover the diffusion data from noisy measurements, similar to that in the in-vivo case. The challenge will employ several acquisitions of a physical phantom data set, with different number of gradient directions and b-values. The phantom was constructed and extensive acquisitions were made by our collaborator, Dr. Frederik Laun, from the German Cancer Research Institute. 

The Challenge: A test data set was acquired with the following parameters: b={1000,2000,3000} s/mm2 with K={20,30,60} gradient directions per shell. This dataset is available in 3 different formats: 1. Nifti, 2. NRRD, 3. matlab (.mat). The challenge consists of two parts: 

  1. Estimation of the diffusion signal over multiple q-values (b-value shells).
  2. Estimation of the Orientation Distribution function (ODF) or fiber ODF from a single shell b-value data.

Participants can choose to participate in any one of the above challenges.

For Challenge #1, participants will be required to reconstruct the diffusion data in q-space (up to a b-value of 5000) and report the following measurements:

  1. Number of fibers at each voxel:  the “total number of fiber bundles” or “the number of peaks” in the ODF.
  2. If more than one fiber bundle, the angle between the fiber bundles
  3. The estimated signal at 405 unique q-space points (we will provide these locations in q-space and the format to report each of these as a .txt file).
  4. rendered image (snapshot) of the ODF's estimated from the data at each voxel, to enable visual comparison across methods at the workshop.

For Challenge #2, participants are required to report the following:

  1. Number of fiber bundles at each voxel, e.g. “the number of peaks” in the ODF. If more than one fiber bundle, the angle between the fiber bundles, (in degrees with four digits after the decimal point). These results need to be tabulated as a matrix (4 x 208) in a txt file.
  2. The estimated ODF (or fODF) at 81 different q-space points (with four digits after the decimal point). The results need to be tabulated as a matrix of size 81 x 208 in a txt file format  (see Data section for more details).
  3. A single rendered image (snapshot) of the ODF's estimated from the data at each voxel, to enable visual comparison across methods at the workshop. This should be one image showing all voxels (13x16).

Additionally, the participants will be required to submit a 1-page document that provides details about the model they used and the parameter settings that were employed.

Submission and deadline: Each submission will consist of a  one-page document along with a set of other files, zipped into a single .zip file. The submission deadline is June 27th June 29th, 2014, 6 PM, EST. All results must be uploaded by this time in the format described here.

Choice of approach: The participants can choose to analyze all of the challenge dataset or the shell(s) that are most useful for their computational method. Thus they may choose whichever sampling scheme they would like, and mix and match the number of gradient directions for each b-value shell to (e.g. for b=1000, K=20, and b=3000, K=60, i.e. they can mix and match the number of gradient directions for each b-value shell). However, the participants must report the results in the provided format. We will also provide a Matlab script that they can use to test whether the format is correct or not. Each participant is welcome to report several results using several combinations of the data, but they will need to provide a submission for each of those.

The error metrics: We will compute the following error metrics by comparison to the “gold standard” dataset:

  1. Error in estimation of the number of fibers.
  2. Error in estimation of angle.
  3. Error in estimation of the signal for each b-value shell b={1000,2000,3000,4000,5000}. (Normalized mean squared error)
  4. Error in estimation of the sum of the signal (i.e. a discrete approximation of the return-to-origin probability).

Gold standard dataset: The “gold standard” data set was acquired with the following parameters: 5 b-values of b={1000,2000,3000,4000,5000}, each with K= 81 gradient directions (multiple shells). 10 repetitions of this data were acquired, which we will average to get create a “gold standard” multishell dataset. Of-course, this will still have some noise, but it will be significantly reduced. In addition, we know the crossing angles in the physical phantom, so we can compute the angular error.

For help, contact: http://projects.iq.harvard.edu/sparcdmri/contact