Data

Three sets of test data are provided for this challenge, which are acquired based on a single slice of a physical phantom with dimension of 13 x 16. The diffusion-weighted signal was acquired with the following parameters: b={1000,2000,3000} and K={20,30,60} gradient directions per shell for the three test sets respectively. Since the data sets also include one measurement with b=0, the size of the three provided data sets are 13 x 16 x N with N = 61, 91, 181, respectively.

Download: The test data sets are provided in three different formats: 1. Nifti, 2. NRRD, 3. Matlab (.mat). The data sets can be downloaded via the links in the following table:

 

20 Gradients

30 Gradients

60 Gradients

Nifti

Gradient_20_Nifti

 

 Gradient_30_Nifti

Gradient_60_Nifti

NRRD

Gradient_20_Nrrd

 

Gradient_30_Nrrd

 Gradient_60_Nrrd

Matlab

Gradient_20_Matlab

 

 Gradient_30_Matlab

 Gradient_60_Matlab

For more details, the content of each zip file for the three formats is explained as follows:

  1. For the Nifti format, there are four files in total for each acquisition. The two nii files are respectively for the signal and the mask. The bvalue.txt file is used to store the scalar b-values and bvec.txt is for the gradient directions along which the samples are obtained.
  2. For the NRRD format, there are four files in total for each acquisition. The mask and the acquired data each have a header (nhdr) file and a corresponding raw.gz (data) file. The Gradient_K.nhdr header file has information for b-value and gradient directions K=20, 30, 60 while the corresponding raw.gz file contains the signal. 
  3. For Matlab format, there are two files in total for each acquisition. The Gradient_K.mat with K=20, 30, 60 contains three variables which are respectively the diffusion weighted signal (named Signal), scalar b-values (named bvalue) and the gradient directions (named bvec). The mask.mat file contains the mask.

Reporting your Results:

Mask for fiber bundles:  The mask for the 13 x 16 voxels is illustrated as in the following figure:

 

The voxels colored in black are known to have isotropic diffusion. Hence, the number of fiber bundles should be set to 0 in these voxels. Note however, that you still need to report the estimated signal in these voxels. 

For each voxel colored in white, the number of fiber bundles is known to take the values of {0, 1, 2, 3}. For these voxels, the participants are required to provide the estimated number of fibers, and the angle (in degrees) between all the fiber bundles if more than one fiber bundle is detected. The estimation result needs to be tabulated in a format as illustrated in the following table (Note the ordering of the voxels). So, these results need to be reported in a matrix of size 4 x 208 (rows x column). This matrix should be stored as a sequence of numbers in a text file (4 rows of 208 numbers per row, with newlines at the end of each row). The table shown below gives the correct ordering of the data in the text file.

We note that the values in the above table are only used to illustrate the format of the result to be reported. For voxels that have less than two fiber bundles, which include the voxels colored in black in the mask figure, the angles are reported as a three-dimensional zero-vector. For voxels with number of fiber bundles equals to two, the angle vector has only one non-zero element which is the estimated angle (in degrees) between the two fiber bundles. The value of the angle needs to be rounded to four digits after the decimal points. For voxels whose number of fiber bundles equals to three, the angles need to be estimated between every pair of the fiber bundles.

In conjunction with the angular information, a rendered image (snapshot) of the orientation diffusion functions (ODF’s) estimated from the data at each voxel is also needed for visual comparisons.

Signal estimation:  For each voxel, the participants are also required to provide the estimated normalized diffusion signal at 405 different q-space points. As usual, the normalized diffusion signal is defined as E(q)=S(q)/S(0) with S(q), S(0)  being the diffusion-weighted signal at q and its value at the origin (q=0) respectively. The q-value relates to the b-value as 

The coordinates for the 405 points in q-space can be downloaded in a txt file via the following link: q vectors for the estimated signal.

The participants are also required to submit the a text file containing the estimated signal in all the 13 x 16 =208 voxels. For each voxel, the signal should be estimated at 405 points in q-space. The signal needs to be estimated for both the voxels colored in white and the voxels colored in black in the mask figure. The expected 405x208 matrix structure of the estimated data is illustrated in the following table (Note the ordering of the voxels):

 

The estimated normalized signal needs to have an accuracy of four digits after the decimal points. The estimated results will be compared with a “gold-standard” data set.

Submission: To summarize, the participants need to submit the following measurements:

  1. Number of fiber bundles in 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 as illustrated in Table 1.
  2. The estimated signal at 405 different q-space points (with four digits after the decimal point). The results need to be tabulated as a matrix of size 405 x 208 in a txt file format as illustrated in Table 2.
  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).

These results, which are documented in three separate files, need to be compressed in a single zip file before submission. We provide a Matlab file that can be used to test the data format in the txt files. 


Single b-value shell challenge:

The provided data sets can also be used in single -shell estimation methods. Take the 20-gradient data set for example. At each voxel, the provided data set has 61 measurements with the first measurement being the signal at b=0. The signals on the shell with b=1000 is given by the 2nd to the 21st measurement while the signal on the shell with b=2000 is given by the 22nd to the 41st measurements.

For participants who use single-shell estimation techniques, the signal needs to be estimated on the corresponding b-value shell along 81 gradient directions that can be downloaded via the link: gradient directions for estimated signal. The number of fiber bundles and the angle between fiber bundles also need to be estimated and documented in the same format as illustrated in Table 1. The participants also need to explicitly provide the b-value of the measurement in the submitted results.

To summarize, the participants need to submit the following measurements:

  1. Number of fiber bundles in 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 as illustrated in Table 1.
  2. The estimated signal on a single b-value shell along 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 as illustrated in Table 2.
  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).
  4. The b-value of the measurements should be clearly indicated in the file name of the three files that document the estimation results. Each file should be named in the format “ABC_bxxxx” where “ABC” represents the participant-chosen name and “xxxx” represents the b-value. For example, the result for the number of fiber bundles could be named as “FiberBoundles_b2000.txt” if the measurements are obtained on the shell with b=2000.  

 

These results, which are documented in three separate files, need to be compressed in a single zip file before submission. We provide a Matlab file that can be used to test the data format in the txt files. 

 

 

Gold-Standard Data:

The gold-standard data set used to compare the estimation methods is the average of 10 acquisition of the same physical phantom with b={1000, 2000, 3000, 4000, 5000} on 81 gradient directions on each shell. The data set has dimension 13 x 16 x 406. We also provide a mask file to indicate the number of fiber bundles in each voxel. The voxels with 1 or 2 fiber bundles are masked by 1 or 2, respectively. In two-fiber voxels, the fiber bundles have a crossing angle of 45 degree. The voxels that are masked by 3 are not considered in this challenged.

The gold-standard data sets are available to download in three formats: 1. Nifti, 2. Nrrd, 3. Matlab (.mat). These can be downloaded via the links in the following table:

GoldStandard_Nifti

GoldStandard_Nrrd

GoldStandard_Matlab

All participants could use the gold-standard data to refine the estimation algorithms. Please do so and submit the updated results. For completeness, all the methods should be compared based on the same data sets. In particular, the participant should prepare the results considering the following requirements:

  • Multi-Shell Challenge: All methods should provide results for all three data sets, i.e. the data sets with a total number of 60, 90, 180 gradient directions, respectively. Extra experiments based on sub-sampling and combinations of the measurements will not be compared with other methods. All results should be submitted a single zip file. 
  • Single-Shell Challenge: All methods should provide results using measurements on the b-shell for b=2000 with 20, 30 and 60 measurements, respectively. Extra results using a different b-value or a different number of measurements will not be compared with other methods. Each method should submit a zip file that contains all three sets of results. 
  • Visualizing ODF: To unify the format for visualizing ODF in future publication, the participants should also submit a text file that contains the value of the estimated ODF on a given set of 642 gradient directions on the sphere. The results should be documented in a table of size 642x208 in similar format as in Table 2. We note that the participants should also submit their own snapshot of ODF's. 

To summarize, each method should submit a zip file that contains three folders for the three sets of measurements (for multi-shell challenge and single-shell challenge). Each folder should contain four files that are for the number of fiber-bundles, the estimated signal, the rendered image of ODF's and the estimated ODF's, respectively. The first three files should follow the same formats as in the first submission. The file of the estimated ODF's has a similar format as in Table 2.

Acknowledgement: Fuding support by the Deutsche Forschungsgemeinschaft (grant no. LA 2804/1-3) is gratefully acknowledged.