The CDMRI workshop was successfully held on September 18th 2014. There were mainly seven multi-shell estimation methods and six single-shell estimation methods participated in the challenge. The results are available in the attached file and are briefly summarized in below.
Muti-Shell Challenge:
The methods that participated the multi-shell challenge are listed as follows:
- Spherical Fourier-Bessel (SFB)
- Spherical Finite Rate of Innovation (SFRI)
- 3D-SHORE
(a) Positive EAP, Laplace-Beltrami
(b) Laplacian, Non-Local Means - MAP with Laplacian regularization
- Constrained Spherical Deconvolution (CSD)
(a) Non-Local Means (NLM)
(b) Non-Local Spatial and Angular Matching (NLSAM) - Sharpening Deconvolution Transform (SDT)
(a) Non-Local Means (NLM)
(b) Non-Local Spatial and Angular Matching (NLSAM) - Self-Adjusted (SA) basis
- Directional Radial Basis (DRB) function
- Spherical Ridgelets with Radial decay (DRR)
We have collected method-wise comparisons and self-comparisons for each method with different number of measurements. The main conclusions are as follows:
- Some of the methods (e.g. MAP) perform extremely well in terms of signal fitting (NMSE), however perform sub-optimally in terms of angular error !
- Other methods perform well in terms of angular error, but not so well in terms of signal fitting (RTOP) !
Single-Shell Challenge:
The methods that participated the single-shell challenge are:
- Spherical Finite Rate of Innovation (SFRI)
- Fiber Orientation Distribution (FOD) using non-negative sparse recovery.
- Constrained Spherical Deconvolution (CSD)
(a) Non-Local Means (NLM)
(b) Non-Local Spatial and Angular Matching (NLSAM) - Sharpening Deconvolution Transform (SDT)
(a) Non-Local Means (NLM)
(b) Non-Local Spatial and Angular Matching (NLSAM) - Self-Adjusted (SA) Basis
(a) Non-Local Means (NLM)
(b) Non-Local Spatial and Angular Matching (NLSAM) - Anonymous method
- Spherical Ridgelets (SR)
The conclusions for single-shell challenge are:
- If connectivity analysis is your only goal and you have time to acquire only 30 directions, then use SFRI method with a b-value of 3000 (about 10% error in signal fit).
- At b=2000, and 30 gradient directions, almost all methods do well (except Sharpening Deconvolution Transform-NLM).
- For b=2000, and 60 gradient directions, CSD methods do very well.
Now, the gold-standard data has been released! All participants could refine their algorithms from this data. Please do so and send us the updated results.