Recommendations for the HCP Young Adult dataset

[Christian Ewert]

Dear all,

I would like to obtain NODDI parameters from the HCP Young Adult multi-shell dataset. For that, I am looking for recommendations regarding the following questions:

1) Biasfield Correction: Is it recommended to run a spatial biasfield-correction, e.g., with SPM, prior to the NODDI computation? Due to the normalization in the voxel-wise NODDI model, this should not make any difference, but when I run NODDI on a slice with and without biasfield correction, some (maybe 0.5%) voxels differ while the vast majority are the same (as expected).

2) b-value clustering: Should the b-values be "clustered"? The distributed b-values deviate around the intended shell, i.e., 995, 1005, 990 instead of 1000, 1000, 1000, both for b=0 and b>0. In response to an older post, it was recommended to set close-to-zero b-values to zero. In fact, not doing this causes an error. But should this also be done for the other b-values, i.e., 1000, 2000, 3000? Similarly to 1), running NODDI with and without this clustering gives the same results for almost all voxels, with a few of them being different.

3) Gradient nonlinearities: In the older post, it was recommended to incorporate the correction of gradient nonlinearities with respect to the b-vectors and b-values (via the grad_dev.nii file), causing them to differ across voxels. I assume that strategy is still recommended. Is there a native way to supply the gradient nonlinearity file to the NODDI toolbox for processing, or do I have to make changes to the code to incorporate this correction (like suggested in the older post)?

Thank you so much for your recommendations!
Christian Ewert

[Christian Ewert]

Point 2) can probably be removed from the list of open questions. Since gradient nonlinearity correction was recommended, the resulting b-values are naturally scattered around the target values. This implies that the number of shells is much greater than three—in fact, with gradient nonlinearity correction applied, we have 270 unique b-values. I initially raised this because FSL2Protocol.m calculates the unique b-values; however, for the default PGSE sequence, this information does not appear to be used elsewhere in the code. I assume that scattered b-values (except zero) do not introduce instability into model fitting, and I will proceed without bias field correction (1), without b-value clustering (2), and with gradient nonlinearity correction applied (3).