It is possible to analyze DWI data with different b-values per direction?

1

Dear experts,

I downloaded a public dataset, in which I found the DWI data had an unusual acquisition for me. Specifically, the b-value is like 0 50 250 450 650 850 100 300 500 700 900 100 300 500 700 900 150 350 550 750 950 200 400 600 800 1000. What I am familiar with is a fixed b-value with many directions (for instance, b-value=1000 with 60 directions). I know this kind of data could not be handled with FSL, which I usually used for DWI processing. My question is that it is possible to analyze this kind of data using other software packages to quantify microstructure or connectivity? If not, what purpose this kind of data is designed for?

Many thanks!

2

Hi @younghoo,

It might be a cartesian grid sampling scheme, which might be more suited towards generalized q-space reconstruction methods, such as implemented in DSI-Studio. You can use qsiprep to pre-and-post-process your data (using SHOREline for head motion correction and one of the dsi_studio reconstruction specifications).

Best,
Steven

3

This looks a bit like it could be an IVIM acquisition. These kinds of acquisitions sometimes have only one direction collected in all of these different b-values, which would pose problems with many preprocessing/reconstruction pipelines. Maybe you can tell us where you found this data? Or at least how many different directions of measurement there are?

4

This dataset is from the OASIS project.
The content of the b-value file:

0 50 250 450 650 850 100 300 500 700 900 100 300 500 700 900 150 350 550 750 950 200 400 600 800 1000

The content of the b-vector file:

0 0.999376 0.93632 0.935713 0.807961 0.578035 0.584476 0.577328 0.497473 0.355648 0.354268 0.321518 0.319299 0.314335 0.447214 0.57735 0.707107 0.707107 0.308046 0.500144 0.500466 0.501255 0.578775 0.808221 0.810142 0.809593
0 0.0249844 0.00450154 0.00489902 0.528953 0.580107 -0.584476 -0.583536 -0.85281 0.934611 0.935139 0.857382 0.0103 0.0062867 0 -0.57735 0 0 -0.854391 -0.529252 0.529782 0.852133 0.583127 -0.529167 -0.00310995 -0.00248723
0 0.0249844 -0.35112 0.352729 0.259632 -0.573891 -0.562829 0.57112 0.158857 0.00413545 0.0031916 0.401898 -0.947598 0.949291 -0.894427 0.57735 0.707107 0.707107 -0.418477 0.685382 -0.684738 -0.150376 0.570072 -0.258382 0.586226 -0.586986
5

I made a little search and it seems that the difference between grid sampling scheme and single-shell/multi-shell sampling scheme is that, for the former, the b-value is more diverse and the number of directions for each b-value is different. Therefore the current dataset I am dealing with seems not acquired using a grid sampling scheme, because the number of directions for each b-value is one (no repetition) and the b-value is generally low (although uniquely low).

6

You could certainly analyze these data using many different processing pipelines and models (e.g., DTI), but I’d caution some care, especially with the lower b-values (particularly b<200), which may be sensitive not only to the effects of diffusion, but also to the effects of perfusion.

7

Thank you for you reply. I had two more questions:

  1. For DTI modeling, the different b-values would be a problem (besides the perfusion effect)? As I only processed shell-like DWI data using FSL, and when fitting DTI model, I was told to use only the data with one b-value. For instance, when there two b-values like 1000 and 2000, use the b=1000 data for DTI fitting.
  2. The current dataset is suitable for tractography and connectivity analysis? For shell-like DWI data, I was told that at least 20 directions were needed using FSL’s bedpostx. For the current dataset, the b-value is generally low and each b-value corresponds to a different direction. I had a very very brief understanding of the physical and mathematical principles underlying DWI data modeling, but I had a feeling that this kind of data may have a bias related to anisotropy, because in each direction the sampling is different. This concern for anisotropy also relates to DTI modeling.
8
  1. In principle, it should be possible to fit the model, so long as you have at least 7 diffusion-weighted measurements conducted in different directions. In practice, odd things might happen if your measurements have some idiosyncratic combination of b-values and directions (for example, all the high b-values are close to each other in q-space and all the low b-values are close to each other and far away from the high b-values). The reason not to use b=2,000 for DTI may have to do more with the fact that the violation of the assumptions of the model are more apparent at higher b-values (where the signal is more likely to reach close to the noise floor).
  2. Maybe? I am not sure where that number 20 comes from for bedpostx, but this Derek Jones paper is a good starting point for thinking about measurement requirements for various diffusion MRI analysis approaches, and suggests at least 30 directions would be needed to get reliable estimates of the principal diffusion direction of the tensor. I’d say whether you can use this data for a particular analysis should depend not only on the number of directions used, but also on the SNR and sensitivity to various tissue configurations (which relates to the b-value), so ymmv.
1 Like
9

So grateful for your kind explainations. As for the number 20, my memory goes wrong and the lower limit is 25 suggested by the FSL wiki page.