Optimal b-vals and number of gradient directions for diffusion data for DTI, DKI, NODDI

[josh]

Questions

1. What are the optimal b-values and respective number of gradient directions to use for a sequence enabling DTI, DKI, and NODDI estimation?

2. Is it preferred to acquire:
   a) all gradient directions in one phase-encoding direction and only a few b=0 in the opposite phase-encoding direction; or
   b) split the gradient directions over two acquisitions with opposing phase-encoding directions (such that you have different but complementary gradient directions)?

Additional Explanations

For Question 1:

For DTI, we usually use a sequence with 64 directions at a b-value of 1000 s/mm². For a planned study, we aim to extend this sequence to also enable DKI and NODDI estimation. We are considering adding 96 directions with a b-value of 2000 s/mm² (for research sequence). For clinical scans, we are planning to use 32 dir at b=1000 and 48 at b=2000. However, the literature shows significant variation in both b-values and the number of directions used.

• DSI Studio recommends 30 directions at b=1500 s/mm² and 60 at b=3000 s/mm², and this was linked in a previous neurostars post (source). However, having only 30 directions at 1500 is rather low for DTI, and not comparable to our previous DTI’s using b=1000.
• UK Biobank: Uses 50 directions each for b-values of 1000 and 2000 s/mm² (source).
• ADNI: Uses b-values of 500, 1000, and 2000 s/mm² with a total of 112–126 directions, distributed as ~6–20, ~48, and ~60, respectively.
• Other Studies: Some use higher b-values up to 3000 s/mm². However, one study demonstrated reliable DTI and NODDI estimates with as few as 4 directions at b=700 s/mm², 14 at b=1000 s/mm², and 32 at b=2000 s/mm² (source).

Therefore, I am uncertain whether a b-value of 3000 s/mm² would be more appropriate than 2000 s/mm², and if a reduced number of gradient directions would still yield reliable results.

For Question 2:

The Tortoise dMRI group strongly recommends acquiring diffusion data with blip-up/blip-down phase-encoding and, for studies sensitive to gradient direction sampling resolution, using different but complementary diffusion gradient directions in each acquisition (source). However, many published studies only acquire a few b=0 images in the opposite phase-encoding direction. Therefore, I am unsure which acquisition scheme would be optimal.

[Steven]

HI @josh and welcome to neurostars!

@josh

• What are the optimal b-values and respective number of gradient directions to use for a sequence enabling DTI, DKI, and NODDI estimation?

The optimal number would be as much as you can feasibly collect; more data is always better. But you may be interested in reading this paper that details minimal number of recommended directions: Shibboleth Authentication Request

@josh

Is it preferred to acquire:
a) all gradient directions in one phase-encoding direction and only a few b=0 in the opposite phase-encoding direction; or
b) split the gradient directions over two acquisitions with opposing phase-encoding directions (such that you have different but complementary gradient directions)?

Typically the latter.

@josh

However, having only 30 directions at 1500 is rather low for DTI, and not comparable to our previous DTI’s using b=1000.

30 directions is more than enough directions for a simple DTI fit, even if you only use the b=1500 shell.

@josh

Therefore, I am uncertain whether a b-value of 3000 s/mm² would be more appropriate than 2000 s/mm²

Ideally you would collect both (more b=3000 than b=2000 due to lower SNR at higher b-val coverage and more space between the high-bval gradients in q-space). Some folks see improved tractography or fixel-based analyses when using high b-val data.

@josh

The Tortoise dMRI group strongly recommends acquiring diffusion data with blip-up/blip-down phase-encoding and, for studies sensitive to gradient direction sampling resolution, using different but complementary diffusion gradient directions in each acquisition (source). However, many published studies only acquire a few b=0 images in the opposite phase-encoding direction. Therefore, I am unsure which acquisition scheme would be optimal.

The “few b=0 images in the opposite phase-encoding direction” is usually done because it is less time to collect and generally produces usable outputs. But it is generally usually preferred to correct two entire sequences with opposite PE dir.

In summary, more data is always better, and it might help to instead know what kind of time constraint you are under, and what planned analyses you have that might guide acquisition protocol.

Best,
Steven