In our lab, we are looking into optimizing our dwi acquisition to be able to perform FBA in ex vivo rat brain data.
We have tried a couple of different approaches, but are unsure how to decide which approach is optimal for what we want to achieve.
As mentioned, we want to perform FBA (to determine e.g. Apparent Fiber Density), since we are interested in WM tracts in the midbrain and brain stem (which contain many crossing fibres).
We have the possibility for scanning time up to 36 hours (over the weekend) on a 9.4T system. We are using a multishot EPI sequence.
Regarding voxel size, we used to acquire our data with voxels of 150um isotropic, but have been playing around with smaller voxel sizes even (62.5um isotropic).
Regarding b-values, since we used 600, 1200, 1800 and 2400 in vivo, we thought to use 1500, 3000, 4500 and 6000 ex-vivo (since I read in another topic that your team recommended the b-values to be 3-4 times higher ex-vivo).
We have 110 diffusion-weighted directions.
From what I understand, more directions, higer b-values and smaller voxel size are better to resolve crossing fibers issues, but high b-values and small voxel size also affect SNR. We do have the scanning time to acquire at 62.5um isotropic, but is it worth it with the SNR trade-off?
My questions are: (1) How do we determine if we should go to 62.5um isotropic or 150um isotropic? (2) Based on what I’ve described above, do you have other advise to optimize the acquisition?
Unfortunately there’s a lot of correlated acquisition parameters offered up here, and it’s quite a difficult space within which to optimise. I’ll make some suggestions having gone through a similar process myself recently, but these are most certainly not hard and fast rules.
If your primary application is FBA, then I would advise that spatial resolution should not be the highest priority. Spatial resolution is a higher priority if you want subject-specific tractography to be more accurate. But in FBA you have both the requirement for spatial alignment between subjects, and an explicit data smoothing step, both of which are going to blur out any very fine details by the time you get to the point of quantitative analysis.
Personally I have chosen to prioritise b-value. While in the original AFD paper simulations were presented arguing that for in vivo, b >= 3,000 is necessary for specificity of the diffusion-weighted signal to intra-cellular volume, more recent evidence from the microstructural modelling community suggests that the requirement is more like 4-5k. I’d also suggest this manuscript from @sgenc for justification for higher b-values that is more pragmatic than theoretical.
Now if ~ b=5k is required for in vivo, does that mean you need ~15-20k for ex vivo? Maybe; I’ve not actually seen raw data directly interrogating that relationship. But I would suggest doing some experimentation here, maybe with a reduced FoV and just one diffusion sensitisation direction to reduce scan time, progressively increasing the b-value and increasing TE at what you think is a reasonable spatial resolution, and see how far you can go. Bear in mind here that individual DWIs don’t need to have a particularly high SNR; as long as you can see at least some structure within the noise, there’s a good chance that CSD will still do pretty well. Also w.r.t. noise, export the complex reconstructed image data if possible, since this improves denoising.
Are you constrained to using a fixed number of diffusion sensitisation directions? Ideally the platform would allow you to specify these directions manually. You can also generally acquire a smaller number of diffusion sensitisation directions at lower b-values; see @jdtournier’s work here. You may also wish to consider the distribution of b-values in between b=0 and your maximum b-value, as doing such linearly in b is not necessarily optimal; more of @jdtournier’s work here.
I hope that’s enough to maybe marginalise out a parameter or two and get some test data for evaluation.
Thanks so much for your extensive answer and literature suggestions! We will try to run some tests to see what b-values we can reach. This definitely gives us something to move forward with.