Feasibility of iFOD2 in single shell DTI


Dear Community,
I am looking for previous publications which support the feasibility of SIFT tractography using iFOD2 to reconstruct whole brain tractography in clinical DTI data (i…e, 3T single shell DTI 55 encoding at b=1000). One pointed out this state-of-the art is only for HARDI which I cannot agree 100%. This method could work with crossing fiber problem although it is suboptimal at single shell data (high sensitivity and low specificity). Any thoughts or references to defend this point would be greatly appreciated. Best, Justin.


Can you post the objection that was raised about this exactly? There’s various things that could apply here, it would help to know which one specifically is being raised as an issue.

Some thoughts:

  • both SIFT and iFOD2 were originally proposed on single-shell HARDI data. There’s no issue there. Both of these rely on the availability of adequate fODFs, so the question really reduces to whether the fODF produced by CSD (and now multi-tissue CSD, even from single-shell data) are suitable for the purposes of SIFT and iFOD2.

  • Your data already qualify as HARDI - 55 directions is well within the HARDI range. The main difference here is the use of a lower b-value than we’ve used in previous publications. This may influence interpretation of the results, but most likely in a relatively subtle way (see e.g. this post). But I certainly don’t see it being problematic either for SIFT or iFOD2.

  • One consequence of lower b-values is that the GM signal is higher than it would be at higher b-values, which introduces more scope for spurious connections through the grey matter. But high b-values are not immune to this problem, and these days, ACT is by far the better approach to dealing with this anyway.

But otherwise, there is nothing inherently wrong with using SIFT or iFOD2 on your data – I’m sure plenty of users already do with similar data… It might not be the absolute best that can be done nowadays, but it’s certainly feasible. How well this works will overwhelmingly depend on your image quality, SNR in particular. You have sufficient angular coverage already (see e.g. this paper on the topic if you’re interested), the thing that matters now is whether you have adequate contrast to noise ratio (CNR). You won’t have as much contrast at b=1000 as you might at b=3000, but if your SNR is sufficient, that will help regain the CNR. I’d give it a shot and see how well it works.


Hi, Dr. Tournier
Actually, one reviewer sent us below comment on our qualification to do diffusion studies. [ “Whole brain streamline tractography was then reconstructed using probabilistic SIFT tractography with second-order integration over fiber orientation distributions (iFOD2) to sample up to three FOD at every voxel .” This is cutting-edge HARDI tracking! Which again really make me realize that authors do not properly understand the difference between dMRI acquisitions and DTI/HARDI modeling.] I assume that the reviewer is thinking of our data is not within the HARDI range, maybe due to single low b-value (1000) and relatively low encoding directions (55). No other reviewers did not mention about this issue. I will make a comment based on your NMR paper which suggested a minimum of 45 DW directions is sufficient to fully characterise the DW signal using CSD based methods like iFOD2 etc . Thank you so much for considerable help!


That’s quite a confusing comment from that reviewer indeed… it’s also not clear to me what point they’re actually trying to make by commenting that on such a sentence in your manuscript. If anything, I’m getting the (maybe wrong) impression they are not understanding the difference between acquisition (strategies) and modelling approaches. They’re making 2 mistakes right there themselves about this point: it’s not “HARDI” tracking, but FOD tracking indeed; the FOD being a model and “HARDI” not being a model, but merely an acquisition strategy. Second, they’re lumping “DTI” and “HARDI” together in a phrase about modelling. Again, HARDI is not a model…

So well, I’m not sure what “difference between dMRI acquisitions and DTI/HARDI modeling” they’re talking about there… :confused:

That said, unrelated to their confusing comment, you can make a minor improvement to your original statement by changing “…to sample up to three FOD at every voxel” into “…to sample up to three FODs at every step”. Those (three) samples don’t per se happen in a single voxel, and it’s also slightly weird to state “three FODs at every voxel”, since there’s only 1 FOD in each voxel. :wink: Of course, generally, interpolation will use most of the time at least the 8 surrounding voxels for any point, and hence use their FODs. But still only a single interpolated FOD is sampled (and that up to three times for each iFOD2 step). Does that make sense?


Thanks a lot for your useful comment. I will correct it in my manuscript. I believe there are many people with different background working in our field. So, it’s often difficult to communicate well.


Just to check: I assume this was what you wrote in your manuscript:

and this was the reviewer response:

Is that correct?

Assuming that’s the case, there’s a few things to point out:

  • the main problem from my point of view is that your description of iFOD2 is incorrect (a point that the reviewer interestingly didn’t pick up…). It “… samples up to …” implies that it might take fewer samples, which isn’t the case: it will always take 3 samples (by default). Also, it’s a bit confusing to say it sample 3 FOD per voxel: it samples the FOD field at 3 different locations (& corresponding tangent directions) along the candidate path segment for the next step. You should reword to something like:

    “… over fiber orientation distributions (iFOD2) to sample the FOD at three equidistant sample points along each candidate path segment for the next step”

  • Your description of SIFT is also a little bit loose. It suggests that SIFT is part of the tractography process, which it isn’t: it’s applied after tracking as a post-processing step. I’d recommend you reword to clarify that.

  • I struggle to understand the reviewer’s response to that – maybe there’s more context around your statement that might explain this? But essentially, there’s a suggestion that because iFOD2 is “cutting-edge tracking” (nice to hear :wink:), it can’t work with non-HARDI acquisitions. I disagree with that, iFOD2 only requires an estimate of the FOD field, which you could obtain using any number of approaches (although we obviously strongly recommend CSD & multi-tissue CSD here). The acquisition only matter in so much as it influences the quality of the estimated FODs – but it doesn’t affect the validity of using iFOD2 on these FODs once you’ve got them.

  • The reviewer’s response also implies that your acquisition doesn’t qualify as HARDI. I think with 55 directions, it clearly already is HARDI – we’ve been using 60 directions routinely for ~15 years, and no-one has yet suggested that this is not HARDI. I don’t think there’s any consensus as to how many directions are required before you can label an acquisition ‘HARDI’, but personally I think anything from around 30 DW directions and up would qualify (it can certainly be processed, especially if the SNR is high). Many users here use these approaches on 30 DW direction data, which is clearly not great, but still works more or less as expected nonetheless (the main consequence is the obvious loss of SNR compared to a more densely sampled acquisition). I’ve also seen reasonable results on much poorer quality data than that – see fig 10 from my original 2007 paper if you need to make that point.

    They may be objecting to the relatively low b-value (b=1000 s/mm² is indeed ‘standard’ for clinical DTI), but that doesn’t prevent these data from being successfully processed with CSD (see Fig 9 from my original 2007 paper). It does reduce the contrast somewhat, but not to the extent that the data can’t or shouldn’t be processed. It also doesn’t impact on whether the acquisition qualifies as HARDI – I don’t recall anyone arguing that HARDI acquisitions should be high b-value. Also, I’ve already made a few points about the b-value issue in a previous post, and they still apply.

Hope this helps,


Hi, Donald
Yes, you’re correct. My best guess is that the reviewer believes our 55 directions acquisition is not HARDI. Thus, iFOD2 is not a good choice for whole brain tractography. He seems to believe HARDI as >100 encoding at multiple b-values like HCP data. Your comments are really useful for my rebuttal, especially Fig 9 and 10 of your 2007 paper- I could find those figures once I had placed my post. I hope this reviewer will also learn a lot with those figures. Best, Justin.