Spherical-deconvolution Informed Filtering of Tractograms (SIFT)

SIFT [Smith2013], or ‘Spherical-deconvolution Informed Filtering of Tractograms’, is a novel approach for improving the quantitative nature of whole-brain streamlines reconstructions. By producing a reconstruction where the streamlines densities are proportional to the fibre densities as estimated by spherical deconvolution throughout the white matter, the number of streamlines connecting two regions becomes a proportional estimate of the cross-sectional area of the fibres connecting those two regions. We therefore hope that this method will attract usage in a range of streamlines tractography applications.

The actual usage of SIFT can be found in the help page of the tcksift command. In this page I’ll outline some issues that are worth thinking about if you are looking to apply this method.

DWI bias field correction

DWI volumes often have a non-negligible B1 bias field, mostly due to high-density receiver coils. If left uncorrected, SIFT will incorrectly interpret this as a spatially-varying fibre density. Therefore bias field correction is highly recommended. We generally estimate the bias field based on the mean b=0 image, and apply the estimated field to all DWI volumes. This can currently be achieved using the dwibiascorrect script, which can employ either the FAST tool in FSL or the N4 algorithm in ANTS to perform the field estimate.

Number of streamlines pre / post SIFT

In diffusion MRI streamlines tractography, we generate discrete samples from a continuous fibre orientation field. The more streamlines we generate, the better our reconstruction of that field. Furthermore, the greater number of streamlines we generate, the less influence the discrete quantification of connectivity has on the connectome (e.g. would rather be comparing 1,000 v.s. 2,000 streamlines to 1 v.s. 2; it’s less likely to be an artefact of random / discrete sampling). So the more streamlines the better, at the cost of execution speed & hard drive consumption.

However we also have the added confound of SIFT. The larger the number of streamlines that can be fed to SIFT the better, as it can make better choices regarding which streamlines to keep/remove; but it also introduces a memory constraint. SIFT can deal with approximately 4-8 million streamlines per GB of RAM (depending on the seeding mechanism used and the spatial resolution of your diffusion images), so ideally you’ll want access to dedicated high-performance computing hardware. On top of this, there’s the issue of how many streamlines to have remaining in the reconstruction after SIFT; the more streamlines that SIFT removes, the better the streamlines reconstruction will fit the image data, but the more likely you are to run into quantisation issues with the resulting tractogram.

So when you design your image processing pipeline, you need to consider the compromise between these factors:

  • Initially generating a larger number of streamlines is beneficial for both the quality and the density of the filtered reconstruction, at the expense of longer computation time (both in generating the streamlines, and running SIFT), and a higher RAM requirement for running SIFT.

  • Filtering a greater number of streamlines will always produce a superior fit to the image data, at the expense of having a lower-density reconstruction to work with afterwards, and a slightly longer computation time.

Unfortunately there’s no single answer of how many streamlines are required, as it will depend on the diffusion model, tractography algorithm, and spatial extent of your target regions / connectome parcellation granularity. There are a couple of papers / abstracts on the topic if you look hard enough, but nothing definitive, and nothing involving SIFT. I would recommend testing using your own data to find numbers that are both adequate in terms of test-retest variability, and computationally reasonable.

Personally I have been using a FreeSurfer parcellation (84 nodes), generating 100 million streamlines and filtering to 10 million using SIFT (I’m a physicist; I like orders of magnitude). In retrospect, I would say that when using white matter seeding, filtering by a factor of 10 is inadequate (i.e. the fit of the reconstruction to the data is not good enough); and with grey matter - white matter interface seeding, a final number of 10 million is inadequate (the streamlines are mostly very short, so the appearance of the reconstruction is quite sparse). Another alternative is ‘dynamic seeding’, which uses the SIFT model during tractogram generation to only seed streamlines in pathways that are poorly reconstructed (see the -seed_dynamic option in tckgen); this provides a better initial estimate, so the percentage of streamlines that need to be removed in order to achieve a good fit is reduced. I will leave it to the end user to choose numbers that they deem appropriate (unless we do a paper on the topic, in which case you will use our published values without question).

Normalising connection density between subjects

An ongoing issue with our Apparent Fibre Density (AFD) work is how to guarantee that a smaller FOD in a subject actually corresponds to a reduced density of fibres. Structural connectome studies have a similar issue with regards to streamline counts; Even if SIFT is applied, this only guarantees correct proportionality between different connection pathways within a subject, not necessarily between subjects. The simplest and most common solution is simply to use an identical number of streamlines for every subject in connectome construction; however this isn’t perfect:

  • The distribution of streamlines lengths may vary between subjects, such that the reconstructed streamlines ‘density’ differs.

  • A subject may have decreased fibre density throughout the brain, but be morphologically normal; if the same number of streamlines are generated, this difference won’t be reflected in the tractogram post-SIFT.

  • If the white matter volume varies between subjects, but the actual number of fibres within a given volume is consistent, then the subject with a larger brain may have an elevated total number of fibre connections; this would also be missed if the number of streamlines were fixed between subjects.

It’s also possible to scale by the total white matter volume of each subject; this would however fail to take into account any differences in the density of fibres within a fixed volume between subjects.

An alternative approach is to try to achieve normalisation of FOD amplitudes across subjects, as is done using AFD. This requires a couple of extra processing steps, namely inter-subject intensity normalisation and use of a group average response function, which are also far from error-free. But if this can be achieved, it means that a fixed density of streamlines should be used to reconstruct a given FOD amplitude between subjects, and then the cross-sectional area of fibres represented by each streamline is also identical between subjects; this can be achieved by terminating SIFT at a given value of the proportionality coefficient using the -term_mu option. One potential disadvantage of this approach (in addition to the issues associated with intensity normalisation) is that using a group average response function instead of the individual subject response may result in spurious peaks or incorrect relative volume fractions in the FODs, which could influence the tracking results.

Ideally, a diffusion model would provide the absolute partial volume of each fibre population, rather than a proportional quantity: this could then be used directly in SIFT. However the diffusion models that do provide such information tend to get the crossing fibre geometry wrong in the first place…

If anyone has any ideas on how to solve this pickle, let us know.

No DWI distortion correction available

SIFT should ideally be used in conjunction with ACT; by passing the ACT 5TT image to tcksift using the -act option, the command will automatically derive a processing mask that will limit the contribution of non-pure-white-matter voxels toward the model. Without this information, non-pure-white-matter voxels adversely affect both streamlines tractography, and the construction of the SIFT model.

If you are looking to apply SIFT without correction of DWI geometric distortions (and therefore without reliable high-resolution co-registered anatomical image data), these are some points that you may wish to consider:

  • The spatial extent of the DWI mask may have a large influence on your streamlines tractography results. Therefore greater care should perhaps be taken to validate this mask, including manual editing if necessary.

  • It is possible to manually provide a processing mask to tcksift using the -proc_mask option. If users are capable of heuristically generating an approximate white matter partial volume image from the DWI data alone, this may be appropriate information to provide to the SIFT model.

Use of SIFT for quantifying pathways of interest

In some circumstances, researchers may be interested in the connection density of one or two specific pathways of interest, rather than that of the whole brain. SIFT is still applicable in this scenario; however the SIFT algorithm itself is only applicable to whole-brain fibre-tracking data. Therefore, the workflow in this scenario should be: * Generate a whole-brain tractogram; * Apply SIFT; * Extract the pathway(s) of interest using tckedit. * Get the streamline count using tckinfo.

The SIFT algorithm is not directly applicable to targeted tracking data. The underlying biophysical model in SIFT assumes that the estimated density of each fibre population in every voxel of the image should be proportionally reconstructed by streamlines; if only a subset of pathways in the brain are permitted to be reconstructed by the tractography algorithm, this will clearly not be the case, so appplication of SIFT in this instance will provide erroneous results.