Connectivity measures to be sampled

[sina-mansour]

Following on the suggestion by @Lestropie (this commit):

RS:

• We had discussed the prospect of modifying the tcksample command to derive, for each streamline trajectory, samples from multiple quantitative 3D images, so that tcksample would not need to be executed once per quantitative measure, which could be slow due simply to having to read through the track file once per quantitative measure.
• One I did not think of at the time is that we could theoretically do the same approach as above, but for connectome construction. tck2connectome could receive as input a 4D image, where each volume of the series is a unique hard parcellation, and the command would generate in a single invocation one connectome per volume. This would however require a decent amount of work on my part, so we'd need to see how many parcellations you want to use and how long tck2connectome takes per execution to know whether or not this is worthwhile.
• As per a prior point above: if the tractogram were stored on a RAM filesystem rather than on a network file share, then the time spent loading streamlines data in both tcksample and tck2connectome would be much shorter.
• Potential additional connectivity metrics:
  • Properties of multi-tissue decomposition (e.g. isometric log-ratios)
  • Pre-calculated NODDI parameters
  • FOD amplitude (existing code can only do exact FOD amplitude at streamline tangent; code for sampling AFD as the integral of the FOD lobe is part of a very large changeset that I started ~ 8 years ago; if it's considered highly desirable I can try to get it into a working state)
  • Note that pre-calculated tensors in UKB (presumably via FSL) will differ to those calculated by MRtrix3's dwi2tensor. Ours uses a thoroughly-evaluated iterative reweighting strategy (see references in command help page). So worth contemplating whether to use the existing pre-calculated tensors, or re-calculate using MRtrix3.
• Note also that for each quantitative metric, taking the mean value along the streamline trajectory and attributing that single scalar value to the streamline is the typical solution, but it's not a unique solution; one can also take e.g. the minimum value along the streamline trajectory (see tcksample -stat_tck option). Many combinations of voxel-wise quantitative metric and streamline-wise statistic will not make sense; but it's nevertheless worth considering the whole space of possibilities.

[sina-mansour]

• I will use the existing tcksample and tck2connectome code and check the runtime, once we have an estimate of how long these steps will take, we could revisit and decide whether we should look for methods to read the track file once and save some time in this step.
• This reminded me of our previous discussion with regards to adding the functionality for tck2connectome to read in a soft parcellation as a 4d image and use that in connectome building, it would be great to have this in MRtrix, I could add extra codes on my side (to the connectome smoothing package) to generate such a 4d image for any arbitrary volumetric parcellation
• The tractograms are saved on the scratch filesystem
• I'll try including as many different measures as would be possible and make sense. I'll also add the option of minimum and maximum and test the time required to generate it all and also the required space. Once this part of the script is updated, I'll add the detailed list of all included measures in this issue thread. Please feel free to add any other suggested measures in the meanwhile.
• I'll open a separate issue on the question of how large a file we can submit to the UKB data center, as that is also another deciding factor (check https://github.com/sina-mansour/UKB-connectomics/issues/11).

[sina-mansour]

• Many combinations of voxel-wise quantitative metric and streamline-wise statistic will not make sense; but it's nevertheless worth considering the whole space of possibilities.

Speaking of the whole space of possibilities; this is a list of all metrics that can be extracted:

• ADC (MD)
• FA
• AD
• RD
• CL, CP, CS
• tenser eigenvalues

(I'm not sure if I've missed anything, please let me know if there are other useful metrics that we could potentially add.)

For each metric above, different statistics could be computed at the level of every streamline: mean, median, min, max, & std. Additionally, we can compute streamline lengths.

Finally, all the streamline level metrics can be aggregated to a connectivity matrix. and for all streamlines contributing to a single connectivity edge, there are different ways we could combine the streamline statistics: mean, min, max, & sum.

I feel that mean would be most meaningful for many of the metrics, but sometimes, median along the streamline length or even the extremum min and max values could be informative.

In the table below I've added a short list of bare minimum measures that we would have. Please feel free to add options you think would be useful:

metric	streamline level aggregation	edge level aggregation
FA	mean	mean
ADC (MD)	mean	mean
RD	mean	mean
AD	mean	mean
CL	mean	mean
CP	mean	mean
CS	mean	mean
length	NA	mean

@mdibiase1 do you think there are relevant metrics for free water that we could add?

[Lestropie]

It's possible to compute ADC without invoking the tensor model, which is how I prefer to do it myself; see dwi2adc (though it's currently rather clunky to use).

Also that list is specifically sampled voxel-wise measures plus length, so it's missing both raw streamline count and FBC.

[sina-mansour]

Oh, yeah, that's right, I missed those in the table above, here's the updated list:

metric	streamline level aggregation	edge level aggregation
Streamline count	NA	sum
FBC (SIFT2)	NA	sum
FA	mean	mean
ADC (MD)	mean	mean
RD	mean	mean
AD	mean	mean
CL	mean	mean
CP	mean	mean
CS	mean	mean
length	NA	mean

TBH, I'm not even sure if CP, CL, CS, RD, & AD are that useful, but I've left them there just in case someone is interested to investigate that on a connectivity matrix.

[Lestropie]

Technically, for all of the voxel-wise measures you can also do not a straight mean across streamlines but a weighted mean where streamlines with greater weights as determined by SIFT2 contribute more to the mean.

There's also the pre-computed NODDI metrics, and the prospect of DKI metrics.

[sina-mansour]

The weighted mean is also a good alternative for edge-level aggregation.

Yes, the NODDI metrics are also pre-computed, these are the NODDI metrics:

• ICVF: intra-cellular volume fraction - an index of white matter neurite density
• ISOVF: isotropic or free water volume fraction
• OD: orientation dispersion index, a measure of within-voxel tract disorganization

We could also add these to the list.

and the prospect of DKI metrics

What DKI measures would you recommend to include?

[Lestropie]

What DKI measures would you recommend to include?

Very much not my area of expertise. Apparently there's not even that much of a consensus in the field w.r.t. how different metrics are defined. Would need to do a bit of a search across the most recent literature.

[mdibiase1]

What DKI measures would you recommend to include?

Very much not my area of expertise. Apparently there's not even that much of a consensus in the field w.r.t. how different metrics are defined. Would need to do a bit of a search across the most recent literature.

[mdibiase1]

Your list of metrics looks good to me. I don't think these DWI data are multishell (?), so DKI measures aren't suitable. I would recommend leaving out FW measures too as the single shell FW code hasn't been made public yet, which is problematic for ukb.

[Lestropie]

I don't think these DWI data are multishell (?), so DKI measures aren't suitable.

No, UKB has b=0,1000,2000.

[mdibiase1]

I don't think these DWI data are multishell (?), so DKI measures aren't suitable.

No, UKB has b=0,1000,2000.

In this case, DKI metrics can be estimated. The following code is quite popular: https://www.nitrc.org/projects/dke

[sina-mansour]

I don't think these DWI data are multishell (?), so DKI measures aren't suitable.

No, UKB has b=0,1000,2000.

In this case, DKI metrics can be estimated. The following code is quite popular: https://www.nitrc.org/projects/dke

What about the FW measures? Is it worth including?

[Lestropie]

I'm not super familiar with how the tensor free water estimation methods work. Realistically they're not all that different to the CSF-like component of the MSMT fit. I'm also not sure of how much utility such a measure is for connectome connectivity information?

[sina-mansour]

So as to finalize the discussion on connectivity measures included, the table below lists all measures for which connectivity is being mapped:

metric	streamline level aggregation	edge level aggregation
streamline count	NA	sum
fiber bundle capacity (SIFT2)	NA	sum
length	NA	mean
fractional anisotropy (FA)	mean	mean
mean diffusivity (MD)	mean	mean
mode of the anisotropy (MO)	mean	mean
raw T2 signal (S0)	mean	mean
NODDI intra-cellular volume fraction (NODDI_ICVF)	mean	mean
NODDI isotropic volume fraction (NODDI_ISOVF)	mean	mean
NODDI orientation dispersion index (NODDI_OD)	mean	mean