[Mrtrix-discussion] Streamtrack and connectivity values

[Robert Smith]

Hi Simon

It's impossible to diagnose what's causing this variability from your
description alone, so I'll rattle off a few ideas for how you can better
diagnose the issue.
But as a quick answer: the mask resolution shouldn't drastically alter the
results of tractography, apart from the obvious differences e.g. partial
volume. This is because streamlines are generated in 'real' / 'scanner'
space, and image information is drawn from the DWIs / ROIs using tri-linear
interpolation as the streamlines are propagated. Therefore, as long as the
information from the various images is aligned in 'real space', it doesn't
matter whether or not the images actually lie on the same voxel grid.



Firstly, given the number of pre-processing steps required in your analysis
with regards to transforms, I would start by having a close inspection of
your data; make sure the images are in fact aligned with one another and
your ROIs are in sensible places. With the 1x1x1 data, use the overlay
feature in mrview to make sure the registration of your T1-based images
into DWI space is correct.

Not sure if you omitted this step in your description for brevity, but just
in case: erosion of the white matter segmentation mask is not adequate for
estimating the response function. In addition to this erosion, a threshold
should be applied to the FA image, to select a subset of white matter
voxels that may be interpreted as containing only a single dominant fibre
population. Failure to perform this step correctly could result in
erroneous FOD estimates, which in turn may result in erratic tracking.
Check the documentation<http://www.brain.org.au/software/mrtrix/tractography/preprocess.html>if
you are unfamiliar with this step.

For the csdeconv command, the -mask option simply tells the command which
voxels to process and which to skip; therefore the mask image must be
identical to the DWI volume in voxel size / image dimensions / image
transform.

If you are constructing a connectome matrix from a FreeSurfer parcellation,
1 million streamlines is a fairly small number. The reason I bring this up
is that if you were to perform repeats of your experiment, using the same
set of input images each time, the resulting matrices would still differ
from one another; this is due to the random placement of streamline seeds,
and (if you are using SD_PROB) the probabilistic nature of the tracking.
Increasing the number of streamlines may reduce this variability.
Alternatively, you could generate multiple reconstructions of 1 million
streamlines, using the same input images each time, and assess the
variability in the resulting matrices; if this variability is comparable to
what you've reported here, then the mask resolution is probably not the
causative factor.

If none of these are the cause of your variability, a couple of final
suggestions:
* Generate a TDI for each of your streamlines reconstructions using the
tracks2prob command, and make sure the tracking is behaving as it should in
both cases. Differences in your connectivity values may be a secondary
marker of a more fundamental problem with the reconstruction(s).
* I'm not sure whether this is possible with the configuration you have for
constructing your matrices, but I always think it's a good idea to take a
look at those streamlines that fail to connect a pair of parcellation
nodes. If a large proportion of streamlines don't connect two grey matter
regions, it's indicative of a larger problem.



Hopefully somewhere in there is an explanation for your matrix variability.
Any further questions / problems let me know.

Rob


--

Robert Smith
PhD Candidate

The Florey Institute of Neuroscience and Mental Health
Melbourne Brain Centre - Austin Campus
245 Burgundy Street
Heidelberg Vic 3084
Ph: +61 3 9035 7128
Fax: +61 3 9035 7301
www.florey.edu.au


On Thu, Jan 17, 2013 at 12:30 PM, Simon Baker <simontebaker at gmail.com>wrote:

> Hi Donald et al.,
>
> It appears that our connectivity output depends on the voxel dimensions of
> the seed, include, and mask images that are used when performing
> "streamtrack".
>
> We have used two approaches to get our seed (binarised and dilated
> FreeSurfer white matter segmentation), include (parcellation based on
> FreeSurfer aparca2009s and FSL FIRST), and mask (combination of seed and
> include) images into FA space.
>
> Prior to implementing Approach 1 or Approach 2, which are described below,
> we put our structural (FreeSurfer brain.mgz), seed, include, and mask
> images into standard space using "fslreorient2std", then we coregistered
> the FA image (generated by MRtrix) to the structural image (in standard
> space) and generated what we will call the fa2std transform matrix using
> FSL's "flirt". The target images refer to the seed, include, and mask
> images (all in standard space).
>
> Approach 1. FSL. We inverted the fa2std transform matrix using
> "convert_xfm -inverse". Then, using the inverted fa2std transform matrix in
> "flirt -applyxfm", we applied spatial transformations to the target images,
> creating outputs that were in FA space and that had the same voxel
> dimensions as the FA image (2.3x2.3x2.3).
>
> Approach 2. MRtrix. Using "mrtransform -inverse", we inverted the fa2std
> transform matrix and applied spatial transformations to the target images,
> creating outputs that were in FA space, but which retained their native
> voxel dimensions (1x1x1).
>
> These outputs were used in subsequent steps as described below.
>
> We eroded the 2.3x2.3x2.3 white matter segmentation to create the single
> fibre mask. Note that we encountered an error when attempting to erode the
> 1x1x1 white matter segmentation, probably because the "erode" step includes
> specification of the FA image (which is 2.3x2.3x2.3) for "mrmult" and
> therefore the voxel dimensions are not compatible.
>
> We used the single fibre mask to estimate the response function
> coefficient.
>
> We performed "csdeconv" with the mask specified as our 2.3x2.3x2.3 hybrid
> brain mask (a combination of the 2.3x2.3x2.3 seed and include images). Note
> that we encountered an error when attempting to specify our 1x1x1 hybrid
> brain mask as the mask, again probably due to incompatible voxel dimensions.
>
> Then we performed "streamtrack" twice: firstly at the 2.3x2.3x2.3
> resolution (i.e., we used the 2.3x2.3x2.3 seed, include, and mask images);
> and secondly at the 1x1x1 resolution (i.e., we used the 1x1x1 seed,
> include, and mask images). We also discovered that "streamtrack" will allow
> any combination of seed, include, and mask images; i.e., the voxel
> dimensions of these images do not need to be identical - why is this?
> Anyhow, we generated one million streamlines using either all 2.3x2.3x2.3
> images or all 1x1x1 images.
>
> From each of the track files we extracted streamlines connecting two or
> more regions (where regions were specified by the include image). Then we
> generated two connectivity matrices (one that was based on the 2.3x2.3x2.3
> images/streamlines and another that was based on the 1x1x1
> images/streamlines) by counting the number of streamlines interconnecting
> each pair of regions. Interestingly, these connectivity matrices were
> noticeably different, suggesting that the connectivity values depended on
> the voxel dimensions of the seed, include, and mask images that were used
> when performing "streamtrack".
>
> We were wondering whether the resolution of the masks affects the
> tractography in anyway? Is there any reason for preferring one over the
> other? We presume that the native 1 x 1 x 1 resolution is preferable, but
> wanted to make sure we are not missing anything here.
>
> Thanks in advance,
>
> --
>
> Simon Baker
> PhD Candidate
> Faculty of Medicine, Nursing and Health Sciences
> Monash University
> _______________________________________________
> Mrtrix-discussion mailing list
> Mrtrix-discussion at www.nitrc.org
> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20130119/3aab494d/attachment.html