Hi Jordan,
1,2,3)
It is possible (although not always necessarily true) that the
probability of a connection ending at a certain target location
increases with the volume of the target location. However, I don't see
this as a problem!
As long as the target locations are "comparable between subjects",
meaning that they refer to the same anatomical or functional location
across subjects, this property of increased probability with size is
not undesirable.
I concede that in some cases, it is not easy to match cortical regions
across subjects, but this is the core of the problem - not the
dependence of tractography on roi size.
Dividing by roi size will result in a quantity that is not very
interpretable. A more proper way to do it would be to have a model for
the probability of connection between A and B given that B has a
certain size, which is very hard to formalise. (there is a similar
issue with the effect of distance between A and B).
One thing you can probably do is to include the target roi size into
your correlation analyses as co-regressors. Note that neither the
result nor the interpretation of such analysis would be the same as
just dividing by roi size. You will instead ask the question of
whether a behavioural measure can be explained by roi size alone, or
whether adding connectivity measures to those rois gives you a better
model. etc.
1)
When considering the values in fdt_paths, one thing to be aware of is
that, in order for these values to be comparable between subjects, the
voxel size need to be the same across subjects.
1)
The connectivity scores that fdt_paths stores are the probability that
the connection through a seed region ends up anywhere in the brain.
Such quantity is not straightforward to interpret in terms of
connection "strength", but rather influenced by many factors, some of
them purely related to the morphology of the connection along its
course. So in general, small values in fdt_paths do not always mean
small connectivity strength. We prefer to use values from the
seed_to_target-type analysis where target rois are chosen carefully
(i.e. represent meaningful brain locations that are comparable across
subjects).
2)
similarly for target masks, the seed mask do not need to have the same
size across subjects, but it is more crucial that it has the same
anatomical meaning across subjects.
3)
Although the seed_to_target results are more comparable across
subjects, there still can be a large variability that is due to
undesirable properties in the data, such as noise/partial volume
effects/etc. which make these measures hard to combine between
subjects. One way to reduce this variability is to use a command line
called proj_thresh, which transforms the seed_to_target values into
proportions (i.e. the sum of connectivities to all targets is equal to
one).
4)
it is useful to consider the tractography procedure as integrating a
differential equation, in which we write that the tangent of a curve
is equal to the local fibre orientation. in this context, the step
size that is used in the tractography is the step size in numerical
approximations to ODEs. The general rule with this parameter is that
the smaller it is, the better (i.e. less error in the integration).
There are ways to also adapt the parameter to be smaller when the
function to integrate changes faster, by e.g. fixing the amount of
error cumulated at each step.
Now, on the other hand, the "function" to integrate in tractography is
discretised (given at each voxel), so we need to interpolate its value
at each location. If the step size is too small, then the
interpolation will basically lead to very little (if any) change in
the value of the function, making the tractography less efficient
(spending too much time making tiny steps in the same direction). So
there is a trade-off, which is dependent on the quality of your data
(especially the voxel size).
Cheers,
Saad.
On 10 Sep 2009, at 04:06, Jordan Poppenk wrote:
> Dear FSL Community,
>
> I am trying to understand the appropriate interpretation of probtrackx
> output. In particular, I have a few questions regarding
> normalization steps
> that may be necessary:
>
> 1) I would like to create connectivity maps seeded from a mask in
> order to
> regress a between-subjects behaviour variable on the maps. However,
> the seed
> mask is based on a different anatomically-determined segmentation
> for each
> participant. To prepare the data for group analysis, should I divide
> each
> fdt_ map by the number of voxels in each subject's segmentation?
>
> 2) Along the same lines, would it be fair to compare two maps, each
> generated from a different seed mask in the same subject, if I first
> divide
> the maps by the number of voxels in each mask?
>
> 3) When using "classification targets" with a seed mask, the number
> of paths
> to each target mask will again be influenced by the number of voxels
> in that
> mask, correct? So if I am going to perform hard classification based
> on
> these values (using find_the_biggest, as done in Cohen et al.,
> 2009), should
> I not first divide the values in each seed_to_target_mask file by
> the number
> of voxels in the target mask (to give each target an equal chance)?
>
> Finally, on the parameters side, there seems to be little consensus
> on an
> appropriate step length. I have read that 1/10th voxel size is
> appropriate;
> that one should not venture below 0.5; and that the smallest size with
> stable tract output should be selected. Is it safe to leave this
> parameter
> as-is at 0.5? I have tried values ranging from 0.25 to 2.5 with a
> voxel size
> of 2.5mm and cannot intelligently identify an optimal output.
>
> Thanks in advance for your help.
>
> Sincerely,
> Jordan Poppenk
> PhD Candidate, Rotman Research Institute
>
--
Saad Jbabdi
University of Oxford, FMRIB Centre
JR Hospital, Headington, OX3 9DU, UK
+44 (0) 1865 222523 (fax 717)
www.fmrib.ox.ac.uk/~saad
|