Hi Everyone,
We've been struggling for some time to come up with an appropriate method to correlate a behavioural measure to probabilistic tractography. Sorry if the post is a bit long...
There have been several options offered on the FSL listserv...
In so far, I've found that we can:
1) take each subject's average tractogram probability and correlate to the behavioural measure
2) take the seed masks # voxels (x 5000) as a denominator, and calculate the ratio that reaches each target
3) Use the waytotals as a denominator for each subject, and calculate the ratio that reaches each target
(2 and 3 are basically the same)...
I did a search in the literature to see how other groups are dealing with the issue, and found a method published in Neuroimage back in 2007 (Cohen et al), and was hoping to get your views on it...
To correlate individual differences in amygdala-seeded tractography with behavioral performance, we correlated across subjects, at each voxel, tractography results with the behavioral results (average trials until learning criteria, probability of loss/switch). Due to inter-subject variability in the spatial distribution of tracts, we included in the analyses only voxels for which at least eight subjects had non-zero tract estimates. This resulted in 7511 voxels being tested (.0285% of 263,340 brain voxels). Because this may have resulted in non-normal distributions over voxels, we used the Statistical non-Parametric Mapping (SnPM) toolbox in SPM5 ([Holmes et al., 1996] and [Nichols and Holmes, 2002]). SnPM uses a non-parametric comparison procedure, in which pseudo-t-values are calculated based on randomization/permutation testing; it is therefore more robust than the standard general linear model in cases of few degrees of freedom or non-normally distributed data. To identify significant regions, we used an uncorrected threshold of p < .01 and a cluster threshold of at least 6 contiguous voxels. We believe this to be an appropriate statistical threshold because (1) although it is one order of magnitude more liberal than the standard whole-brain threshold of .001, there are two orders of magnitude fewer voxels; and (2) all voxels are already defined as being a priori of interest by the output of an orthogonal analysis (that is, all voxels are identified by amygdala-seeded tractography, the results of which could not have been influenced by functional connectivity or behavioral performance). This is also the default setting for uncorrected tests in SnPM. One subject had corrupted diffusion data, and was not included in these analyses.
Specifically, I was under the impression that values in individual voxels were meaningless- and that correlating them to a behavioural measure would not provide a physiologically relevant finding. What does it mean if a focal probability of a tract fluctuates with behavioural finding? Doesn't the target of the tract- and the probability of reaching it make a better measure?
Your insight in this dilemma is greatly appreciated.
Regards,
Massih
|