 |
 |
Thanks for the response! This definitely answers my thoughts. Tim From: Tim Behrens [mailto:[log in to unmask]] Sent: Thursday, May 15, 2008 1:28 AM To: [log in to unmask] Subject: Re: [FSL] The meaning of distance correction in probtrackx Dear Tim and David - The short answer is this - the distance weighted solution has no reasonable interpretation or definition, nor was it intended to. It simply upweights things a long way from the seed. Whilst I don't recommend this idea in normal tractography studies unless you can think of good reason, it is often very useful in the tractography-based-clustering-for-grey-matter parcellation that we do, as it ensures the parcellation is driven by remote connectivity and not by local connectivity. The meaning of the numbers in the original unweighted case is at least defined, if not quite what you would want. T On 14 May 2008, at 21:13, Laumann, Timothy (NIH/NIMH) [F] wrote: Dr. Gutman, thanks for your (and your cat's) response. The point that the probability of connection should not be equated with 'strength' of connection is well taken, as crossing fibers will change the probability but not necessarily the 'strength' of a region to region connection. What I'm still trying to understand, however, is how the distance weighting adds a further complication to the interpretation of the probtrackx results. With distance weighting you no longer simply receive the probability of connection, but a value that reflects both the probability of connection (in the number of tracks out of the total possible which get from region A to region B) and the total length of each of the tracks which made it. When I then compare the distance-corrected output values for (region A to region B) and (region A to region C), I can no longer say that one had 0.5 probability of connection and the other has 0.8; I instead have two values which don't tell me the relative probabilities of these two connections (a 100 value could mean 1 out of 1000 tracks went 100 voxels, or 100 out of 1000 tracks went 1 voxel), nor do I even know the relative distances tracks had to travel in each case (unless you perform probtrackx again without distance correction and divide this number (total tracks) into the distance-corrected value - which may not work exactly as the outputs may not perfectly overlap due to the probabilistic nature of the calculation). I'm just not sure how to interpret a comparison of these values across subjects when the underlying sources of the values could have such variant meanings. Thanks again for your thoughts! Tim From: David Gutman [mailto:[log in to unmask]] Sent: Tuesday, May 13, 2008 7:54 PM To: [log in to unmask] Subject: Re: [FSL] The meaning of distance correction in probtrackx Nuts-- cat decided to send my message before I was done. Anyway-- be careful about interpreting the raw numbers as strength of connection-- they are actually probability's of connection, and the distance correction attempts to somewhat control for the decrease in probability as you get farther away from the seed region. For example, assuming your using more than one seed voxel, to have values at a given voxel > 5000, for example if all of the threads from neighboring voxels pass through a given point, that value will be >> 5000. More relevant, the probability of connection is not directly equatable to a strength of connection. Imagine a tract that is going straight to the frontal pole-- say with 1000 threads running down the highway to the frontal pole. Now in subject 1 the tract goes straight to the frontal lobe; in this analogy imagine this as a 5 line highway. In subject 2, the 5 lane highway also has an exit/offramp. Of the 1000 threads that were going to go down to the frontal pole, 200 of them take the "offramp" and the other 800 continue straight. You've done nothing really to the underlying "strength" of connection (i.e. both subjects have a 5 lane highway), but the probability of getting to the frontal lobe has now decreased because you have more options, 20% of the time you get off the exit, 80% of the time you've gone straight. Thus when framing questions, you have to be careful how you interpret the numbers. While the probability of getting from point A to point B might have changed (and this is interesting perhaps), the actual "strength" of the connections has not necessarily changed. Tim and Saad can likely provide oodles more detail about this dg On Tue, May 13, 2008 at 7:47 PM, David Gutman <[log in to unmask]> wrote: Timothy, Dr. Behrens or Dr. Jbadi can likely comment on this much better than I can, but On Tue, May 13, 2008 at 6:57 PM, Timothy Laumann <[log in to unmask]> wrote: Dear List, I was wondering if anyone could comment on the meaning of the results of a 'distance-corrected' probtrackx calculation as a metric of connectivity strength, relative to a non-'distance-corrected' probtrackx calculation. I understand that the latter will represent the total number of tracks which go from a seed to a target mask (units = tracks), while the former will represent the total distance these tracks traveled to reach the target (units = voxels). The non-distance-corrected output will always have a standard proportional value based on the number of tracks used, i.e. 1000 out of 5000, say, tracks made it from a seed to a target. This means I will always have an absolute measure of connectivity relative to the maximum possible strength (in this case 5000 out of 5000). With the distance-corrected output, however, there is no single upper limit to the value it could output for tracks going from a seed to a target, and that limit will change depending on where my target is. The distance correction does enhance the results for targets far away from a seed which would be washed out in a non-distance-corrected analysis, but how can I meaningfully compare these values across subjects (or even within a subject)? Is a 100 generated by one track traveling a long distance equivalent to a 100 generated by 10 tracks traveling a short distance, in terms of strength of connectivity? Further, if, for example, we compare fronto-occipital connections in two groups, one with long brains and the other with short brains, there will be a difference between the two in non-distance corrected data (assuming distance does diminish the likelihood of a track successively reaching its target), while there may not be one in the 'distance-corrected' data. However, the distance corrected data are now expressed in units of length of connectivity, which does not sound like a very meaningful biological measure, while the probability of connections (non distance-corrected) is more appealing. Another related question: what determines the reduction in probability of connection with increasing distance from the seed? Is it only the probabilistic nature of the algorithm (with increasing likelihood of deviating from the main path as the procedure is repeated on successive voxels) or also the architecture of crossing paths along the way? Any thoughts or comments would be greatly appreciated. Thanks for the help! Sincerely, Timothy Laumann -- David A Gutman, M.D. Ph.D. Department of Psychiatry & Behavioral Sciences Emory University School of Medicine -- David A Gutman, M.D. Ph.D. Department of Psychiatry & Behavioral Sciences Emory University School of Medicine