That’s right. Matrix1 seeding strategy (seeds on the white matter surface) has a distance bias that is may ways similar to the one in real tracer data though via a different mechanism. Matrix3 seeding strategy (seeds in the white matter) does “correct” for a distance bias in the sense that longer paths have more seeds than shorter ones. In the end, it is not at all clear whether one should correct for a distance bias. Matrix1 outperformed matrix3 in matching to tracers, but only slightly. Perhaps if I understood the specific application, I could give a more informed opinion.
Peace, Matt. From: Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> Date: Monday, October 9, 2017 at 3:49 PM To: "Gopalakrishnan, Karthik" <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> Cc: Matt Glasser <glass...@wustl.edu<mailto:glass...@wustl.edu>>, "hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>" <hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>> Subject: Re: [HCP-Users] Distance between surface ROIs in MMP On Sun, Oct 8, 2017 at 11:41 PM, Gopalakrishnan, Karthik <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> wrote: Thanks Matt and Tim! Matt, I also wanted to know — in the dDT1 method to obtain WGM-to-WGM surface matrix reported in the paper you referenced, the number of streamlines isn’t corrected for distance bias, right (I believe it’s just for dDT3)? If it has, could you please share how? Tim, you mentioned the longer a probabilistic streamline gets, the wider the area that it could have hit gets, but much of this area gets intercepted by pieces of cortex before the streamline gets as long as it "should” be - therefore this spreading effect causes long streamlines to be rarer than they should be, by virtue of the streamline length itself. If I were to separately run probtrackx with a subset of ROIs from the Glasser parcellation that are pairwise “far” from each other using computed pairwise distances based on the approach Matt mentioned (all-to-all connectivity, followed by weighted average of vertex distances with # of streamlines as the weights), then would the newly computed streamlines be less “rare” vis-a-vis the computed streamlines with all ROIs (and this is reflected in the # of streamlines)? Because the other pieces of cortex that intercept the area a streamline could have hit are no longer part of the set of ROIs we’re tracking with (from and to)? Using seed masks doesn't change how each streamline behaves, nor where it stops - ones that stop by hitting something that isn't selected for reporting just means that streamline is ignored, wasting that computation. The relative strength reported between different ROIs from the same ROI should always be the same regardless of whether the full surface or only some ROIs are tracked - it should be equivalent to taking the full area to area matrix and zeroing parts of it. I want to make sure that the WGM-to-WGM matrix I obtain with the MMP parcellation and the subject's diffusion data accounts for distance bias in the best possible manner. Specifically, I want to make sure that any two connection strengths (or # of streamlines, or connection probability, whatever we call it) I pick from the WGM-to-WGM matrix are truly comparable, which they aren’t if there’s distance bias and consequently I wouldn’t be able to apply a global threshold to obtain a neuro-biologically significant unweighted graph from the matrix. I don't think we have decided what the "best possible manner" is for dealing with the distance bias (or we would already have dealt with it). Regards, Karthik On Oct 6, 2017, at 7:32 PM, Glasser, Matthew <glass...@wustl.edu<mailto:glass...@wustl.edu>> wrote: We did it for all to all connectivity at the vertex level and then did a weighted average according to number of streamlines (as the more streamlines, the more robust the distance measure). Peace, Matt. From: Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> Date: Friday, October 6, 2017 at 7:30 PM To: Matt Glasser <glass...@wustl.edu<mailto:glass...@wustl.edu>> Cc: "Gopalakrishnan, Karthik" <gkart...@gatech.edu<mailto:gkart...@gatech.edu>>, "hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>" <hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>> Subject: Re: [HCP-Users] Distance between surface ROIs in MMP Would that require them to run a separate probtrackx for each seed area? If the hits are recorded on the surface, is the distance also reported on vertices? If you can only get the distances in white matter voxels, or only the distances from the seed point, things could get challenging if you want to use different seeding strategies. Tim On Fri, Oct 6, 2017 at 6:24 PM, Glasser, Matthew <glass...@wustl.edu<mailto:glass...@wustl.edu>> wrote: --ompl option in probtrackx2. Peace, Matt. From: "Gopalakrishnan, Karthik" <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> Date: Friday, October 6, 2017 at 7:05 PM To: Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> Cc: Matt Glasser <glass...@wustl.edu<mailto:glass...@wustl.edu>>, "hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>" <hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>> Subject: Re: [HCP-Users] Distance between surface ROIs in MMP This was useful and I shall make sure to go through the paper you referenced, Matt, thank you both Matt and Tim! Tim, I notice you mention that tractography reports distances as well, which shouldn’t have the same bias as the reported number of streamlines/connection strength — I wasn’t really aware of this and I’ve only been working with connection strengths so far. My network inference procedure has just involved finding the right global threshold on the number of streamlines so far, so any actual distances available to me would be immensely useful w.r.t finding a better global threshold on the number of streamlines. Could you please share how I could obtain these distances? I suppose the files are already generated in my servers where I executed probtrackx, so could you maybe also share what the file names would look like? Regards, Karthik On Oct 6, 2017, at 5:40 PM, Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> wrote: Right, I wasn't very precise in my wording. I was thinking of the "tractography distance bias" as the amount of the bias that is above and beyond the real biological distance relationship. Tim On Fri, Oct 6, 2017 at 4:36 PM, Glasser, Matthew <glass...@wustl.edu<mailto:glass...@wustl.edu>> wrote: It is worth noting that there IS a biological distance bias in connections that has been found with invasive tracers, though the mechanism by which this occurs in in tractography is different from the biological mechanism as Tim says. There’s more discussion of this in the paper I referenced. Peace, Matt. From: Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> Date: Friday, October 6, 2017 at 5:30 PM To: "Gopalakrishnan, Karthik" <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> Cc: Matt Glasser <glass...@wustl.edu<mailto:glass...@wustl.edu>>, "hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>" <hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>> Subject: Re: [HCP-Users] Distance between surface ROIs in MMP Tractography's distance bias is in its reported strengths. The distances reported by tractography should not have a significant bias in the same way - while it takes longer paths less often, it doesn't often take paths that are even more windy and longer than the real path, and it generally can't take a shorter path, right? Since these paths are through the white matter, and generally follow real fiber directions, they are more plausible than any other available method of computing connection distances between areas (3D distance is wrong because connections don't go through CSF, geodesic distance is wrong because long-distance connections aren't transmitted through gray matter the whole way). Moreover, the tractography-reported distances should have a much better relationship to the tractography strength biases. To put it another way, the distance bias of tractography is not a *biological* effect, it is an effect of the *method* of tractography. In particular, the longer a probabilistic streamline gets, the wider the area that it could have hit gets, but much of this area gets intercepted by pieces of cortex before the streamline gets as long as it "should" be - therefore this spreading effect causes long streamlines to be rarer than they should be, by virtue of the streamline length itself (not as a function of the biological tract length). Tim On Fri, Oct 6, 2017 at 4:15 PM, Gopalakrishnan, Karthik <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> wrote: Hi Matt/Tim, My goal is to improve network inference from tractography data by better accounting for the distance bias in tractography, so I want to use some proxy for actual connection distance between ROI pairs. Using tractography itself to account for its own bias against long-distance connections doesn’t make sense to me. Do you have any suggestions on how I could best compute this proxy? Regards, Karthik On Oct 5, 2017, at 8:50 AM, Glasser, Matthew <glass...@wustl.edu<mailto:glass...@wustl.edu>> wrote: Indeed I think we would need to know what you needed the distance for to know how best to compute it. For things like MR artifacts, a 3D distance might be most appropriate. For something like smoothing, a geodesic distance would be appropriate. For something neurobiological, the tractography distance might be most appropriate. Peace, Matt. From: <hcp-users-boun...@humanconnectome.org<mailto:hcp-users-boun...@humanconnectome.org>> on behalf of Timothy Coalson <tsc...@mst.edu<mailto:tsc...@mst.edu>> Date: Tuesday, October 3, 2017 at 6:30 PM To: "Gopalakrishnan, Karthik" <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> Cc: "hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>" <hcp-users@humanconnectome.org<mailto:hcp-users@humanconnectome.org>> Subject: Re: [HCP-Users] Distance between surface ROIs in MMP Since ROIs are not points, distance between them becomes a trickier question. Since areas are connected through white matter rather than gray matter, that also implies that the easy ways to calculate distance may not be all that biologically relevant. This would point to using tractography to find distances. So, I don't think there is an easy answer, sorry. If you want to compute distance along the gray matter anyway, a possibility is to find the center of gravity of each ROI, translate them back to surface vertices (the centers will not actually be on the surface anymore, so you may want to double check them), and then find geodesic distances between those points (you can use -surface-geodesic-distance, running it once per area - you can then get the values from the other vertices near the centers to build the all-to-all matrix a row at a time). Note, however, that this will not give you a distance to areas in the other hemisphere. Tim On Tue, Oct 3, 2017 at 5:06 PM, Gopalakrishnan, Karthik <gkart...@gatech.edu<mailto:gkart...@gatech.edu>> wrote: Hi, I’m working with the Glasser multi-modal parcellation and I’d like to know if there is some prevalent notion of distance between any two surface ROIs in the parcellation? If there is, could you please tell me how I could obtain it or point me to a source? Thanks a lot! Regards, Karthik _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org<mailto:HCP-Users@humanconnectome.org> http://lists.humanconnectome.org/mailman/listinfo/hcp-users _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org<mailto:HCP-Users@humanconnectome.org> http://lists.humanconnectome.org/mailman/listinfo/hcp-users _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
