[caret-users] ROI definition by connection distance

Wed Jul 11 17:22:17 CDT 2007

This email on connection distance raises a very important topic concerning
cortical "distance" and probabilistic mapping in general.

I was initially very excited to use probabilistic cytoarchitectonic
mapping for my imaging data. I studied carefully the work of the European
group (Zilles, Amunts, Eickhoff, etc.) and the American group (Mazziotta,
Toga, Thompson, etc.) and agreed completely with the majority of their
points and reasons for needing probabilistic anatomy. I was particularly
excited that the approach was disseminated as a toolbox for SPM in Matlab
(Eickhoff et al. 2005).

However, I find a major flaw with the approach as currently implemented,
such that I cannot yet use or recommend the use of probabilistic mapping.
This flaw is suggested in the email on connection distance, and the
essence of the problem is that a 2-D problem has been approached as a 3-D
problem. The consequence is that area 46 (part of Broca's area in the
IFG), for example, has a non-zero probability of landing in the anterior
temporal lobe. This never happens in real brains and we all know that the
"real" cortical distance from IFG to the temporal lobe is much further
than their 3-D proximity suggests.

Amunts writes very disdainfully of "gross" macroanatomy (i.e. lobes,
sulci, and gyri). However, I think this disdain went too far and overtook
reason: the probabilities should be constrained, or conditioned, for
obvious gross anatomical facts, like the fact that a point in the temporal
lobe has no chance of overlapping area 46. Not to mention that points
outside the brain in the scalp or hair should have a zero chance. Consider
also the finding of Van Essen et al. (2005) that variability in the
location of area 17 largely follows variability in the location of the
calcarine sulcus - that is, microanatomically-defined fields do indeed
follow macroanatomy.

Conditioning the probabilities by macroanatomy is one approach, but
ultimately the best approach would be to recognize the problem as a 2-D
one from the get go. That is, the original reference brains (the ~10
post-mortem specimens that undergo cytoarchitectonic mapping and MRI
scanning) should be flat mapped in the manner developed by Van Essen and
colleagues and implemented in CARET. If the probabilistic fields slide
around on a 2-D map  then they would never end up outside of the brain or
in the wrong lobe, only within nearby sulci, for example, as indeed
happens in real brains. (Note: a 2-D approach should use spherical coords
and the "spherical standard surface", not the fully flattened map).

That cortical maps are essentially 2-D objects has been eloquently put
forth for a number of years by Van Essen (Van Essen and Maunsell 1980).
Furthermore, the CARET framework already makes use of macroanatomy. Thus,
I was very excited to hear that CARET had recently implemented
probabilistic cytoarchitectonic mapping, hoping that a 2-D approach or
macroanatomical conditioning might have been implemented. Unfortunately,
that is not yet the case (unless a recent release has addressed this?).
But the necessary data and the CARET framework are poised to implement a
2-D approach, and I look forward to the day when that is in place.

Disclaimer: my particular imaging data (ECoG in neurosurgical patients)
requires mapping individual subjs in a serial case study approach. The
current 3-D approach for probabilistic mapping doesn't work here, but it
is useful on average for typical fMRI/PET studies that first average
across many subjs before relating activations to probabilistic fields. So
those readers using probabilistic mapping for fMRI data should not be
discouraged by the above arguments; but do note that a 2-D approach would
also improve your use of the method.

Erik Edwards
U. Washington

> Interesting. What would you need it for? Do you want to weight the
> correlation between voxels by the their axon distance? Maybe we can
> recruit
> a student who wants to do a project and has programming skills to do this?
>
>
>
> ----------------------------------------------------
>
> Dr. Leon Y Deouell, MD, PhD
>
> Department of Psychology
>
> The Hebrew University of Jerusalem
>
> Jerusalem 91905
>
> Israel
>
> Tel: +972-2-5881739
>
> Fax: +972-2-5825659
>
>  http://pissaro.soc.huji.ac.il/~leon/Lab
>
>
>
>   _____
>
> From: caret-users-bounces at brainvis.wustl.edu
> [mailto:caret-users-bounces at brainvis.wustl.edu] On Behalf Of Alon Keren
> Sent: Wednesday, July 11, 2007 5:57 PM
> To: caret-users at brainvis.wustl.edu
> Subject: [caret-users] ROI definition by connection distance
>
>
>
> Hi.
>
>
>
> I have been using CARET a lot lately, and found it very powerful and
> useful
> for a variety of applications. There is a feature that I am interested in,
> and I think is not available at the moment. I am interested in a
> sophisticated measure of distance between nodes:
>
> The "real" distance, or functional or connection distance between two
> nodes
> is neither euclidean nor geodesic. Two nodes on opposite walls of a sulcus
> are further away in the connectivity sense then two nodes on opposite
> walls
> of a gyrus, with white matter bridging the gap. In the first case a
> geodesic
> measure is suitable, whereas in the second an euclidean measure is more
> adequate. I imagine it is much more complicated to implement, but a
> measure
> of distance that goes around CSF but cuts through WM would be very useful
> and more accurate for functional purposes. For instance I would like to
> estimate the relationship between this functional distance and
> co-activation
> of nodes.
>
>
>
> A question on the same topic:
>
> Is there an automatic way to produce a matrix of node to node geodesic
> distances (of size #ofNodes^2)?
>
>
>
> Thanks,
>
> Alon.
>
> _______________________________________________
> caret-users mailing list
> caret-users at brainvis.wustl.edu
> http://pulvinar.wustl.edu/mailman/listinfo/caret-users
>