Friday, March 12, 2010

Mis-localization of fMRI activation at the temporal-parietal boundary

In the many fMRI studies we have now done looking at sensory-motor integration for speech and related functions we have repeatedly observed that sensory-motor area Spt (i) shows up more robustly in individual subjects than in group averages and (ii) localizes to the posterior planum temporale region in individual subjects but sometimes localizes to the inferior parietal lobe in group averaged normalized data.

The results below are an example of this (data represent the work of my graduate student, Lisette Isenberg). This is a study involving 17 subjects performing a listen->covert repeat task (~shadowing) minus a listen_only condition. Data were warped to Talairach space following typical procedures and projected onto a template brain (left image below). Notice that the focus of activation is relatively small and squarely on the supramarginal gyrus. When we looked at individual subjects, however, there was no activation in the SMG but rather it was consistently ventral to this in the Sylvian fissure.

The image on the right uses ANTS (Advanced Normalization ToolS) which performs a diffeomorphic transformation to align the images based on anatomical features. The structural image you see here (right) is actually the 17 brains of the participants nicely aligned to each other. With their sulci and gyri neatly aligned, the activation in the right image is now substantially larger (functional regions are lining up better) and correctly localized to the posterior Sylvian region.

What seems to happen is that the standard template brains have a Sylvian fissure that is sometimes too flat compared to sample population, which can slope up more at the posterior boundary. Activations at the posterior margin then, as in Spt activations, end up mis-localizing to the parietal lobe. ANTS alignment seems to solve the problem though.

For this reason, anytime someone reports an SMG or inferior parietal lobe activation, I question whether it might actually be in the posterior Sylvian fissure.


Anonymous said...

can you give us a bit more detail of the standard normalisation you used? which software package?

Greg Hickok said...

I believe it was done using AIR.

Kenny Vaden said...

I'm not sure which statistical package was used, but I helped develop the ANTS normalization for this when it was just a few subjects (N=2-4). We used pretty heavy regularization even then, to get the subjects into tight spatial alignment. The challenge was that the sylvian fissures were all over the place; pretty normal distribution, I've since observed.

This result raises a lot of questions about standard normalization and displaying results. There are still too many papers with group-level results superimposed on the "Colin brain". (Myself included, once.)

This sort of result also raises interesting questions about incidental spatial smoothing in group results. That is, after combining subjects, to what extent does cluster size reflect unaligned structures (since averaging across misaligned boundaries will effectively spread out group-level clusters).

More things to keep neuroimaging researchers up at night...