Hi Heath
yes, I think it's a bad idea for thickness just as it would be for VBM. ADNI is a bit of a special case as they went to great lengths to make their acquisitions uniform across scanner platform, but I would still look to see if there is a significant scanner effect in analyzing that data.
cheers Bruce
On Tue, 11 Sep 2012, Heath Pardoe wrote:
Hi all, I was wondering if there was a consensus view on the optimal way to approach morphometric analysis (cortical thickness, volumetry & associated analyses) of MRI data acquired on multiple scanners.
From VBM literature & our own experience it appears that comparing MRI scans of patients from one scanner with controls scanned on another is a bad idea, because there are systematic differences that will manifest as very significant GM volume/concentration/density differences following statistical analysis when these differences are unlikely to really exist ie. we get a lot of false positives.
Definitive findings on similar analyses using cortical thickness mapping seem harder to come by. There's the Han study from 2006 and Dickerson 2008 but both of these seem to be empirical analyses of mean absolute error or related measures rather than pitting scanner 1 subjects against scanner 2 subjects in a statistical analysis (please correct me if I haven't looked hard enough). Wonderlick et al 2009 is closer in terms of statistical analysis but these were done on one scanner so don't really address the multicenter issue.
To my mind it seems likely that cortical thickness analysis would have the same problem as VBM and would require controls to be scanned on the same scanners as patients to account for inter-scanner variability. Does anyone have any opinions on this? Or, even better, any data so we can tell how much of a problem this is?
In a related question, the ADNI study uses a phantom to correct an image for gradient nonlinearities in different scanners, and then pools this corrected data from multiple scanners into each control or patient group, and subsequent morphometric analyses are done without worrying about the multi-site issue. Does anyone have an opinion on the validity of this approach?
I guess one way to investigate it would be to look at vertex-wise across-subject variance in coregistered cortical thickness maps both with and without the phantom-based correction procedure; if the variance was lower in the phantom-corrected variance map this would provide good evidence that it's worth the effort of undertaking this procedure. I would be keen to do this analysis but unless I'm mistaken the ADNI website only provides images that have already been processed to correct for gradient nonlinearities.
An alternative approach that has been proposed is to scan a group of the same controls on multiple scanners to characterise between-scanner differences. The practicality of this approach breaks down as the number of centers increases. Again I would be interested to hear if anyone has experience in using this approach compared to others.
In summary, there seem to be three ways to overcome inter-scanner variance:
- scan different controls at each site
- scan a phantom at each site (and probably scan controls too anyway)
- scan the same group of controls at all sites
I would be interested in hearing which approach people think is best, and why.
Thanks if you're still reading! Best wishes Heath