Dear SPM:

Here is a summary of some of the responses I received both on and off the 
list about this issue. Thanks to everyone who responded.  Unedited 
responses are below, and my take on the responses with a summary is at the 
end. Since some people wrote to me off the list I have stripped names.
----------------------------------
        I guess the plan would be to assume that the residual variance is the same 
for both scanners, and simply treat scans from the different scanners as 
different subjects would be treated in a conventional SPM analysis.  This 
would preclude comparisons of e.g. Alzheimers subjects from one scanner 
with controls from another, just as a comparison of condition A scans from 
one subject with condition B scans from another would not be possible.
---------------------------------
        I have dealt with both situations you mention, studies with different 
scanners in the same institution and from different institutions. These are 
PET studies, but the basic idea is the same, the fancy new machine has 
better resolution and acquisition characteristics.
        For a case where we moved to a new scanner on site we were lucky enough to 
have similar numbers of patients and controls scanned with both machines. 
So we did a separate VBM analysis with the data collected by each machine 
to prove to ourselves that the results were going in the same direction on 
each machine and after this proved to be the case we pooled the data and 
analyzed in a standard fashion. We compared mean and standard deviation 
images for the control and patient groups as well as the resulting t-maps 
from the VBM modeling of control vs patient.
        We were already losing resolution by registering to a MNI template, but we 
otherwise would have resampled the higher resolution data into the same 
voxel dimensions as the lower resolution data.
        For the study across institutions it was not so simple. In fact I am still 
wrestling with that data, unfortunately the other  institution has both a 
lower resolution PET scanner and MR scanner. The design of the study is a 
pretreatment scan with multiple follow up scans, thus keeping the data from 
our scans and their scans separate should not affect our results. If we do
normalize the data to a template I will be resampling our data into the 
lower resolution provided by the other group.
-------------------------------------
        What I was told by some experts is that it is all about doing a perfect 
registration to overcome distortion differences between different scanners.
        I tried a gender analysis with equal males/females from two different 
scanners (both Siemens Vision 1.5 T), results were bad and completely 
explained by scanner effects.
        I also did a musician/non-musician comparison using equal numbers from 
both scanners and never got even close to the results that I got when I 
analyzed data from each scanner separately.
        I know that the group at UCLA has not tried this either. If you look 
through their papers, they always use images from one scanner, although 
they are sitting on many datasets from different scanners as well.
------------------------------------
        The Australian UHR group did some structural MRI on scans from two different
scanners:
        Phillips L.J., Velakoulis D., Pantelis C., Wood S., Yuen H.P., Yung A.R., 
Desmond P., Brewer W., McGorry P.D. 2002. Non-reductions in hippocampal 
volume is associated with higher risk of psychosis. Schizophr. Res. 58 145-158.
        There are no great details in this paper, but it may be of use.

[ WE DON'T GET THIS JOURNAL AT OUR LIBRARY SO I HAVE NOT READ THIS PAPER. drg]
-------------------------------------

MY COMMENTS:
        My take on all the responses is that VBM should probably not be done on 
scans from different scanners.

        Is there ever a time when it is possible, however? The only way I can 
think of is if one had equal numbers of the populations to be assessed on 
two different scanners, i.e., 10 patients with disease (a) on scanner A and 
10 on scanner B. Then 10 normals on scanner A and 10 on scanner B. Perhaps 
a non-sphericity correction would be appropriate here since scans on one 
scanner vs. the other might be considered to have some dependence.

        One thing that people didn't mention is that some diseases may actually 
interact with the type of scanner. As an example let's say you were doing a 
study on multisystem atrophy (MSA) vs. normals on 1.5 and 3 T scanners. MSA 
frequently shows signal changes in the basal ganglia partially due to 
deposition of iron. These changes may be more pronounced at 3T because of 
greater sensitivity to inhomogeneity effects. Thus one could get a disease 
by scanner interaction.

Unfortunately, no one had any ideas on how to make this problem go away 
other than reacquiring scans on the new scanner.

Best regards to all,
Darren


-------------------------------------------------------------------------
Darren R. Gitelman, M.D.
Cognitive Neurology and Alzheimer¹s Disease Center
Northwestern Univ., 320 E. Superior St., Searle 11-470, Chicago, IL 60611
Voice:   (312) 908-9023     Fax:  (312) 908-8789
-------------------------------------------------------------------------