Hi Marco,

No, of course, you need the same number of scans for each subject, but if one of the six is messy you just acquire another one, and add only a few minutes to your (and the subjet's) total scan time, as opposed to using one long (multi-echo or some such) sequence which you have to start all over with if you discover at the end that the subject moved somewhere in the middle. No, it does not get blurry after averaging, on the contrary, the images are quite beautiful (although I realize that this is not a valid estimator of S/N).

Best
Andreas

Andreas Meyer-Lindenberg, M.D., Ph.D., M.Sc.
Investigator, Genes, Cognition and Psychosis Program
Unit for Systems Neuroscience in Psychiatry and Neuroimaging Core Facility
National Institute of Mental Health, National Institutes of Health
10-3C101, 9000 Rockville Pike, Bethesda, MD 20892-1257, USA
phone: 301 4969672, fax: 301 4800169, email: [log in to unmask]
web: snp.nimh.nih.gov






-----Original Message-----
From: Marko Wilke [mailto:[log in to unmask]]
Sent: Tuesday, May 16, 2006 8:56 AM
To: Ged Ridgway
Cc: [log in to unmask]; Meyer-Lindenberg, Andreas (NIH/NIMH) [E]
Subject: Re: [SPM] Averaging before VBM - valid ?

Dear All,

interesting discussion evolving here :)

> It seems to me that either: affinely registering the original repeat
> scans and averaging them; or averaging the (s)mwc images, should be
> valid*

I still believe that averaging before segmentation will decrease your resolution as the accuracy of the fit (see below) will be lower than the original resolution of the data.

> I wonder if it might be better to keep all n*k scans in an anova
> model?

Sounds more intuitive to me.

> The contribution of the repeats will be a combination of some SNR
> improvements where they are well-aligned, some extra variance where
> they aren't, and the extra degrees of freedom from having more scans.

Then the question is, does the SNR improvement outweigh the extra variance. In this case, which does without averaging, I would tend to agree with you.

> Statisticians: please shout me down as appropriate!

Yes, please do :)

> *(Marko, responding to your reply, don't you think that if the
> intra-subject registration can't be relied upon then the idea of VBM
> itself is questionable, since it relies on acceptable inter-subject
> registration -- something much harder to achieve!)

True, but spm has never tried to be really good at matching scans in the first place but rather chose to cover up the inevitable inaccuracies by smoothing the hell out of the data. I think it is an approach that has worked well.

Andreas:

> It can be a useful strategy if you have a patient population that has a hard time keeping still, becaues you can discard individual volumes that are messed up.

I see that you had six image volumes per subject, and I believe that averaging those will increase SNR, but does that not result in even visible blurring of structures? Plus, if you discard some volumes in some subjects, does that not systematically bias your analyses w.r.t.
more signal to noise in subjects lying still longer? This is true anyway but if you have more scans in some subjects I would be hesitant to use 6 scans in one subject and 1 scan in the next.

It's interesting to see how different opinions can be, and I would love to hear that somebody has compared different approaches and can tell us "the truth" :)

Best,
Marko
--
=====================================================================
Marko Wilke (Dr.med./M.D.)
                 [log in to unmask]

Universitäts-Kinderklinik University Children's Hospital
Abt. III (Neuropädiatrie) Dept. III (Pediatric neurology)
             Hoppe-Seyler-Str. 1, D - 72076 Tübingen
Tel.: (+49) 07071 29-83416 Fax: (+49) 07071 29-5473
=====================================================================