A -
>I haven't had problems with structurals so far, on the
>contrary - for 3T structurals, SPM5 seems much more robust
>than 99, 2 or mpr - do you happen to have v large fields of
>view (i.e. uncropped images)?
>
>Just curious - how did you define "better" ;-)?
"better" in terms of the SPM for a well-known contrast (faces vs houses),
which I have performed many times previously, in a group 2nd-level analysis
on 17 different brains. I have not done a proper SPM of the direct
comparison of the two methods, but an informal glance at the two SPMs
reveals many more voxels with higher Z scores after normalisation via
segmentation in SPM5.
Yes, I think the large field of view is likely to be one cause of "standard"
SPM5 segmentation failing dramatically for some subjects; another cause may
be large amount of image inhomogeniety (attenuation bias), possibly due to
our use of parallel coils.
Note also that, as I said below (not very clearly), the large
FoV/attenuation may also be a reason why the SPM2-type normalisation, while
never failing as dramatically as SPM5 segmentation, did not give very
convincing normalisations for some subjects.
A fairer comparison (if I have time!), will be to manually crop and
reposition all the structurals first, and then compare the two types of
normalisation, but at the moment I wanted to get an automated (ie batchable)
pipeline working, which I have done thanks to John's suggestion of running
segmentation twice: first without initial affine coregistration on the
original structural, then secondly with the initial affine component, but on
the attenuatation-corrected image (ms*) output from the first stage. This
hasn't failed yet for ~30 subjects, without requiring manual
repositioning/cropping. (However, perhaps I should stop being lazy, and
achieve even better normalisation by manual intervention in all cases... ;-)
BW,R
>-----Original Message-----
>From: SPM (Statistical Parametric Mapping)
>[mailto:[log in to unmask]]On
>Behalf Of Rik Henson
>Sent: 01 April 2006 15:24
>To: [log in to unmask]
>Subject: Re: [SPM] SPM5 spatial normalization
>
>
>Alexander/Eric -
>
>If it's any help, I can report that for a group fMRI analysis
>on 17 subjects comparing faces vs houses (ie a
>well-characterised contrast), the results in SPM5 were much
>better when determining normalisation parameters via John's
>"unified segmentation", rather than via the more "traditional"
>approach (i.e, the "normalise" button, as in SPM2).
>
>Having said this, the initial affine part of SPM5's
>segmentation sometimes failed on our structurals, unless they
>were manually re-positioned, possibly because of attenuation
>bias or excessive neck (to overcome this automatically, I
>adopted a two-pass segmentation approach, which seems to work
>on ~30 subjects so far).
>
>Presumably for the same reason, some of the normalisation
>parameters determined via the more traditional normalisation
>were clearly wrong, and I must admit that I didn't play around
>with the regularisation, which might have given better results.
>
>Rik
>
>--------------------------------------------------------
>DR RICHARD HENSON
>MRC Cognition & Brain Sciences Unit
>15 Chaucer Road, Cambridge,
>CB2 2EF England
>
>EMAIL: [log in to unmask]
>URL: http://www.mrc-cbu.cam.ac.uk/~rik.henson
>
>TEL +44 (0)1223 355 294 x522
>FAX +44 (0)1223 359 062
>MOB +44 (0)794 1377 345
>--------------------------------------------------------
>
>
>>>-----Original Message-----
>>>From: SPM (Statistical Parametric Mapping)
>>>[mailto:[log in to unmask]] On Behalf Of Hammers, Alexander
>>>Sent: 30 March 2006 11:44
>>>To: [log in to unmask]
>>>Subject: Re: [SPM] SPM5 spatial normalization
>>>
>>>
>>>Dear Eric,
>>>
>>>
>>>It's kind of you to thank me even though I seem to have rather
>>>added to your woes ;-)!
>>>
>>>I haven't really looked at the data with a comparison of the
>>>various SPMs in mind, and I think one would have to think
>>>pretty hard about how to do that - detection rate / yield?
>>>Extent of abnormalities? Overlap of abnormalities with some
>>>predefined idea of where they should be? Effect sizes? False
>>>positives or lack thereof?
>>>
>>>Maybe someone has tackled this with simulations? It sounds
>>>like the kind of thing Matthew Brett would do ;-)?
>>>
>>>As for your reviewer, ok, you may be losing some sensitivity
>>>but if you're involved in ongoing / long-term studies you
>>>might actually have had a reason to have stuck to somewhat
>>>dated software - and the improvements I've seen are moderate
>>>not revolutionary (say 1/15 individuals more in terms of
>>>detection rate) so while it's certainly reasonable to
>>>acknowledge the possibility of reduced sensitivity it seems a
>>>bit harsh to me to trash a study for it (if that's what's
>happened)...
>>>
>>>Good luck, A
>>>
>>>-----Original Message-----
>>>From: SPM (Statistical Parametric Mapping)
>>>[mailto:[log in to unmask]]On
>>>Behalf Of Eric Zarahn
>>>Sent: 30 March 2006 10:53
>>>To: [log in to unmask]
>>>Subject: Re: [SPM] SPM5 spatial normalization
>>>
>>>
>>>Dear Alexander,
>>>
>>>Thanks very much for your answer. I was mainly asking because a
>>>reviewer claimed that by using SPM99 instead of SPM5 one is
>>>reducing the sensitivity of group analyses of functional imaging
>>>data. I could certainly imagine this being possible, but I wished
>>>to know if there was any empirical evidence of this. Again, thanks
>>>very much for providing that empirical substantiation.
>>>
>>>Best,
>>>Eric
>>>
>>>Quoting "Hammers, Alexander" <[log in to unmask]>:
>>>
>>>> Dear Eric,
>>>>
>>>>
>>>> 99 vs 2: yes, for ligand PET (single scans per subject, i.e.
>>>> conceptually the same as an fMRI second level analysis). The
>>>> second (SPM2) study is submitted - if you can wait a little I can
>>>> hopefully let you have a reference.
>>>>
>>>> 5 vs rest: Impressionistically, it's again an improvement (we
>>>> didn't expect otherwise ;-) ) - we don't have hard data yet but
>>>> we have test case datasets (real data with histological
>>>> correlates) which could be used to directly compare if you're
>>>> very keen to have hard evidence or back up the impression with
>>>> some numbers.
>>>>
>>>> All the best,
>>>>
>>>> Alexander
>>>>
>>>> -----Original Message-----
>>>> From: SPM (Statistical Parametric Mapping)
>>>> [mailto:[log in to unmask]]On
>>>> Behalf Of Eric Zarahn
>>>> Sent: 30 March 2006 03:13
>>>> To: [log in to unmask]
>>>> Subject: [SPM] SPM5 spatial normalization
>>>>
>>>>
>>>> Dear John and Group,
>>>>
>>>> Does anyone have any empirical evidence, published or otherwise,
>>>> that
>>>> spatial normalization in SPM5 or SPM2 leads to better sensitivity
>>>> in
>>>> second-level analyses than in SPM99?
>>>>
>>>> Thanks,
>>>> Eric
>>>>
>>>
>>
>
|