Hi Chris and Bing
Fang,
I'd just
like to add that Emmanual Stamatakis presented a poster at HBM2000 In
San Antonio
(S527), "Nonlinear spatial normalisation of SPECT images with SPM99"
and concludes
that using linear plus non-linear options (with 2 x 2 x 2 basis functions),
appears
to
give the least error between template and object image. He used a custom
SPECT
template.
However, my
experience of using the supplied SPM PET template with SPECT images has been
satisfactory - I
explored the difference between a custom SPECT and supplied PET
template
some time
ago and found no major differences between
normalisation results. I know there are
some on this forum
who advocate using a custom made template for very abnormal brains, but we
also used the
supplied
PET
template in a published study of elderly depression and dementia with
great success, and
showed differences
between both diagnostic groups and a group of elderly controls
which were in
agreement with the gold standard ROI analysis.
See:
"A voxel-based
analysis of cerebral perfusion in dementia and depression of old
age"
Ebmeier KP, Glabus
MF, Prentice N, Ryman A, Goodwin GM
Neuroimage. 1998
Apr;7(3):199-208.
Regards -
Mike
--
----------------------------------------------------------------
Mike
Glabus, PhD,
Visiting Fellow in Functional Neuroimaging
Unit on
Integrative Neuroimaging,
Clinical Brain Disorders Branch, NIMH, NIH
Building 10, Rm 4C101, 9000 Rockville Pike
Bethesda, MD, 20892-1365, USA
Tel: + 301 496 7864 FAX: + 301 496 7437
[log in to unmask]
Dr Fang-- you raise several
important questions regarding SPM analysis of SPECT data.
For spatial
normalization, there is no definite answer regarding the use of a custom
template. With the latest version of SPM [SPM99], the algorithm for
spm_spatial has changed considerably, and is much more robust-- especially for
noisy images, or ones whose power spectra are very different from the
templates provided in SPM [both of which are generally true for SPECT data,
and I will leave the consideration of what fan-beam collimation does alone for
the moment]. This newer method can be "regularised"-- restricted so that it
does not warp images in unlikely or bizarre ways-- which means you can try out
levels of regularisation on your images to see how it performs.
Several simple considerations:
you mention Coregister, which is
used for matching functional data [e.g., SPECT] with MRI data. If you have
anatomic scans for your controls, then your task is much easier-- you register
each SPECT scan to its corrsponding MRI, and then normalize the MRI volumes to
the SPM-MRI template [presumably T1 images], applying that transformation to
the coregistered SPECT data as well. This process takes advantage of the much
higher resolution of MRI data. As an aside, I also note that your selection
criterion of "no assymetry >7%"--presumably in a set of predefined
regions?-- appears unusual to me. Unless you have other data indicating that
the range of asymmetry in SPECT scans of the [actual] normal population is
less than 7%, I would say this requirement may produce comparison data that
does not necessarily reflect the normal range.
If you do not
have anatomic images, then you must wrestle with what is "best" for your
reference dataset. One very helpful published study deals with both this issue
and the statistical concerns you raise:
Validation of
Statistical Parametric Mapping (SPM) in Assessing Cerebral
Lesions: A
Simulation Study
E. A.
Stamatakis,* M. F. Glabus,† , ‡ D. J.
Wyper,§ A. Barnes,§ and J. T. L. Wilson*
NeuroImage
10, 397-407 (1999)
Article ID
nimg.1999.0477, available online at http://www.idealibrary.com
in this paper, the authors note that using their
own template of an average SPECT image made no appreciable difference in
spatial normalizatio,compared with using the PET template. Note, however, that
these were Strichman images, and the spatial normalization and statisitcal
analyses were performed in SPM96. More importantly, this paper also addresses
the question of sensitivity-- the ability of SPM [96] to detect a lesion of
known size and degree in a single SPECT image compared to a set of
"normative" scans.
[Note that for this type of single-subject analysis,
you would use "single subject:replication of conditions" as your model, since
in SPM terms this is equivalent to one scan from each of many subjects in only
one condition. ]
the lead author of that paper recently presented other
data regarding SPECT data at the HBM conference in San Antonio [search
"Stamatakis" at http://www.academicpress.com/www/journal/hbm2000/auindex.htm,
or I can send you the abstract as it appears there] which demonstrated an
important concept: the process of spatial normalization can significantly
alter a markedly abnormal signal in a given scan. If you are interested in the
updated results of this work, I suggest you contact Dr Stamatakis. If he does
not reply to this message, I will put you in touch with him.
"How many
subjects" is a difficult question in high-dimensional statistical comparisons,
even with massive univariate tests. Ultimately, you are asking about power, of
course, and you see there have been several questions on this topic of late on
this list. My other reply today on this topic lists some papers on this
subject, and Matthew Brett's reply today also links to some help for
estimating power in SPM96. A related paper which also touches on many of your
concerns is:
Influence of ANOVA Design and
Anatomical Standardization
on Statistical Mapping for PET
Activation
Michio Senda,*
Kenji Ishii,* Keiichi Oda,* Norihiro Sadato,† Ryuta Kawashima,‡ Motoaki
Sugiura,‡
Iwao Kanno,§ Babak Ardekani,§ Satoshi Minoshima,¶ and Itaru
Tatsumi\
NEUROIMAGE 8, 283-301
(1998)
ARTICLE NO. NI980370
I hope this is helpful to you, and
I look forward to additional responses from the community at large--
At
03:46 PM 06/26/2000 , you wrote:
>I am a physicist and try to help NM
physician to set up a protocal for
>analyzing pediatric ADHD patient
using SPM. My knowledge on SPM is very
>basic. I would like to ask your
spmers to comment on the setting I post
>below and help me answer some
questions . We use SPECT (Fan Beam) data.
>1). Collect a control group
of subject under two criterions a).No
>asymmetry >7% with manual 9
pixel analysis by a experienced NM
>physician; b).ADHD is not primary
diagnosis with Axis I standard.
>2). Since these are pediatric patient
(age from 5-18), Should I create a
>average image from the control group
instend of using SPM's PET template
>in Coregister and Reslice? Does it
matter which one to be used as
>modality and target image?
>3).
If such a average image is prefered in my study, should I
average
>on raw .img images or on snr .img images. If I should use snr
.img
>images to do the average, should I use SPM's PET template or
random
>selecting one of patient in my control group as the modality and
target
>image in Coregister and Reslice?
>4). How many patients do
I need to create such an average image under
>statistical consideration?
If the subjects in this control group will be
>used as Group 1 for the
design type "Compare - group; 1 scan per
>subject" in Statictics, how
many patients do I need to reach a
>statistical
satisfaction?
>
>Thank you all in advance.
>
>Bing
Fang, B.M.,M.S.
>
********************************************************
Christopher Gottschalk,
MD *
Assistant Professor of Neurology &
Psychiatry *
Yale School of
Medicine *
*
Mailing
Adress: *
VAMC
[116-A] tel
[203] 932-5711 x4329*
950 Campbell
Avenue
FAX 937-4791*
West Haven, CT
06516 *
********************************************************