>Question: When using SPM through Medx, the default "gray matter threshold"
>is .80 for PET/SPECT. Is this the recommended level for SPECT analysis?
>Or, is there another recommended level to use?
I agree with the other comments. If you're new to SPM, one of the most
useful features in SPM99 regarding threshold decisions is the mask.img
which gets written out during the analysis and shows you exactly which
voxels were included in the analysis.
Choosing a threshold (which will determine your mask.img) depends a lot on
what you're interested in and which tracer you use; there isn't a
"recommended level for SPECT". The default .8 threshold will probably be
fine for CBF studies. If you lower the threshold further, you include a lot
of voxels outside the brain, your search volume goes up, and sensitivity
and specificity suffer (see Koen van Laere's email). The best way to
quantify this is to look at your voxel and resel count at the bottom of
your result window. Note the relationship is not linear (have a look at
Matthew Brett's excellent web pages by googling him & Cambridge, the term
to look for is "lego brick effect"). The best way of getting a qualitative
impression is to do a few runs with varying thresholds and look at the
resulting mask.images.
We were interested in 11C-flumazenil (PET) binding in the _white_ matter
where signal is low, and this got excluded to varying degrees with the .8
(or even .5) threshold. To get around the problem and look at all the areas
we were interested in but not others, we have used explicit masking. The
rationale and technique are briefly described in Hammers et al., Brain in
press (email me for a draft if you're interested). I have a feeling this
would also give more consistent results than lowering the blanket threshold
in the type of study Emmanuel Stamatakis was talking about.
You can do such masking via the "Full Monty" option, but that's a bit
cumbersome. Fortunately Andrew Holmes has written an extremely useful bit
of code which allows you to use explicit masks with the standard models.
You can find it in message number 3992 in the archives (which contain all
sorts of useful information on this theme anyway).
If you intend to include low signal areas, however, you need to be careful
and check (e.g. with ROI analyses on the normalised & smoothed images) that
the signal you're looking at is far enough away from zero for the
parametric assumptions to hold.
Hope this helps,
Good luck, Alexander
---------------------------------
Dr Alexander Hammers, MD
Department of Clinical and Experimental Epilepsy
National Hospital for Neurology and Neurosurgery/ Institute of
Neurology, UCL
33 Queen Square
London WC1N 3BG
and
Clinical Sciences Centre, MRC Cyclotron Building
Faculty of Medicine, Imperial College of Science, Technology and
Medicine
Hammersmith Hospital
DuCane Road
London W12 0NN
Telephone +44-(0)20-8383-3162
(ext. 2651 or 3704)
Fax +44-(0)20-8383-2029
Email [log in to unmask], [log in to unmask]
|