Ged,
> Okay, but don't you think there may be some dangers here too? For
> example biasing the analysis in favour of easily-segmentable areas
> against regions that are segmented with less accuracy/confidence
> (such as hippocampi)?
>
> And doesn't the comment on "sub-threshold but broad patterns"
> potentially apply to patterns found in the lower probability regions
> -- for example a pattern found around the outer surface of a
> structure?
Sure... this is a good point. That just because a region is low
probability doesn't mean it is necessarily deserves to be tossed out.
I see this happen with PET studies of neurodegenerative diseases; some
regions are so low that they fall below the proportional analysis
threshold. For that setting, the easy solution was to lower the
analysis threshold, but a better solution (which I ended up going
with) was to use MRIcro and hand-edit an explict mask that included
good voxels and excluding non-brain (or possibly out-of-field-of-view)
voxels.
> Having just looked again at some smwc1 images, an absolute threshold
> of 0.05 seems very low anyway*, it includes everything that looks
> like it should be present, along with quite a bit that probably
> shouldn't, so I suppose this is fairly harmless (if perhaps slightly
> ineffectual). However, it didn't seem clear to me that a higher
> threshold would get rid of the correct bits (at least not on my spm5
> smwc1s) so I'm still not convinced that thresholding is desirable...
>
> (*and 0.08 relative to the smwc1 mean happens to be fairly close to
> 0.05 in the images I've looked at, which I expect is by design!)
Again... go explicit... you'll sleep better at night.
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|