Dear Geraint,
>Richard's intuition that small areas of fMRI 'activation' may represent
>false positives seems sensible; but I wanted to raise a few possible
>concerns for discussion.
>
>First, is it really true that 'real' biological effects will be spread over
>more than five voxels? In numerical terms, five voxels at typical
>neuroimaging sizes (3x3x3) in gray matter will contain ~13 million neurons.
>In cortical terms, the distance between cortical areas isn't always that
>great; about 3cm (ten voxels) between foveal representation in posterior V1
>and area V5 in humans. And single cortical areas aren't that large either.
>V1 is about 4x8cm in humans so even without cortical magnification 5 voxels
>aligned linearly might represent one fifth of the visual field horizontally.
>But perhaps these concerns are rendered moot because functional imaging
>data, as Richard points out, are typically very smooth anyway?
Except that if one had a priori reasons to expect such spatially
restricted clusters, one would perhaps be careful to scan with a
small voxel size, to allow a reasonable spatial resolution after
smoothing.
>A second concern is that the (apparent) spatial extent of a blob will vary
>with height threshold and so a threshold of (say) five voxels will have very
>different implications at different Z thresholds. So I'm not sure how we can
>justifiably claim that a spatial extent threshold of (say) five voxels is
>grounded in biological plausibility, unless that biological effect is
>determined at a particular height threshold.
>
>So I guess my bottom-line question is: if we are interested in taking into
>account the spatial extent of a cluster in deciding whether to report it,
>shouldn't we be using the cluster level statistics that take into account
>both peak height and spatial extent? Rather than applying an (inevitably
>somewhat arbitrary) spatial threshold to voxel level statistics?
I would imagine that most people would feel uncomfortable with
rejecting clusters which gave high z scores. In my experience the
tiny clusters only just exceed significance, so I don't worry too
much if I have 'cleaned up' the image with a limited spatial extent
threshold (but OK, I admit it, to date I haven't made a formal
decision to do this before analyzing the data).
But, sure, I may well be missing 'unexpected' clusters which
represent a real biologically important effect. But if they are
clusters which only just exceed significance, I would have 'missed'
them anyway since this experiment would be insufficient to convince
me of this real biological effect. If applying a spatial extent
threshold had any effect at all on the way that I discussed the
results, then I would be worried!
>Or is everyone generally comfortable with the small-spatial-extent
>threshold approach?
It's certainly an issue which deserves discussion.
Best wishes,
Richard.
--
from: Dr Richard Perry,
Clinical Lecturer, Wellcome Department of Cognitive Neurology,
Institute of Neurology, Darwin Building, University College London,
Gower Street, London WC1E 6BT.
Tel: 0207 679 2187; e mail: [log in to unmask]
|