| the appended graph summarizes Dr. Japee's inquiry of today into the
question of
| affine mapping following your suggestions. It shows the ratio of
| normalized/nonnormalized volumes ('count') and integrals ('sum') as a
function
| of the threshold. Ideally the ratio should be 1,71 (the value of the
determinant
| of the matrix). Basically, we see that volumes are underestimated if the
| threshold is too high; and if you want to count voxels, as we do, you'll
get an
| overestimation of the volume after normalization if the threshold is too
low
| (<0.2), presumably due to 'noise' voxels added by the sinc interpolation.
| Thresholds between 0.4-0.6 appear to be good.
The graph shows pretty much what I would have expected.
|
| For me, this raises two questions about which I'd appreciate your
comments:
| 1) Would another interpolation method be better?
Sinc interpolation is probably the best form of interpolation available
within SPM. However, for spatially normalising segmented GM images, it
may be a little un-necessary in terms of its benefits versus its
computational
cost.
| 2) If you do use sinc, these results suggest to me that thresholding
| segmentation maps at 0.05, as has been suggested on the list, may lead to
the
| inclusion of a lot of unwanted voxels, and that a higher threshold might
be
| better. Comments?
The threshold of 0.05 suggested on the list refers to the GM threshold
entered
during the statistics part of SPM, rather than a way of thresholding the
GM segments to produce binary images. I would suggest either thresholding
at 0.5
(voxels that are more likely to be GM than not GM) or not thresholding and
simply
using the grey matter maps as produced by the segmentation.
A third alternative may be to assume that a voxel is grey matter if its
probability
of being GM is greater than it is of being any of the other tissue classes
in the
model (WM, CSF and non-brain). You can do this with the <ImCalc> button,
selecting
images seg1, seg2 and seg3 (as i1, i2 and i3 respectively). The function
that
would be evaluated would be:
(i1>i2) & (i1>i3) & (i1>(1-i1-i2-i3))
Which roughly translates into "voxels that are more probably GM than WM, and
also more probably GM than CSF, and also more probably GM than non-brain
tissue".
The choice is yours.
All the best,
-John
|