Dear SPMers,
we would like to use SPM for segmentation, eg to correct MRS data for csf. As novices to SPM we encounter some problems when trying to
segment 30 T1-weighted images of 256 x 256 x 64, 1 x 1 x 2 mm resolution, acquired at 3 Tesla. We tested different combinations of input: the
normalization option is always run; no, little or lots of inhomogeneity correction; no smoothing or smoothing with a kernel of 3 3 3 or 3 3 6 or even
10.
1. The relative proportions of gm, wm, and csf (mean values) depend on the input options used and are 0.156 - 0.179 (gm), 0.086 - 0.108 (wm)
and 0.064 - 0.091 (csf), generally with smoothing increasing the value for csf and imhomogeneity correction decreasing that for wm. Unless we
smooth with a kernel of 10 the gm/wm ratio appears unreasonably large; can this be connected with (i) that the intensity distribution in the images
at 3 T (with longer T1 and shorter T2) differs from that at 1.5 T where the template has been acquired (which I actually do not believe!) or (ii) that
the lower part of the cerebellum is missing in the images (only 12.8 cm total slice thickness)? or (iii) what else?
2. When we extract the data for orthogonal regions of interest, say, a cube of 20 x 20 x 20 mm3 containing the hippocampus, the results are even
more dependent, eg for csf the differences are up to 300 %, those for wm 200 %. Aside from the uncertainty regarding the results to be taken as
the most reliable (with no quality criterion at hand) the use of such results for csf correction will hardly improve the data to be corrected ...
Any idea how to interpret these findings?
Thanks and regards, Florian Schubert
|