Dear everyone,
thanks a lot for the tips. I disabled the brain mask
and the problem with the incorect segmentation in
upper regions was fixed. What happens in the
algorithm, if I disable the brain mask. I read the
paper (Multimodal Image Coregistration and
Partitioning ..) but I can't find a hint concerning a
brain mask ? Does any more detailed paper exist ?
I like to do a quantitative analyze of the segmented
data over the time. That is why I do need preferably
the same segmentation results for datas of the same
patient acquired at different times.
Gregor
Medical School Heidelberg
--- John Ashburner <[log in to unmask]> schrieb:
> I took a look at the tiff file that you sent with
> the last email,
> and the problem does look like it is due to the
> affine
> registration. The segmented image is reduced in
> size because
> the prior probability images do not accurately
> overlay onto the
> image to segment.
>
> One way of making this part more robust would be to
> disable
> the brain masking, which is found under the spatial
> normalisation
> part of the defaults.
>
> The remainder of the email answers the previous
> questions ...
>
> | I tried the spatial normalization and it works
> well.
> | Although I do not know, how I can disable the
> | nonlinear wraping ?
>
> This can be done via the defaults button (spatial
> normalisation, estimation), where you would specify
> the number of DCT basis functions as [0 0 0].
>
> | The segmentation is not generaly bad, only
> segmented
> | areas in upper regions are incorect. I do not
> think,
> | that any function is not working, because there
> are no
> | error messages during segmentation. Exists any
> | limitation for segmenting ?
>
> The images segment best when you can get a good
> registration between them and the prior probability
> images. If the head is a slightly different shape,
> then
> the affine registration may not model the overall
> head shape
> that well. This is particularly important if the
> contrast
> is not so high, in which case, the segmentation
> relies more
> on the prior probability images.
>
> | I use SPM'99 to segment 3-d gradient-echo
> datasets.(TR
> | = 19ms, TE = 4,4 ms, flip = 8, FOV = 23 cm,
> thickness
> | 3mm, gap = 0, slices = 60). I get good results,
> when I
> | use the PD-weigthed template. Usually the
> segmentation
> | is almost perfect. Sometimes (on some heads) the
> | segmented areas are much too small, especialy in
> upper
> | regions.
>
> On some images, the intensity is lower further away
> from the
> centre of the magnetic field. This can be
> problematic to
> the segmentation, which is why SPM99 has an optional
> intensity
> nonuniformity correction built into the
> segmentation. Although
> this usually works well, it sometimes has problems
> because the
> image contrast can be lower further away from the
> centre of
> the magnet.
>
> | I tried quite a lot to find the reason
> | for that but I can't find a systematic. I have a
> | couple of ideas: Does the ratio between FOV and
> head size
> | have any influence on the segmentation ?
>
> This can have a small effect on the segmentation.
> When assigning
> probabilities for the voxels belonging in the
> different clusters,
> the segmentation uses Bayesian rules. Part of these
> rules base
> the belonging probabilities on the number of voxels
> that the cluster
> is estimated to contain.
>
>
> | How does
> the
> | position of the head (i.e. rigth-left) influence
> | the quality of the segmentation ?
>
> This should not have any specific effect.
>
> Best wishes,
> -John
>
>
__________________________________________________________________
Do You Yahoo!?
Gesendet von Yahoo! Mail - http://mail.yahoo.de
Yahoo! Mail auf Ihrem Handy? - http://mobil.yahoo.de
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|