This old response (written shortly after our UK referendum on changing
our voting system) may help here:
Like voting systems, if there are more than two alternatives, there is
no single best approach for picking the winner. To avoid such holes,
you may wish to consider a first past the post system, where you
assign membership based on what class has the largest probability.
However, this may not be so effective if you consider a voxel with a
35% probability of being GM, 25% probability of being WM and a 40%
probability of being scalp. This voxel would be labelled as scalp,
but is more likely to be brain than non-brain.
FSL segmentation incorporates an MRF model, which uses neighbourhood
information to inform class membership. This has the effect of
shifting the belonging probabilities closer to either 0 or 1, thus
reducing the ambiguity.
On 1 June 2011 19:54, Greggory Rothmeier <[log in to unmask]> wrote:
> I'm using SPM8 and 'New Segmentation' to segment the single_subj_T1.nii file
> in the canonical folder. I then import them into matlab and use round() to
> set each element to a binary tissue (1) or no tissue (0). If I do this for
> each of the tissue types and then add them together, I end up with gaps in
> the data. I assume that this is because SPM didn't have enough confidence
> in it being a particular type of tissue or it's a mixture of different
> types. I've also done segmentation with FSL
> (http://www.fmrib.ox.ac.uk/fsl/) and it outputs segmented files that fit
> together like I'm looking for, but it doesn't have as many tissue types. Is
> there a setting within SPM8 to force it to make a decision about the tissue
> type so that the resulting segmented files fit together without gaps?
> Thank you,
Personally, I wouldn't do any thresholding. Instead of counting GM or
WM voxels, I would sum up the probabilities (and multiply by the
volume of a voxel).
On 20 July 2011 07:19, Seyed Batouli <[log in to unmask]> wrote:
> Dear All,
> I have segmented 750 T1 scans of 1.5T scanners, using the “New Segment”,
> SPM8. This is a volumetric study.
> I am going to count the number of GM and WM voxels to calculate regional
> brain volume. But it seems to me that I should perform a thresholding on the
> GM and WM images, since visually I can say some non-zero voxels can never be
> brain tissue. On the other hand, some voxels are regarded both as GM and WM.
> I have tried several thresholds, starting from 5% to 70%, but to me and just
> visually 20% seemed the best. I just wanted please to ask you if you know
> any certain level for this thresholding? My results of regional brain volume
> seem dependant on this threshold level, and although the changes are not
> very significant, since I have collected data from 3 different scanners it
> becomes more important for me.
> Thanks for your response in advance,