One of the important thing is that the model assumptions hold for your data.
If there are artifacts (other than those that are corrected by the bias
correction) then the data is not so good. Similarly, if the resolution is
low, then you will probably encounter more partial volume (which is not
modelled).
It is not enough to look only at the standard deviations. You also need to
consider the mean intensities of the different classes. If you rescale an
image by 0.001, then the standard deviations will be rescaled by the same
amount. You also need to consider the differences between the class means.
I guess that one way of getting an idea about how well the contrast separates
the classes is to look at a histogram of the values in e.g. a seg1 image. If
they are all very close to either zero or one, then you have pretty good
separation. If the data had not been digitised to 8 bit, then they could be
logit transformed ( http://mathworld.wolfram.com/LogitTransformation.html ).
t=log(p./(1-p));
Then the variance of t could be used as an indication of how well separated
the intensities are. Here is a little demo to try to illustrate this:
% Simulate some random (transformed intensities)
t=randn(10000,1)*8-1;
% Inverse transform
p=exp(t)./(exp(t)+1);
% A histogram - supposed to resemble that of a seg1 image
hist(p,256);
Best regards,
-John
> Has anyone looked into how to decide objectively on the performance of
> various structural sequences for VBM? I have read over the recent paper by
> Deichmann for example and he cites improvements in CNR and SNR, but to some
> extent the evidence for the goodness of a sequence is visual ("it looks
> good"). Are there any tools for quantitating this goodness? Is there a
> measure for deciding when the bias field or inhomogeneity is minimized?
>
> Let's say i have a T1 sequence and a T2 sequence. I segment each to their
> respective templates and find that the T2 sequence shows narrower Standard
> deviations for all three compartments (gray, white and CSF) in the
> segmentation printout. Does this suggest that this sequence was better
> because it produced better separated tissue clases?
|