This is a matter of empiricism, and depends on artefacts, image
quality etc. It is essentially a model selection issue that involves
selecting the best model for your particular data.
Assessing what works best would be done by comparing model predictions
against ground truth, typically via some form of cross-validation.
Note that assessing what gives the biggest blobs in an SPM analysis is
not a comparison against ground truth.
Typically, you would assess how well the registration model could be
used to predict the contrast images in other subjects. To do this,
you would spatially normalise all the images (not using contrast
images to drive the registration) to give deformations.
Then you would combine the forward and inverse deformations so that
you can overlay the contrast images from N-1 subjects on to the
native-space contrast image of the left-out subject. You can also do
some form of smoothing of the warped scans (but not the native space
one) if you think it improves the match.
Then you would assess the accuracy based on some measure of difference
between the native space contrast of one subject and all the warped
(and possibly smoothed) contrast images of the other subjects.
Then repeat, by overlaying the other contrast images onto another
native space contrast, doing this for all subjects in turn.
Approaches such as Bayesian model selection among generative models
are not appropriate for this kind of analysis, as coregistration,
spatial normalisation and GLM have not been combined into the same
model.
Best regards,
-John
On 4 November 2011 10:17, Peng Syu-Jyun <[log in to unmask]> wrote:
>
> Dear exports
>
> Which is better normalization method between normalizing by using EPI
> templates and by using T1 image unified segmentation? Why?
>
> Best regards,
>
> Syu-Jyun
|