in the end I use a Gaussian Mixture Models (GMM) to obtain a segmentation and then build the 5tt image myself.
If this is something that you find useful, and the relevant processing is available at the command line, you could write a new
5ttgen algorithm wrapping that functionality
does it change something if for mostly of the subjects I use fast segmentation algorithm and for few of them the GMM?
Depends entirely on your downstream analysis. As a generic argument, performing different processing for different subjects always has the potential to be a problem when trying to make statistical inferences. You could at least include this variable as a nuisance regressor. Conversely, if all you’re doing is generating and visualising streamlines, then it doesn’t really matter at all. So it depends on where within that spectrum your intentions lie.