It is extensible if you have the training set. A more recent criticism is that it only uses the curvature of the cortical surface to do its work, whereas more recent atlases such at the Glasser HCP use multimodal data.

Exactly this, I just want to use it with Glasser or the like but it's not really feasible. And since DKT just doesn't capture what we need, FreeSurfer gets left behind.

What does “extensible with different atlases” mean?

The idea predates modern usage, back to 1967: https://en.wikipedia.org/wiki/Talairach_coordinates, where they were trying to define a common way of referring to structures.

In this freesurfer sense it is the result of a training set, where an expert has delineated cortical and/or subcortical structure on a set of one or more images. If you have multiple images and you map them all into the same atlas space, you can derive a probabilistic model from this, or collapse it to a highest likelihood map or whatever you want.

The cortical and subcortical labeling algorithms used in freesurfer use such a (probabilistic) atlas as one of the inputs. GP is wishing that they could replace this with a new, presumably improved, version.

it's like what? 10-15 years now?