Adding onto Paul's useful advice - the probabilistic atlases are generally not thresholded at all. The probabilities are typically a fractional overlap of some sort. The overlap is based on the template target for the multiple subjects and the alignment methods used to move all the subjects. The numbers actually used may be from 0-1, 0-100, 0-250, 0-79 for 79 subjects, ..., almost anything the atlas developer thinks is appropriate. In AFNI, each sub-brick volume within the dataset contains the probability map for a particular region. and the subbrick label and atlas_points_list attribute in the header should reflect which region is associated with it.

These kinds of atlases are often simplified into a stricter definition for each region with a "maximum probability map" or MPM atlas. At each voxel, the region that has the largest overlap (the maximum probability) "wins" that voxel. We typically don't look at the actual maximum probability, but the maximum probability is not necessarily a very large fraction. Typically, there is a minimum threshold like 5 or 10%, but that varies across atlases. The lower the threshold, the larger regions will tend to be where there is little overlap with neighboring regions.

In AFNI, there are a variety of probabilistic atlases available - the Desai atlases, the Eickhoff-Zilles cytoarchitectonic atlas, the Haskins Pediatric atlas, the Princeton Visual atlas, the Ventro-prefrontal Cortical atlas and the UNC infant atlases. These can also be made for surface analysis like the Princeton atlas. Those are included in the links Paul left and a little more specifically in these links:

[afni.nimh.nih.gov]
[afni.nimh.nih.gov]