The atlases in FSL can be divided into two classes: probabilistic and non-probabilistic. In a probabilistic atlas, the number in each voxel represents the probability that the voxel belongs to a structure. Take the Harvard-Oxford atlas, for example. A voxel at the MNI coordinates 18, -102, 6 has a value of 76. This number means the odds are 76% that the voxel coordinates will be part of the occipital cortex, for a subject taken at random.
In contrast, take a look at a voxel in the frontal areas, which show more variability from subject to subject. At the voxel coordinates 2, 46, 6, the atlas reports that the odds for a random subject are 44% cingulate gyrus and 44% paracingulate gyrus - equal odds. This also shows that structures in a probabilistic atlas can overlap.
A non-probabilistic atlas, on the other hand, does not have any overlap between any of the structures; the voxel values in these atlases represents what structure that voxel belongs to. Take the Juelich Histological DTI-81 atlas for example. Clicking on the MNI coordinates 2, 30, 6 returns a value of 3, which, according to the atlas, means that this voxel belongs to the genu of the corpus callosum. Any voxel with the value 3, as defined by this atlas, belongs to the genu of the corpus callosum; it does not belong to any other structure.
To create masks from a non-probabilistic atlas, add all of the structures from that atlas onto your overlay and note which number corresponds to the structure you want. Assume that we wish to create a mask of the entire corpus callosum from this atlas. We note that the numbers 3, 4, and 5 represent the genu, body, and splenium, respectively. We then highlight the atlas structure in the overlay window and save it as JuelichAtlas.nii.gz.
We then extract only the numbers 3, 4, and 5 by using the -thr and -uthr options of fslmaths:
fslmaths JuelichAtlas.nii.gz -thr 3 -uthr 5 CC.nii.gz
This zeroes out any number below 3 or above 5.
Lastly, we need to binarize the image, again using fslmaths:
fslmaths CC.nii.gz -bin CC_bin.nii.gz
When you overlay this on a template in fslview, you will see that the structures of the corpus callosum all have values of 1. You can then use this mask for an ROI analysis.