First of all, in which space is your ROI? Have you processed it at all?

I guess that your FA image is in the MNI space if you followed the TBSS processing. Also, your original ROI image should be in the subject's native space but acquired in a different run than the DTI acquisition. Did you have an anatomical scan before the MRS? If yes, then you can use "flirt" command to align that anatomical scan to the MNI space. Then you can use this transformation matrix to transform your ROI from the native space to the MNI and thus be aligned with the FA image. Just remember to use "nearestneighbour" interpolation for the ROI, since you want to preserve the binary values.

It might be sensible for you to wait an expert's input on this, since it could be better to register first both the DTI and MRS scans in the same anatomical image before doing any further processing.

After having done the above registration, getting the FA value from this areas is easy. If you need all the skeleton voxels that fall inside your ROI, then run:

fslmaths "ROI" -mul mean_FA_skeleton_mask "output"

And you can use that output as a mask for randomise call.