I suggest to first compute a fully parameterized model of the fiber populations in each voxel,
and then to use appropriate per-fiber measures, such as the FA of a fiber.

This is a link to our 2012 paper on how to do it (Scherrer, Warfield PLOS One 2012):
http://www.plosone.org/article/info%3Ado...

and our 2013 MICCAI paper on what to do if you only have a single shell HARDI (Taquet, Scherrer et al. MICCAI 2013):
http://link.springer.com/chapter/10.1007...