I think your strategy is sensible. FA will be underestimated by the single-tensor model, but MD might be more stable.

With a single shell of DWI, there's no unique solution to the multi-tensor parameters:

http://crl.med.harvard.edu/publications/...