Wolfgang is correct on all counts here. There have been several papers showing noise at highly anisotropic values give negative eigenvalues. Unfortunately, the FA values are not really defined for those cases. GE (I hadn't known this) and DTI Studio change the negative eigenvalues to 0.0. Bob Cox has come up with a modified algorithm that takes into account the noise. I've been working on an update to this software to implement that algorithm.
In the meantime, Wolfgang's suggestion may be the best choice. Do a manual masking of the negative eigenvalues and recalculate the FA.