It seems to me that the problem is essentially that the tensor estimated is not positive-definite – i.e. there are negative eigenvalues. This would explain the black glyphs being displayed in these locations. dwi2tensor uses a weighted least-squares fit, which doesn’t constrain the tensor parameters to be positive-definite. If dtifit imposes such a constraint, that would be enough to explain the discrepancy. You could verify with tensor2metric dt.mif -num 1,2,3 -value ev.mif, see if any of the eigenvalues at that location are negative.
As to what might be causing these issues: these typically boil down to the b=0 signal being lower than the DW signal in certain directions. This can often happen due to Gibbs ringing for example: voxels near CSF can often have lower b=0 signal than anticipated because of this, and this is often enough to cause problems of this nature. If you’re not already, you could try using mrdegibbs before your motion and eddy-current correction (i.e. dwipreproc), see if that helps.