When DTI data sets are scanned other than transverse on a Philips system (coronal or sagittal) the bvec that is produced by DCM2Nii is sometimes wrong (i think when the slices are scanned double oblique).
The overall distribution of the gradient directions is correct however they are all rotated towards a wrong direction. This happens both for multi frame/enhanced and classic DICOM.

I have several data sets that show this behavior and can reproduce this with different versions of DCM2NII.
 
The rotation of the gradient directions that are stored in the Philips dicom header should be done based on the ImageOrientation tag.
This method should be independent for each acquisition orientations, which does not seem to be the case.
I suspect one of the columns of the rotation matrix based on the ImageOrientation are permuted.