1) Using a lesion mask in segmentation should only affect normalization
parameters, it should not set any voxels in the normalized image to zero.
The mask image should be 0 in the lesion and 1 everywhere else, be in
register with the image to be normalized, and is specified in the GUI under
Custom->Masking Image. If your segmented image is zeroed in the lesion, it
sounds like you've applied the mask by hand to the original structural image.
2) Implicit masking should only apply when a voxel is zero for all subjects
(i.e. where all patients have lesions). Voxels are excluded if a single
subject has the value 'NaN', but this applies in 2nd-level fMRI analysis
where NaN is the result of implicit masking in the 1st-level analysis; I'm
not aware of any way segmentation would set a structural/GM voxel to NaN.
Setting voxels to zero may be used to selectively exclude values from
damaged tissue on a subject-by subject basis in a 2nd-level fMRI analysis
using a one-sample t-test vs. zero. As Stephen pointed out, using this
approach in a two-sample t-test is problematic. One solution may be, in each
voxel, to set values for patients with damage to the mean over patients with
that voxel intact. This approach of selective exclusion may be preferable to
excluding all voxels where one or more patients has damage, as you keep more
information, but it would be tricky to implement. Strictly, you'd need to
vary the degrees of freedom in each voxel, but using DF for the whole group
would at least be conservative and should not introduce false positives.
Hope this helps,