Yes, you should not need 3dfractionize.
I forgot one option to 3dAllineate though. Since you are
warping a mask, you will want nearest neighbor interpolation,
so add the option '-final NN'.
The output values in this dataset should be 0, 1 and 2, the
same values that you started with in +tlrc space.
I don't know where the ~90 comes from. What dataset are
you running 3dROIstats on? What pre-processing was done?
If I had to shoot in the dark I would guess that these are
residual time series with baseline, but not including motion
(or other regressors of no interest).
- rick