Not completely sure of what is going on, but a few scattered ideas:
1.Cluster these images first (3dClusterize/3dclust/3dmerge/3dkmeans) if the gadolinium is separable from the head. Motion correction on each cluster separately (3dvolreg/align_epi_anat.py/3dWarpDrive/3dAllineate) and then join them back together at the end (3dcalc).
2. Since these are T1 images, you might try a rigid or rigid_equiv alignment to a template (@auto_tlrc/align_epi_anat.py).
3. Skullstrip each volume first (@SSwarper/3dSkullstrip) to separate the brain from the gadolinium or 3dAutomask.