These surfaces are extracted using cube marching algorithm + Laplacian filtering + post-processing steps from a volumetric BEM segmentation available directly in the MRI Neurodevelopmental Database (NMD) and it fit well to the original BEM segmentation file:
However, when I compare with the MRI directly, I observe something similar than you:
So the problem is apparently with the NMD BEM segmentation for this subject. I can forward this issue to John Richards who is in charge of the NMD. I'll put you in CC.