Model-driven brain shift compensation