Linear registration is an essential first step for image registration. However, linear registration often fails when the brain shapes, locations, orientations of the target and template images are severely different. To solve this problem, we proposed a knowledge-based approach, in which a large number of MR images were prepared as intermediate images, which were semi-automatically registered to the template a priori to ensure accurate registration. A new target image was first registered to all intermediate images and best intermediate image was selected based on a goodness-of-fit metric. The final transformation was then calculated by combining the pre-determined intermediate-to-target transformation.