Deformable image registration is a fundamental technique in computational neuroanatomy. An iterative multilevel block matching technique with the use of several recent inventions is proposed here. A symmetric multimodal similarity measure allows to register subject images to an arbitrary digital brain atlas. Smooth deformations produced by scattered data interpolation based on compactly supported radial basis functions suppress gross inter-subject differences and preserve the localized anatomical variability which may be further studied with selected automated morphometry methods. Four similarity measures are tested in an experiment with image data obtained from the Simulated Brain Database and a quantitative evaluation of the algorithm is presented.