Abstract: Multi-model non-rigid medical image registration plays an important role in diagnosis work. The approach most widely used is by marking control points manually first and then calculating the transform coefficients. The method however is poor in efficiency and accuracy. A new control point selection approach based on Hierarchical Mutual Information and Thin-plate Spline is presented in this paper. By the approach the image is divided into pieces, and then two control points are chosen with specific rules, to make automatic registration. Experiments show that it is an effective non-rigid registration approach.