Abstract: Mutual information representing the anatomic features of medical image is computed with feature points．The feature points are extracted from images using region growing and dynamic clustering．During maximization of the mutual information to accomplish image registration，man-machine interaction is used to facilitate this procedure and a ransack method is also adopted to avoid local extrema．Preliminary results on twodimensional robust alignment of MRI-MRI images and MRI-CTimages are presented．The registration strategy presented here might be suitable to clinical applications，especially for images without gray level information．