Reconstructing 3D models from 2D images is an important problem in visualization, particularly in medical and biological visualization. We present a new algorithm for 3D model reconstruction from 2Dimages. In our approach, a 3D atlas is deformed to fit the feature data set in the target images. Compared with existing 3D reconstruction techniques, our approach is more efficient and easier to control. The resultant model has good inherent connectivity and smoothness. The resolution of the final model is determined by the resolution of reference model rather than the dataset. The model generates highly detailed view with a limited number of images. This algorithm has been successfully applied to medical simulation.