A fully automatic and non-rigid hierarchical registration separation ROI (Region of Interest) method based on optimal Atlas image search and local weighted B spline transform is proposed. The experimental results show that the accuracy of the registration of the proposed algorithm is 95.6%, the normalized mutual information value is 1.8432, the root mean square error is 1.12%, and the correlation coefficient is increased by 18.33%. Compared with other registration methods, the registration accuracy and precision of this registered method have obviously improved, which is of great significance for clinical assistant diagnosis.