Automatic alignment of 3D CT and MRI datasets with the highly deformable and flexible spine regions is challenging but critical for image-guided surgery systems. The proposed system presents a fully automatic 4D registration and fusion system, including a L-SVM vertebra detection, a vertebra localization method, a corresponding landmark detection approach and an elastic 4D registration approach. This method can help to register accurately 3D spine images from CT and MR even they are not taken at the same time or even their anatomical structure is different. In evaluation, a preliminary test conducted to compare nine registration methods with the presented registration approaches, and the top two benchmark methods with high registration accuracies and computing speed selected for full evaluation. In full evaluation, three proposed registration methods were compared with the selected top two benchmark methods using 78 pairs of 3D CT and MRI datasets from two group where each group has 39 pairs. For quantitative evaluation, corresponding landmarks on the CT and transformed MRI datasets are identified by human first and then the distances between the landmark positions are measured as registration errors. The results show that the proposed method is significantly better than the benchmark methods (p ≤ 0.001) in both groups. For the first group, the proposed method obtains the lowest error 12.3624 pixels (4.185mm) on average. In comparison, benchmark methods obtain notably larger mean error distances (41.8085 pixels (14.153mm) and 50.8003 pixels (17.1971mm)).