The similarity metrics of voxel-based image registration are usually Normalized Mutual Information (NMI), and its good registration properties can make images to achieve sub-voxel registration. However, the local extrema and the narrow capture range of global maximum for multi-modal images are easy to cause registration to fail. The structural similarity function has been used to assess image quality. It may reflect the difference of the visual effect and structural information between images, and is associated with the statistical distribution of the voxel gray. When the spatial location between images is changed, the structural similarity also will be changed. We modify this function to make it available for image registration. Simulation results demonstrate that the registration curves of this modified structural similarity (MSSIM) used as a new registration metric show a good convex upward function, and have no significant local extrema. Moreover, the global maximum is located exactly. Especially, the capture range of the global maximum is wide, and hence its robustness is strong. In addition, it is sensitive to strong noise and its operation speed is slow. The metric MSSIM can achieve sub-pixel registration accuracy for three-dimensional image registration even if a 10-parameter affine transformation.