Deformable Registration Combined with 3D SIFT Matching and Moving Least Squares

Free-form deformation (FFD) is widely used in deformable image registration. FFD uses a regular grid of control points to generate image deformation. Accurate optimization of the control-point displacement relies on an appropriate initial-deformation of the regular grid. In this work, a hybrid registration of landmark-based and free-form deformation is proposed and applied to lung CT images. Corresponding landmark-pairs are detected and matched by 3D SIFT (scale-invariant feature transform). Using the landmark pairs, Moving Least Squares (MLS) is applied for deforming the regular grid. Utilizing the deformed control-point grid, a landmark-constrained FFD registration obtains a final registration result. Since this landmark-based deformation approach can obtain smooth and local deformation of the control-point grid, the proposed registration can directly start with the initialized control-point grid, without the need for any coarse-to-fine processing. It was experimentally demonstrated that the proposed hybrid-registration method outperforms conventioal FFD registration in terms of efficiency and accuracy.