Optical Flow with Learning Feature for Deformable Medical Image Registration

Deformable medical image registration plays a vital role in applications, such as placing different temporal images at the same time point or modality into coordinate system. Various strategies have been developed to satisfy increasing needs of deformable registration. One popular method is estimating displacement field by computing optical flow between two images. The motion (flow field) computed based on either gray-value handcrafted descriptors scale-invariant feature transform (SIFT). These methods assume that illumination constant However, may not always this assumption. In study, we propose metric learning-based estimation called Siamese Flow for We train learners using network, which produces an patch descriptor guarantees smaller distance similar anatomical structures and larger dissimilar structures. proposed framework, features close real deformation due excellent representation ability network. Experimental results demonstrate outperforms Demons, SIFT Flow, Elastix, VoxelMorph networks regarding accuracy robustness, particularly with large deformations.