CycleMorph: Cycle consistent unsupervised deformable image registration

Image registration is a fundamental task in medical image analysis. Recently, many deep learning based methods have been extensively investigated due to their comparable performance with the state-of-the-art classical approaches despite ultra-fast computational time. However, existing still limitations preservation of original topology during deformation vector fields. To address this issues, here we present cycle-consistent deformable registration, dubbed CycleMorph. The cycle consistency enhances by providing an implicit regularization preserve deformation. proposed method so flexible that it can be applied for both 2D and 3D problems various applications, easily extended multi-scale implementation deal memory issues large volume registration. Experimental results on datasets from non-medical applications demonstrate provides effective accurate diverse pairs within few seconds. Qualitative quantitative evaluations fields also verify effectiveness method.