Unsupervised Deformable Image Registration Method with Cyclic Consistency

To solve the problem that the reversibility of the deformation field is neglected in deformable image registration, which leads to a decrease in registration accuracy, an unsupervised deformable image registration method with cyclic consistency of deformable field is proposed. Firstly, a deformable image registration frame based on unsupervised learning is designed, which consists of an encoder-decoder for learning image features and a spatial transformation network for generating a sampling grid to guide the accurate movement of the floating image towards the reference image to complete the image alignment. Secondly, a deformation field cyclic consistency loss function is proposed to ensure the consistency of the deformation field during the alignment process. Finally, the deformation field is penalized by combining the Jacobi’s loss function and L₂ norm to ensure the smoothness of the deformation field, which leads to the accurate and realistic deformation field output. Based on the PyTorch framework, the proposed network is evaluated using 2D synthetic and 2D MR datasets. The experimental results show that this method improves the Dice value by 1.77% and decreases the percentage of negative values of the deformation field Jacobi determinant by 35.71% compared with several advanced registration methods, achieving better alignment results.