Unsupervised nonlinear deformable registration network for 4D CT lung imaging

eformable image registration is crucial in various image-guided radiation therapy applications, particularly in registering computed tomography (CT) lung image modalities during extreme inspiration and expiration phases. The primary challenge is effectively addressing nonlinear deformation between these phases. Advancements in deep learning techniques provide efficient alternatives to traditional approaches, promising enhanced accuracy and computational efficiency. This study presents an unsupervised approach called the Image Patch Decomposition Registration Network (IPDRN) designed for 4D CT lung image registration, which operates independently of ground truth data. The IPDRN effectively learns multi-resolution and multi-scaling features, facilitating more accurate deformation vector field (DVF) computation. The integration of regularization, image dissimilarity and Jacobian determinant loss functions enhances its ability to capture small and complex deformations in image subjects. The effectiveness of the proposed registration method is assessed by comparing the target registration error (TRE) with those produced by conventional algorithms and state-of-the-art (SOTA) unsupervised registration methods.