Adaptive deformation decomposition network for unsupervised medical image registration

In recent years, deformable registration techniques have been widely used in medical image analysis. However, most studies assume that the data have been affine pre-processed and focus on large deformable registration on this basis. This reliance on preprocessed data ignores the integration of affine and deformable registration, limiting the applicability and scalability of existing methods to unpreprocessed clinical data. The aim of this study is to investigate the challenges faced by existing methods to directly align datasets without prior affine registration, and to design a registration method that can perform complex deformations and large deformations directly on unprepared datasets. This paper introduces an Adaptive Deformation Decomposition Network (ADDNet) for unsupervised non-rigid registration, capable of decomposing complex deformations across multiple scales in unaligned images. Specifically, we propose a Deformation Decomposition Module (DDM) that captures fine-grained, long-range dependencies and deep semantic information, enabling accurate large displacement capture and detail preservation. To further enhance deformation mapping interactions, we design an Adaptive Contextual Fusion (ACF) module that adaptively fuses global and local semantic features based on regional deformation complexity, thereby improving registration efficiency. Our experiments demonstrate the clear superiority of ADDNet in handling both large and small deformations, particularly on datasets without the affine pre-alignment. ADDNet achieves the highest registration accuracy on the LPBA, ABCT, IXI and Mindboggle datasets, with best-in-class HD95, and MSE metrics. ADDNet provides a more realistic deformation field. These results confirm ADDNet's effectiveness and efficiency in complex registration tasks.