Medical image registration of lung large deformation with hybrid cost volume

Deformable image registration of lung CT remains challenging due to the coexistence of large global displacements and local sliding deformations induced by respiratory motion. Existing multi-level registration methods based on cost volumes usually adopt a unified search strategy across different resolution levels, which makes it difficult to simultaneously achieve robust global motion perception and accurate local anatomical alignment. To address this issue, we propose a hybrid cost volume network for large deformation lung CT image registration. At the low- and middle-resolution levels, an axis-guided sparse cost volume (ASCV) is introduced to organize and aggregate the correlation responses of three-dimensional candidate displacements according to the principal anatomical axes. By preserving the complete three-dimensional search range while reducing the dimensionality of the explicit correlation representation, ASCV facilitates the modeling of large-range respiratory motion. At the high-resolution level, a standard dense cost volume (CV) is employed to provide richer local correlation information, thereby supporting the alignment accuracy of fine anatomical structures. In addition, we systematically analyse the search radius and correlation representation strategy of ASCV on registration performance. Experimental results on two public lung CT datasets demonstrate that proposed method achieves favorable registration performance compared with the evaluated deep learning-based methods while producing smooth and anatomically plausible deformation fields.