2D/3D fast fine registration in minimally invasive pelvic surgery

The 2D/3D rigid registration between preoperative 3D CT and intraoperative 2D X-ray is a crucial step in minimally invasive pelvic surgery. The deep learning-based 2D/3D registration methods address the inefficiencies of traditional approaches. However, the wide range of spatial transformation parameters and other complexities pose significant challenges for achieving accurate registration in a single step. Additionally, the stylistic differences between Digitally Reconstructed Radiographs (DRRs) used in training and real X-ray images limit the practical applicability of most methods. To overcome these challenges, we propose a 2D/3D fast registration framework comprising a coarse registration network, fine registration based on key point tracking and alignment, and domain adaptation. Coarse registration using plug-and-play attention is introduced to preliminarily estimate transformation parameters. Then we design a key point tracking network to match key points between different images, and leverage points alignment to achieve fine registration. To address the stylistic differences between DRR and X-ray images, we investigate a domain adaptation network. The experiments were conducted on DRR and X-ray images, respectively. Our method achieved a mean absolute error of 0.58 on DRR and a structural similarity of 78% on X-ray, outperforming baseline methods. Extensive ablation studies demonstrate that fine registration and domain adaptation significantly improve registration performance.