DVAP-Reg: Dual-view anatomical prior-driven cross-dimensional registration for spinal surgery navigation.

2D-3D cross-dimensional registration serves as a critical technology in spinal surgery navigation, with profound implications for enhancing surgical precision, reducing radiation exposure and mitigating surgical risks. Its core objective is to achieve visual navigation by accurately aligning preoperative high-resolution 3D vertebrae with intraoperative 2D X-rays. However, this technology remains constrained by prominent challenges, primarily arising from the inherent semantic and dimensional discrepancies. Traditional registration methods, typically relying on classical iterative optimization strategies, suffer from low computational efficiency. Meanwhile, deep learning-based approaches struggle to accommodate the spatial randomness inherent in intraoperative 2D X-rays and 3D vertebrae. In this paper, we propose a dual-view anatomical prior-driven cross-dimensional registration method for spinal surgery navigation. First, a direct regression network based on dual-view X-rays, integrated with a proposed spatial correlation mechanism, is employed to enhance geometric consistency constraints and mitigate inter-patient anatomical variability. Then, corresponding vertebrae's anatomical priors, extracted via the proposed Face-GCN module as conditional information enhancement for high-level, generalized spatial perception, are fused into the regression network for spatial pose alignment guidance. Finally, a clinical cross-dimensional image dataset is released using the developed interactive registration platform. The proposed network has been validated in real-world spinal surgical navigation scenarios across diverse lumbar spine pathologies, utilizing authentic intraoperative 2D X-rays and preoperative 3D CT images. In these clinical settings, our method achieves real-time 2D-3D image-assisted navigation with rotational and translational accuracy of 2.81° and 1.82 mm, demonstrating its ability to keep pace with the time-sensitive, ever-changing nature of intraoperative workflows. Our project's dataset and source code are available at: https://github.com/TMMU-KLPOP/DVAP-Reg.