PRSCS-Net: Progressive 3D/2D rigid Registration network with the guidance of Single-view Cycle Synthesis.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Navigation

Background:

• Accurate 3D/2D image registration is crucial for image-guided spine surgery.
• Conventional methods are slow, and existing learning-based approaches struggle with large misalignments and high computational costs.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for 3D/2D rigid registration of pre-operative CT and intra-operative X-ray images.
• To address limitations of existing methods, including computational expense and poor performance on large misalignments.

Main Methods:

• Proposed PRSCS-Net, incorporating differentiable projection operators for single-view cycle synthesis to overcome limited views.
• Utilized a self-reconstruction path for 3D CT data and performed pose estimation in a shared 3D feature space.
• Implemented a progressive registration path with two sub-networks to handle large misalignments through two-step warping.

Main Results:

• PRSCS-Net achieved state-of-the-art performance on public (CTSpine1k) and in-house (C-ArmLSpine) datasets.
• Demonstrated superior registration accuracy, robustness, and generalizability compared to existing methods.
• Significantly reduced computational complexity and improved handling of large misalignments.

Conclusions:

• PRSCS-Net offers a promising solution for accurate and efficient 3D/2D registration in image-guided spine surgery.
• The method has potential for enhancing clinical spinal disease surgical planning and navigation systems.