Abstract
2D/3D medical image registration is crucial for image-guided surgery and medical research. Single-view methods often lack depth information, while multiview approaches typically rely on large-scale training datasets, which may not always be available in clinical scenarios. This paper explores the feasibility of multiview registration with insufficient data. We propose an innovative multiview 2D/3D image registration model with three key components: a points of interest (POIs) tracking network based on style transfer (POITT), a triangulation layer, and shape alignment based on branch and bound (SA-BnB). POITT constructs a POIs tracking network on Digital Reconstructed Radiographs images (DRRs) and performs points matching on X-ray images via style transfer. The triangulation layer establishes high-precision spatial mapping relationships between X-ray images and the patient. SA-BnB calculates the optimal transformation matrix between the CT volume and the patient. We collected clinical X-ray and CT data from 8 patients and generated 300 pairs of simulated DRRs from two views for each patient to validate the proposed model. The average fiducial registration error was reduced to 6 mm (less than 10 mm is considered successful) with a 96.67% average registration success rate on DRRs. The average SSIM reached 0.77 between X-ray images and the aligned DRRs. All experimental results demonstrate significant advantages over traditional 2D/3D registration. This study may provide a reliable solution for minimally invasive pelvic surgery navigation. The key code is available at https://github.com/TUYaYa1/Points-of-Interest-tracking-network.