Sparse Branch-and-Bound 2D-3D Registration Based on Prior Correspondences of Dense Points

In Image-Guided Intervention, 2D-3D registration between pre-operative 3D Computed Tomography (CT) or Magnetic resonance imaging (MRI) images and intra-operative X-ray 2D images is a key step for accurate surgical navigation. The goal of 2D/3D registration is to find the best rigid transformation parameters that best align the 3D offline CT and 2D online X-ray images. The multi-modality 2D/3D registration is a challenging problem, partly due to the different dimensions and modalities. To handle the issues, we propose a novel globally Branch-and-Bound (BnB) method incorporating point set neural network, to speed up the estimation of transformation. The pre-trained Deep BlindPnP model can provide global and local prior information for the rotation and translation search in the BnB framework. And different densities of point sets are used in different cases, balancing the capture global information and computation time to find the correspondence. The proposed registration framework combing learning-based and optimization-based methods have been comprehensively validated on real-world datasets, which demonstrate significantly improved accuracy and speed.