A Spatial Registration Method Based on Point Cloud and Deep Learning for Augmented Reality Neurosurgical Navigation

Abstract

BACKGROUND: In order to achieve spatial registration for surgical navigation, a spatial registration method based on point cloud and deep learning is proposed. METHODS: Neural networks are used to register medical image point clouds and patient surface point clouds to complete spatial registration of surgical navigation. An image processing method is designed to convert medical images into point clouds, and a structured light robot is used to extract patient surface point clouds. RESULTS: Coarse registration was conducted through a neural network, followed by fine registration using the ICP algorithm, achieving a rotational registration error (RRE) of 0.961° and a translational registration error (TRE) of 0.118 mm. In phantom experiments, the surface registration error was 0.622 mm, and the target registration error was 0.748 mm. CONCLUSIONS: The proposed spatial registration method based on point cloud and deep learning improves the accuracy and efficiency of neurosurgical navigation.