2-D/3-D Medical Image Registration Based on Feature-Point Matching

Abstract

Two-dimensional/3-D medical image registration has a wide range of applications in intraoperative image-guided navigation, which can not only assist surgeons in accurately locating lesions but also serve as a key link for surgical robots to locate the surgical site. Current methods for 2-D/3-D spine image registration are susceptible to getting stuck in local optimization, struggling to extract gradient information from noisy real data, and exhibiting slow processing speeds. Recently, deep learning methods have suffered from insufficient training data, poor generalization performance, and a tendency to produce incorrect solutions. We propose an optimized model that significantly improves the speed and accuracy of 2-D/3-D registration by deeply integrating a feature-point matching network. This network demonstrates exceptional robustness in processing high-noise imagery and is adept at coarse registration, providing the initial solution for the optimized model and thereby abbreviating the time required for coarse registration. It also facilitates updates of parameter location modules within the optimized model, diminishing the overall computational demand. Additionally, by harnessing grayscale and spinal feature information, we formulate an objective function enriched with a feature-point similarity metric to govern the descent trajectory, culminating in heightened precision and expedited convergence. Our empirical findings indicate that this method achieves a mean accuracy of 0.2550 mm on real data, substantiating the efficacy of our approach.