基于生成对抗网络的椎骨关节约束配准方法研究OA

Objective:To propose a 2D/3D medical image registration method based on generative adversarial network(GAN)for improving spatial alignment accuracy of images in minimally invasive spinal surgery navigation.Methods:A registration framework for spinal preoperative 3D CT and intraoperative 2D X-ray images was constructed based on the LIDC-IDRI dataset.Deformation parameters conforming to vertebral joint constraints were generated using a generative adversarial network.Digitally reconstructed radiographs technology was employed to project preoperative 3D CT data onto a 2D plane to generate synthetic X-ray images.A stage-wise parameter decoupling method was adopted for iterative parameter optimization.2D/3D medical image registration was achieved.The performance of the proposed algorithm was validated using Elastix as the benchmark method.Results:The GAN-based registration method outperformed the open-source medical image registration framework Elastix in terms of mean absolute error(MAE),normalized cross-correlation(NCC),and normalized mutual information(NMI).The GAN-based registration achieved an MAE that was only 52.4%of that of the open-source medical image registration framework Elastix,while its NCC and NMI were 10.2%and 42.8%higher than Elastix,respectively.Conclusions:By synergizing adversarial learning with gradient optimization,the GAN-based registration method achieves high-precision and robust intraoperative image registration.