CT-X-Ray Registration Via Spatial-Projective Dual Transformer Network Fused With Target Detection

Registration of CT-X-rays is crucial in high-precision orthopedic surgery. In this study, a deep learning network integrating convolution and transformer modules is proposed as a model for measuring image similarity for the registration of CT-X-rays. By training the network model to approximate the geodesic distance of Riemann space, the model has the property of convex function, to avoid falling into a local optimum. To further reduce the translation error of registration, this study introduces a spine detection network based on Yolov5, detects the spine of the target image and the image to be registered, obtains the spine position information and readjusts the translation component of the pose. The method used in this study has been tested, and the translation error and rotation error are lower than 3.05 mm and 1.96°, respectively.