Multi-View Depth-Aware Rigid 2-D/3-D Registration

Overlays of CT images onto X-ray images are used in minimally invasive procedures to provide navigation assistance. Image registration is usually required to align the images. Registration using a single view is often not sufficient, especially for a limited field of view. Therefore, multi-view registration is used. Recently, a depth-aware 2-D/3-D registration framework was proposed and shown to achieve high accuracy and robustness for single view registration. In this paper, we extend this method to multiple views by reformulating the point-to-plane correspondence model. We perform experiments on a publicly available cerebral angiography dataset as well as on clinical spine data. Furthermore, we investigate the effect of the viewing direction and the angle between the views to achieve optimal registration results while minimizing the obstruction to the clinical workflow. For the native 2-D images of the angiography dataset, our method achieves a success rate (SR) of 99.22% and an accuracy of 0. 21mm, outperforming the baseline method (SR of 90.73 % mm and accuracy of 0.23 mm). For registration of single vertebrae, our method shows increased robustness (SR of up to 95.44%) compared to a “naive” multi-view extension. We demonstrate that registration robustness reaches a plateau for 30° between views for single vertebra registration. Our evaluation also shows a dependency of registration accuracy and robustness on the actual viewing direction.