Fully automated initialisation of 2D–3D image registration

Abstract

The manual interaction required by image-based 2D-3D registration systems can be a major limitation for their use in routine clinical practice. A main source of such interaction is the process of providing an initial estimate for the pose of the 3D data. We propose here a fully automated method for computing this estimate, based on the preoperative computation of 2D templates over a wide range of 3D poses which encompass the majority of standard fluoroscopy gantry poses. These templates are used to apply the Generalised Hough Transform to the intraoperative fluoroscopy image. The sought 3D pose estimate is selected using the generated accumulator arrays and a Gradient Difference Similarity Measure. We show how this method can be incorporated into a vertebra-based, CT-to-fluoroscopy registration system, to provide an initial pose estimate and automatically verify the produced results. The proposed method is tested on data from four clinical procedures with low dose, i.e. low quality, high noise fluoroscopy images, producing correct pose estimates for all 65 images taken within the operating range of interest. The estimates for the 9 images outside this range were correctly rejected by the verification criterion. We believe that the robustness, automation and speed (45 sec computation time) achieved by the proposed method is a significant step to wider clinical use of 2D-3D registration technology.