Automatic localization of target vertebrae in spine surgery using fast CT-to-fluoroscopy (3D-2D) image registration

Localization of target vertebrae is an essential step in minimally invasive spine surgery, with conventional methods relying on “level counting” – i.e., manual counting of vertebrae under fluoroscopy starting from readily identifiable anatomy (e.g., the sacrum). The approach requires an undesirable level of radiation, time, and is prone to counting errors due to the similar appearance of vertebrae in projection images; wrong-level surgery occurs in 1 of every ~3000 cases. This paper proposes a method to automatically localize target vertebrae in x-ray projections using 3D-2D registration between preoperative CT (in which vertebrae are preoperatively labeled) and intraoperative fluoroscopy. The registration uses an intensity-based approach with a gradient-based similarity metric and the CMA-ES algorithm for optimization. Digitally reconstructed radiographs (DRRs) and a robust similarity metric are computed on GPU to accelerate the process. Evaluation in clinical CT data included 5,000 PA and LAT projections randomly perturbed to simulate human variability in setup of mobile intraoperative C-arm. The method demonstrated 100% success for PA view (projection error: 0.42mm) and 99.8% success for LAT view (projection error: 0.37mm). Initial implementation on GPU provided automatic target localization within about 3 sec, with further improvement underway via multi-GPU. The ability to automatically label vertebrae in fluoroscopy promises to streamline surgical workflow, improve patient safety, and reduce wrong-site surgeries, especially in large patients for whom manual methods are time consuming and error prone.