Fast tracking of catheters in 2D fluoroscopic images using an integrated CPU-GPU framework

Catheter tracking has become more important in recent interventional applications for atrial fibrillation (AF) ablation procedures. It can provide real-time guidance for the physicians and be used for motion compensation by overlaying a 3D left atrium model on live 2D fluoroscopic images. To achieve that, this paper has two main contributions. We first propose a new approach to generate tracking hypotheses based on catheter electrode detection. The novelly-designed tracking hypotheses are evaluated by a Bayesian-framework that fuses learning-based detection and template matching. The second contribution is a novel integrated framework that efficiently distributes computation between a GPU (graphics processing unit) and a CPU. Our framework implements Probabilistic Boosting-Tree (PBT)-based [7] classification for object detection in 2D data on the GPU. Quantitative evaluation has been conducted on a databases of 1073 clinical fluoroscopic sequences. The new framework achieves robust performance with the median error at 0.5mm and the 95^th percentile error at 1.0mm. The speed of tracking the coronary sinus (CS) catheter reaches more than 30 frames-per-second (fps) on most evaluation data. The achieved speed is faster than most real-time fluoroscopy frame rates.