中国科学院自动化研究所机构知识库

The ability of real-time instrument tracking is a stepping stone to various computer-assisted interventions. In this paper, we introduce a two-stage framework for real-time guidewire endpoint localization in fluoroscopy images during the percutaneous coronary intervention. In the first stage, in order to predict all bounding boxes that contain a guidewire, a YOLOv3 detector is applied, and following the detector, a post-processing algorithm is proposed to refine the bounding boxes produced by the detector. In the second stage, an SA-hourglass network modified on stacked hourglass network is proposed, to predict dense heatmap of the guidewire endpoints that may be contained in each bounding box. Although our SA-hourglass network is designed for endpoint localization of guidewire, in fact, we believe the network can be generalized to the keypoint localization task of other surgical instruments. In order to prove our view, SA-hourglass network is trained not only on a guidewire dataset but also a retinal microsurgery dataset, and both achieve the state-of-the-art localization results.