中国科学院自动化研究所机构知识库

In endovascular surgery and cardiology, interventional therapy is currently the treatment of choice for most patients. Robust guidewire detection in 2D X-ray fluoroscopy can greatly assist physicians in interventional therapy. Nevertheless, this task often comes with the challenge of the extreme foreground-background class imbalance caused by the slenderer guidewire structure compared to other interventional tools. To address this challenge, a novel efficient network architecture, termed Fast Recurrent Attention Network (FRA-Net), is proposed for fully automatic mono-guidewire and dual-guidewire segmentation and tracking. The main contributions of the proposed network are threefold: 1) We propose a novel attention module that improves model sensitivity to guidewire pixels without requiring complicated heuristics. 2) We design a recurrent convolutional layer that ensures better feature representation. 3) Focal Loss is reinforced to better address the problems of extreme class imbalance and misclassified examples. Quantitative and qualitative evaluation of various datasets demonstrates that the proposed network significantly outperforms simpler baselines as well as the best previously-published result for this task, achieving the state-of-the-art performance. To the best of our knowledge, this is the first end-to-end approach capable of real-time segmenting and tracking mono-guidewire and dual-guidewire in 2D X-ray fluoroscopy.