A Real-Time Multifunctional Framework for Guidewire Morphological and Positional Analysis in Interventional X-Ray Fluoroscopy

• Published in: IEEE transactions on cognitive and developmental systems 2021-09, Vol.13 (3), p.657-667
• Main Authors: Zhou, Yan-Jie, Xie, Xiao-Liang, Zhou, Xiao-Hu, Liu, Shi-Qi, Bian, Gui-Bin, Hou, Zeng-Guang
• Format: Article
• Language: English
• Subjects: Algorithms; Catheters; Contrast agents; Convolution; Decoding; Deep learning; Elongated structure; Feature extraction; Fluoroscopy; Formability; guidewire; Localization; Morphology; Pixels; Qualitative analysis; Radiation dosage; Real time; Real-time systems; robot-assisted intervention; Robots; Segmentation; segmentation and localization; Task analysis; X-ray fluoroscopy
• ISSN: 2379-8920; 2379-8939
• DOI: 10.1109/TCDS.2020.3023952

In endovascular and cardiovascular surgery, real-time guidewire morphological and positional analysis is an important prerequisite for robot-assisted intervention, which can aid in reducing the radiation dose, contrast agent, and procedure time. Nevertheless, this task often comes with the challenge of the deformable elongated structure with low contrast in noisy X-ray fluoroscopy. In this article, a real-time multifunctional framework is proposed for fully automatic guidewire morphological and positional analysis, namely, guidewire segmentation, endpoint localization, and angle measurement. In the first stage, the proposed fast attention recurrent network (FAR-Net) achieves real-time and accurate guidewire segmentation. In the second stage, the endpoint localization and angle measurement algorithm robustly obtain subpixel-level endpoint and angle of the guidewire tip. Quantitative and qualitative evaluations on the MSGSeg data set consisting of 180 X-ray sequences from 30 patients demonstrate that the proposed framework significantly outperforms simpler baselines as well as the best previously published result for this task. The proposed approach reached {F_{1}} -Score of 0.938, mean distance error of 0.596 pixels, endpoint localization and angle measurement accuracy of 97.8% and 95.3%, and an inference rate of approximately 13 FPS. The proposed framework not only addresses the issues of extreme class imbalance and misclassified examples but also meets the real-time requirements, achieving state-of-the-art performance. The proposed approach is promising for integration into robotic navigation frameworks to various intravascular applications, enabling robotic-assisted intervention.