An Improved Image Segmentation Model based on U-Net for Interventional Intravascular Robots

Robot-assisted intervention vascular procedures have become an important research focus in recent years. The robotic systems involve separate devices designed for master-slave operation, and this greatly reduces exposure of interventionists to radiation. However, image-based visualization and tracking of endovascular tools have received minimal efforts. In this study, an end-end deep learning model is developed for segmentation of guidewire in X-ray angiograms acquired during robot-assisted intravascular catheterization. The encoding module in original U-net was adopted while a simplified decoding network module was employed for binary pixel classification. For this purpose, custom isomorphic master-slave robotic system and C-arm cone beam computed tomography (CBCT) imaging system were employed for auricle-to-femoral cardiac catheterizations done in rabbits, while a binary dataset including X-ray angiograms was recroded and used for training the improved U-Net model. Validation of the modified U-Net model shows it performs better than other semantic segmentation networks. These include MIoU and feedforword speed (Ffs) of 2.23% and 113.5%, respectively; these are higher than those achieved by the original U-Net. This study provides a new binary dataset for guidewire segmentation.