Multi-Lateral Branched Network for Tool Segmentation During Robot-Assisted Endovascular Interventions

Robot-assisted endovascular intervention has emerged for improving the outcomes of cardiovascular interventions. However, the current segmentation methods are affected with low and varied contrast values of endovascular tools in the angiogram, and background noise, both of which affect segmentation performance. Thus, surgical scene analytics are characterized with slow tool visualization and response during endovascular navigation. In this study, a multi-lateral branched network (MLB-Net) is proposed for pixel-level segmentation of guidewire in angiograms recorded during robot-assisted cardiovascular catheterization. The network has an encoder with lateral separable convolutions and depth-wise attention, and decoder with improved loss function. Feature maps extracted during end-to-end fully supervised training were optimized for guidewire segmentation. The MLB-Net was trained and validated with multiple angiogram sequences obtained during series of robot-assisted catheterization in rabbit model. Validation studies show a robust performance, characterized with mean IoU of 84.89% and area under curve of 90.64%. In addition, the model offered fast (15.28 frame/second) and reliable segmentation performance in new angiograms obtained during additional trials carried out in pig and human phantom models. Furthermore, we evaluated the MLB-Net by comparing it with existing state-of-the-art networks. Based on our rabbit dataset, the MLB-Net offers better segmentation experience over DeepLabV3+, SegNet, and U-Net which are commonly used for medical image segmentation. Also, MLB-Net generalized well under incremental training. This study contributes a new model for fast tool segmentation, tracking and visualization and during endovascular catheterization.