An Efficient Local-Global Fusion Framework for Accurate Vascular Segmentation

Accurate vascular segmentation is crucial for assisting interventional surgical robots by providing detailed anatomical information for navigation and treatment planning. However, the complex topology of vascular structures and the demand for real-time performance pose significant challenges, especially in clinical environments with limited computational resources. In this paper, a efficient vascular segmentation framework based on a novel Local-Global fusion with Gated attention and Residual connection Block (LGGR-Block) was proposed. This encoder design synergistically combined local detail preservation and global context modeling through a window-based attention mechanism, depthwise convolution, and a learnable gated fusion module. Moreover, channel and spatial attention (CBAM) was integrated to further enhance feature discrimination. The proposed LGGR-Block enabled the network to capture fine-grained vessel structures while maintaining high efficiency, making it well-suited for deployment in resource-constrained robotic platforms. The method was evaluated on three public retinal vessel datasets and one coronary angiography dataset. The performance gains over prior methods validate the effectiveness of the proposed LGGR-Block design and its ability to enhance feature representation in complex segmentation tasks. This made the framework especially suitable for real-time, robot-assisted vascular interventions.