Global and Local Attention Network for Automated Retinal Vessel Segmentation

The segmentation of retinal vessels is a pivotal step in the realm of fundus disease diagnosis. Current methodologies, however, often overlook the disparity in semantic content between deep and shallow features. This oversight hampers their ability to concurrently capture the global and local traits of fundus images, notably affecting the segmentation of vessels. To address this, we introduce a global and local attention network (GLANet), which incorporates a global and local attention (GLA) module and a global and local uncertainty (GLU) fusion module. GLA enriches the model by integrating diverse feature scales, while GLU strategically reduces uncertainty in the decoding phase. Evaluated on DRIVE, STARE, and CHASE_DB1 benchmarks, GLANet achieves impressive AUC scores of $98.63 \%, 98.56 \%$, and 98.75%, respectively, marking a significant advancement in retinal vessel segmentation. The results underscore the effectiveness of GLANet in enhancing segmentation accuracy by leveraging both global and local features, thereby providing a robust solution for retinal vessel segmentation tasks.