<p>This paper proposes a novel retinal artery and vein segmentation network, termed Trans-EMAUNet, which integrates a Transformer architecture with a direction-aware Efficient Multi-scale Attention (EMA) module. By leveraging the Transformer’s capability for global dependency modeling and the EMA modules spatial and channel-wise directional sensitivity, the proposed method significantly enhances the representation of both vascular topology and local details, such as tiny vessels, vessel edges, and crossover regions. The network adopts a multi-scale encoder-decoder architecture, enabling the effective fusion of global and local features across different semantic levels. To further improve segmentation performance, a cascaded training strategy is introduced to iteratively optimize arteriovenous segmentation. Experimental results demonstrate that Trans-EMAUNet outperforms existing state-of-the-art methods in terms of accuracy, sensitivity, specificity, and F1 score.</p>