SLKF-UNet: Strip-Enhanced Encoding with Directional Long-Range Dependency and Large-Kernel Skip Fusion for Retinal Vessel Segmentation

Retinal vessel segmentation is a fundamental component in computer-aided diagnosis of ophthalmic diseases, and its accuracy directly affects the automatic screening and quantitative analysis of various conditions such as diabetic retinopathy. Due to the slender tubular structure of blood vessels, segmentation requires two key priors: directional long-range dependency modeling and multi-scale contextual fusion. However, U-Net is limited by the restricted receptive field of 3 × 3 convolutions and insufficient feature fusion in skip connections, making it difficult to adapt to the anisotropic morphological characteristics of vessels. Based on the encoder - decoder symmetric architecture and skip connection mechanism of U-Net, and leveraging its strengths in spatial detail recovery and multi-scale feature fusion, this study further introduces targeted improvements for directional modeling and contextual fusion of fine vessels. A lightweight and efficient improved network, termed SLKF-UNet, is proposed. In the encoder, a strip convolution module (StripBlock) is introduced to establish anisotropic long-range dependencies along horizontal and vertical directions, thereby effectively enhancing the continuity modeling of slender vessels. In the decoder, a large-kernel fusion module for skip connections (LKFBlock) is designed, which incorporates multi-scale large-kernel convolutions at skip connections to achieve efficient fusion of global and local context, thereby compensating for boundary and detail loss during upsampling. To ensure scientific rigor, the standard segmentation protocol of the DRIVE dataset is adopted, and a three-level validation framework consisting of comparative experiments, single-module ablation studies, and positional ablation studies is constructed to evaluate module effectiveness, combined gains, and optimal architecture. Extensive experiments on the DRIVE dataset demonstrate that both modules yield consistent performance improvements, and their combination outperforms the baseline U-Net and other representative improved methods across multiple metrics, including Dice coefficient, mIoU, IoUfg, IoUbg, and overall accuracy, thereby validating the effectiveness of the proposed model. The connect-first, then-refine design paradigm proposed in this study provides theoretical support and a novel technical pathway for the segmentation of slender tubular structures in medical images, and lays a foundation for the clinical application of retinal vessel segmentation.