Rethinking Hybrid U-Shape Network with Pixel-Level Feature Learning for Retinal Vessel Segmentation

Retinal vessel segmentation is a critical, non-destructive medical imaging task in computer vision, essential for diagnosing fundus diseases. Although deep learning methods dominate this field, existing U-shaped encoder-decoder networks with skip connections face limitations when handling discrepancies in multi-scale features. Shallow encoder and decoder stages produce high-resolution but low-dimensional feature maps, effectively capturing fine vessel details, whereas deeper stages (such as the bottleneck) generate lower-resolution, high-dimensional feature maps rich in semantic information. Traditional U-shaped architectures often struggle to effectively integrate these distinct types of features. To address these challenges, this paper introduces a redesigned U-shaped network that incorporates modified convolution and transformer layers tailored specifically for segmenting slender and tortuous retinal vessel structures. A Multi-Core Channel-Spatial Attention (MCCSA) block replaces conventional skip connections, enhancing the extraction of high-frequency texture features in shallow stages. For deeper stages, a Pixel-level Vision Transformer (P-ViT) is introduced to model semantic interconnections among pixels, thereby improving semantic feature recognition. Furthermore, a Pixel-level residual dynamic adaptive Convolutional Neural Network (P-CNN) is proposed to better capture the intricate curved topology of blood vessels. The proposed method is evaluated on two publicly available benchmark datasets, demonstrating significant segmentation performance improvements compared to existing U-shaped methods. Our contributions include enhanced multi-scale feature integration, improved semantic feature learning, and refined extraction of vessel topology.