U-Net-based model with integrated SE module and asymmetric convolution for vessel segmentation

The paper introduces SeA-Unet, an advanced retinal vessel segmentation algorithm based on an improved U-Net architecture, designed to address the limitations of traditional models in segmenting complex retinal vessel structures. Analyzing the morphological features of retinal blood vessels is crucial for understanding the onset and progression of various diseases. The acquisition of retinal images generally involves two stages: imaging and segmentation. Among the common retinal imaging tools, Optical Coherence Tomography Angiography (OCTA) and fundus cameras are widely used. However, due to its technological maturity, lower production costs, and ease of use, the fundus camera remains the most widely adopted tool for retinal imaging. Retinal vessel segmentation plays a vital role in the early diagnosis of ophthalmic diseases. Yet, traditional U-Net models still face challenges related to feature redundancy and the difficulty of detecting fine capillaries and branching points. These challenges highlight the need for improved segmentation methods that can address the complexities of small and fine structures. To tackle these issues, SeA-Unet combines multi-scale feature fusion with asymmetric convolution to enhance vascular feature representation, capturing both global context and fine structural details. The integration of the Squeeze-and-Excitation (SE) module further optimizes performance by dynamically recalibrating the importance of features across channels, highlighting key vascular features while suppressing irrelevant information. Evaluated on retinal vessel datasets, SeA-Unet achieves an accuracy of 95.50%, surpassing the baseline U-Net’s 94.85% by 0.65%. This improvement demonstrates SeA-Unet’s enhanced ability to accurately segment fine vascular structures.