关键词: 视网膜血管分割, U-Net, 深度学习, 图像处理

Abstract: Research on retinal vessel segmentation aims to facilitate the early diagnosis and pathological analysis of fundus diseases, providing crucial support for doctors to assess patients?? ocular health. The rapid advancement of deep learning technologies has introduced novel approaches and breakthroughs in the segmentation performance of retinal vessel images. Among these, U-Net has emerged as a mainstream segmentation model in this field due to its outstanding performance. This paper comprehensively reviews recent progress in the application of U-Net and its improved models in retinal vessel segmentation. It firstly introduces commonly used datasets and evaluation metrics for retinal vessel segmentation, then gives an overview of the U-Net model and its primary structural enhancement strategies. Furthermore, the paper categorizes U-Net variants into single-network models and multi-network models. From the perspective of single-network models, it elaborates on improvements such as attention mechanisms, residual structures, multi-scale feature modules, and convolutional modules. For multi-network models, it examines enhancements like cascaded U-Net, dual-path U-Net, the integration of generative adversarial networks (GANs), and the incorporation of Transformer and Mamba models. A comparative analysis is conducted to summarize the improvements and limitations of various studies in terms of model architecture, feature extraction, performance optimization, and experimental results on public datasets. Based on this analysis, the paper discusses current challenges and future prospects in the field.

Key words: retinal vessel segmentation, U-Net, deep learning, image processing