YEN-CHING CHANG2026-10-162026-10-162025-01-01Pure ID: 15354115Pure UUID: 765bc430-c368-48d7-a97b-9162b4e6ac4810.1109/iCAST68191.2025.11300236https://www.scopus.com/pages/publications/105038315214https://scholars.csmu.edu.tw/handle/123456789/22818An accurate segmentation of retinal blood vessels is critical for the diagnosis and treatment of various ocular and systemic diseases. Over the past few years, deep learning-based models such as U-Net, SegNet, and HaarNet have demonstrated promising performance in medical image segmentation. However, single-path architectures often struggle with capturing fine vascular structures and balancing semantic and spatial features. Therefore, this study proposes a dual-path segmentation framework that integrates the strengths of the previous three models and investigates multiple fusion strategies, including element-wise addition and channel-wise concatenation followed by 1×1 convolution. A total of 12 dual-path models are constructed and evaluated on the DRIVE and STARE datasets. Experimental results show that the U-Net + HaarNet model using full-stage fusion with channel-wise concatenation achieves the best performance, with a Dice coefficient of 66.04%. All dual-path models significantly all outperforms three single-path models: HaarNet (49.09%), U-Net (52.94%), and SegNet (54.19%). These findings confirm that dual-path architectures with carefully designed fusion strategies can effectively enhance segmentation performance, particularly in capturing small and low-contrast retinal vessels.enjournal article