Scholarlink

<p>Segmentation of retinal vessels is critical for the identification of diseases like diabetic retinopathy and glaucoma but is time-consuming and requires a lot of manual effort. Classic U-Net architecture excels in processing local details but lacks global context awareness, while transformer-based models are context aware but extremely computationally costly. The trade-off between speed and accuracy in segmentation is challenging. We propose a modified U-Net that integrates Swin Transformer blocks and shift windows to enable context-aware vessel segmentation with lower computational costs. It was trained and evaluated on the DRIVE database, comparing with six different U-Net variants (e.g., Residual, Attention, U-net++) and three hybrid models (e.g., Vision Transformer U-Net). The proposed model in this paper achieved 89.12% accuracy and 0.6023 IoU,0.6917 DSC and 0.6466 a +7.92% accuracy and +0.0891% IoU gain compared to the baseline U-Net with only 11.9M parameters used versus 399M for Vision Transformer U-Net. These findings suggest that the Swin Transformer U-Net is a good trade-off between segmentation quality and computational cost and may open the door for future clinical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retinal vessel segmentation using a swin transformer-based encoder-decoder architecture

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Elma Shumsun Kamal,
  • Nusrat Sharmin

摘要

Segmentation of retinal vessels is critical for the identification of diseases like diabetic retinopathy and glaucoma but is time-consuming and requires a lot of manual effort. Classic U-Net architecture excels in processing local details but lacks global context awareness, while transformer-based models are context aware but extremely computationally costly. The trade-off between speed and accuracy in segmentation is challenging. We propose a modified U-Net that integrates Swin Transformer blocks and shift windows to enable context-aware vessel segmentation with lower computational costs. It was trained and evaluated on the DRIVE database, comparing with six different U-Net variants (e.g., Residual, Attention, U-net++) and three hybrid models (e.g., Vision Transformer U-Net). The proposed model in this paper achieved 89.12% accuracy and 0.6023 IoU,0.6917 DSC and 0.6466 a +7.92% accuracy and +0.0891% IoU gain compared to the baseline U-Net with only 11.9M parameters used versus 399M for Vision Transformer U-Net. These findings suggest that the Swin Transformer U-Net is a good trade-off between segmentation quality and computational cost and may open the door for future clinical applications.