Scholarlink

<p><?tk 1?>Early detection of Diabetic Retinopathy (DR) plays a crucial role in preventing permanent vision loss and supporting ophthalmologists in delivering prompt and accurate treatment strategies. However, the segmentation of DR lesions is challenging due to their high inter-class similarity in terms of size, texture, and contrast. To address the limitations of traditional deep learning methods in lesion segmentation, this study proposes the Texture-Enhanced Multi-Scale Interaction Swin-UNet (TEMI-SwinUNet) network. This approach utilizes the Swin Transformer as a feature extraction module and incorporates the Texture Complement Enhancement Module (TCEM) to effectively recover and enhance the perceptual boundaries and texture features lost during downsampling. The Multi-Scale Channel Interaction Module (MSCI) replaces the original bottleneck layer in Swin-UNet, enhancing the network’s ability to extract local spatial features from images, further improving segmentation accuracy. Experiments on the public datasets demonstrate that the model exhibits superior performance in both Dice and AUPR metrics, with a significant improvement in lesion segmentation accuracy. The code is available at <a href="https://github.com/songqiq/TEMI-SwinUNet">https://github.com/songqiq/TEMI-SwinUNet</a>.</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

TEMI-SwinUNet: a transformer-enhanced multi-scale integration network for multi-lesion segmentation in diabetic retinopathy

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Qiqi Song,
  • Guoyun Lian,
  • Jingyu Du,
  • Xinyu Ouyang,
  • Mengting Zhou,
  • Jinfeng Yang

摘要

Early detection of Diabetic Retinopathy (DR) plays a crucial role in preventing permanent vision loss and supporting ophthalmologists in delivering prompt and accurate treatment strategies. However, the segmentation of DR lesions is challenging due to their high inter-class similarity in terms of size, texture, and contrast. To address the limitations of traditional deep learning methods in lesion segmentation, this study proposes the Texture-Enhanced Multi-Scale Interaction Swin-UNet (TEMI-SwinUNet) network. This approach utilizes the Swin Transformer as a feature extraction module and incorporates the Texture Complement Enhancement Module (TCEM) to effectively recover and enhance the perceptual boundaries and texture features lost during downsampling. The Multi-Scale Channel Interaction Module (MSCI) replaces the original bottleneck layer in Swin-UNet, enhancing the network’s ability to extract local spatial features from images, further improving segmentation accuracy. Experiments on the public datasets demonstrate that the model exhibits superior performance in both Dice and AUPR metrics, with a significant improvement in lesion segmentation accuracy. The code is available at https://github.com/songqiq/TEMI-SwinUNet.