Scholarlink

<p>For the traditional Unet model, directly connecting features with different depths is prone to semantic gaps, and thin blood vessels cannot be recognized by the network, resulting in missing segmentation. To solve this issue, we propose an updated TransMA-Unet with a U-shaped network for segmenting retinal vessels. On the one hand, the encoder incorporates a Transformer module. Since the data volume of the eye dataset is relatively small, the Transformer module is more suitable for datasets with a small number of data points. At the same time, an attention mechanism module is added to the jump connection. The new Transformer and attention modules make full use of the feature information between the upper and lower layers, addressing the problem of incomplete use of feature information and missing segmentation. On the other hand, a multi-scale residual connection block (MSR) is introduced to the decoder, which reduces gradient disappearance and explosion in the network, increases the model’s generalization ability, and improves segmentation performance. The proposed architecture is evaluated using metrics such as the dice coefficient, sensitivity, specificity, and accuracy on the open-source database DRIVE. The results show that the proposed approach is beneficial, especially for extracting peripheral vascular anatomy. Using the proposed architecture, the DRIVE has a dice coefficient of 0.8630, sensitivity of 0.9506, specificity of 0.9884 and accuracy of 0.9790. The dice coefficient, sensitivity, specificity, and accuracy of TransMA-Unet increase by 7.67%, 3.41%, 1.84%, and 1.14%, respectively, when compared to the original Unet, demonstrating that TransMA-Unet has better all-around performance.Additionally, the suggested TransMA-Unet network performs and achieves more than the technologies already in use, and the proposed method is promising in other vascular segmentation. The generated codes, models, and the mobile application are available for enthusiastic users at <a href="https://github.com/sunshine-elf/TransMA">https://github.com/sunshine-elf/TransMA</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

TransMA-Unet: transformer and multi-scale residual connection attention network for vessel segmentation

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Lihong Zhang,
  • Chongxin Xu,
  • Yingbo Liang,
  • Yuzhuo Li,
  • Tong Liu

摘要

For the traditional Unet model, directly connecting features with different depths is prone to semantic gaps, and thin blood vessels cannot be recognized by the network, resulting in missing segmentation. To solve this issue, we propose an updated TransMA-Unet with a U-shaped network for segmenting retinal vessels. On the one hand, the encoder incorporates a Transformer module. Since the data volume of the eye dataset is relatively small, the Transformer module is more suitable for datasets with a small number of data points. At the same time, an attention mechanism module is added to the jump connection. The new Transformer and attention modules make full use of the feature information between the upper and lower layers, addressing the problem of incomplete use of feature information and missing segmentation. On the other hand, a multi-scale residual connection block (MSR) is introduced to the decoder, which reduces gradient disappearance and explosion in the network, increases the model’s generalization ability, and improves segmentation performance. The proposed architecture is evaluated using metrics such as the dice coefficient, sensitivity, specificity, and accuracy on the open-source database DRIVE. The results show that the proposed approach is beneficial, especially for extracting peripheral vascular anatomy. Using the proposed architecture, the DRIVE has a dice coefficient of 0.8630, sensitivity of 0.9506, specificity of 0.9884 and accuracy of 0.9790. The dice coefficient, sensitivity, specificity, and accuracy of TransMA-Unet increase by 7.67%, 3.41%, 1.84%, and 1.14%, respectively, when compared to the original Unet, demonstrating that TransMA-Unet has better all-around performance.Additionally, the suggested TransMA-Unet network performs and achieves more than the technologies already in use, and the proposed method is promising in other vascular segmentation. The generated codes, models, and the mobile application are available for enthusiastic users at https://github.com/sunshine-elf/TransMA.