Scholarlink

<p>Medical image segmentation plays a crucial role in clinical diagnosis and treatment. Its accuracy and efficiency are vital for diagnosing, treating, and monitoring diseases. To address the challenges posed by high-resolution and complex medical images, where traditional Transformers may struggle to fully extract diverse global features, we propose a Direct2D Transformer with Linear Angle Attention network for multi-organ medical image segmentation (DTLA-Net). This innovative structure retains the UNet structure. To overcome the computational complexity of combining traditional convolutional neural networks (CNN) and Transformers, we introduce the Direct2D Transformer (D2D-Former) structure. It integrates the advantages of CNN to more effectively extract and process complex features in medical images within the encoding section of the network. Additionally, given that medical images often contain noise and irregular shapes, we incorporate Linear Angle Attention (LA-Attention). This can more effectively capture and utilize the rich detail features in medical images, performing more stably and reliably in regions with complex structures and subtle local variations. Extensive experiments on the Synapse and ACDC public datasets demonstrate that DTLA-Net achieves significant improvements in medical image segmentation tasks, showcasing its potential and effectiveness in handling complex and high-resolution images.The source code is available in <a href="https://github.com/0zeroer/DTLA-Net.git">https://github.com/0zeroer/DTLA-Net.git</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

Dtla-net: a direct2d transformer with linear angle attention network for multi-organ medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年4月23日

查看全文
  • Yepeng Liu,
  • Feiyu Liu,
  • Jinjie Xiao,
  • Xiaofeng Zhang,
  • Fan Zhang

摘要

Medical image segmentation plays a crucial role in clinical diagnosis and treatment. Its accuracy and efficiency are vital for diagnosing, treating, and monitoring diseases. To address the challenges posed by high-resolution and complex medical images, where traditional Transformers may struggle to fully extract diverse global features, we propose a Direct2D Transformer with Linear Angle Attention network for multi-organ medical image segmentation (DTLA-Net). This innovative structure retains the UNet structure. To overcome the computational complexity of combining traditional convolutional neural networks (CNN) and Transformers, we introduce the Direct2D Transformer (D2D-Former) structure. It integrates the advantages of CNN to more effectively extract and process complex features in medical images within the encoding section of the network. Additionally, given that medical images often contain noise and irregular shapes, we incorporate Linear Angle Attention (LA-Attention). This can more effectively capture and utilize the rich detail features in medical images, performing more stably and reliably in regions with complex structures and subtle local variations. Extensive experiments on the Synapse and ACDC public datasets demonstrate that DTLA-Net achieves significant improvements in medical image segmentation tasks, showcasing its potential and effectiveness in handling complex and high-resolution images.The source code is available in https://github.com/0zeroer/DTLA-Net.git.