Scholarlink

<p>Recently, adversarial learning-based U-Net models have achieved encouraging performance in MRI brain tumor segmentation. However, existing works still have limitations in capturing global dependencies and inter-channel semantic information of brain tumor images. To address these issues, this work proposes a novel cycle generative adversarial Transformer network, i.e., CycTransU-Net, that simultaneously integrates adversarial learning and Transformer U-Net models to improve brain tumor segmentation performance. More specifically, CycTransU-Net adopts part of CycleGAN as the adversarial learning framework, and it constructs an improved Transformer U-Net as a generator to obtain brain tumor segmentation results while utilizing PatchGAN as a discriminator to evaluate the segmentation results. In the generator, a dynamic token sparsification Transformer module is embedded into the U-Net. This module not only removes irrelevant background information in brain tumor images via dynamic sparsification but also establishes long-range dependencies to capture comprehensive global information, leading to improved tumor segmentation accuracy. Additionally, we have combined tied block convolution that facilitates the exchange of semantic information across channels, to optimize the generation of brain tumor images. In summary, our network architecture differs from previous approaches by introducing Transformer modules and bound block convolution in the CycleGAN architecture. CycTransU-Net not only facilitates the exchange of feature information across interactive channels, but also empowers the network to capture a wealth of global feature. We extensively evaluate CycTransU-Net on the public BraTS 2019–2021 and MSD datasets. Through online evaluation for BraTS 2019–2020 validation sets, it achieves dice values of 90.5%/90.6%, 85.2%/85.1%, and 77.6%/77.8% for whole tumor, core tumor, and enhancing tumor segmentation. The results demonstrate the effectiveness and competitiveness of CycTransU-Net compared with state-of-the-art works.</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

Cycle generative adversarial Transformer network for MRI brain tumor segmentation

  • 期刊论文

  • 发表日期：

    2025年3月17日

查看全文
  • Muqing Zhang,
  • Qiule Sun,
  • Yutong Han,
  • Bin Liu,
  • Jun Wang,
  • Mingli Zhang,
  • Paule-J. Toussaint,
  • Jianxin Zhang,
  • Alan C. Evans

摘要

Recently, adversarial learning-based U-Net models have achieved encouraging performance in MRI brain tumor segmentation. However, existing works still have limitations in capturing global dependencies and inter-channel semantic information of brain tumor images. To address these issues, this work proposes a novel cycle generative adversarial Transformer network, i.e., CycTransU-Net, that simultaneously integrates adversarial learning and Transformer U-Net models to improve brain tumor segmentation performance. More specifically, CycTransU-Net adopts part of CycleGAN as the adversarial learning framework, and it constructs an improved Transformer U-Net as a generator to obtain brain tumor segmentation results while utilizing PatchGAN as a discriminator to evaluate the segmentation results. In the generator, a dynamic token sparsification Transformer module is embedded into the U-Net. This module not only removes irrelevant background information in brain tumor images via dynamic sparsification but also establishes long-range dependencies to capture comprehensive global information, leading to improved tumor segmentation accuracy. Additionally, we have combined tied block convolution that facilitates the exchange of semantic information across channels, to optimize the generation of brain tumor images. In summary, our network architecture differs from previous approaches by introducing Transformer modules and bound block convolution in the CycleGAN architecture. CycTransU-Net not only facilitates the exchange of feature information across interactive channels, but also empowers the network to capture a wealth of global feature. We extensively evaluate CycTransU-Net on the public BraTS 2019–2021 and MSD datasets. Through online evaluation for BraTS 2019–2020 validation sets, it achieves dice values of 90.5%/90.6%, 85.2%/85.1%, and 77.6%/77.8% for whole tumor, core tumor, and enhancing tumor segmentation. The results demonstrate the effectiveness and competitiveness of CycTransU-Net compared with state-of-the-art works.