Scholarlink

<p>Automatic segmentation of medical images plays a significant role in improving clinical diagnosis accuracy, enhancing treatment plans, and enabling pathological analysis. Although deep convolutional neural networks (DCNNs) are increasingly used in medical image segmentation, the inherent complexity of medical images can sometimes lead to suboptimal segmentation results. To tackle this issue, we designed a DCNN-Transformer network (DT-Net) to capture information. Specifically, we introduced DCNN to mine local information, while utilizing Transformer to capture the relationships between pixels at a global scale to better understand contextual information in the image for more accurate segmentation. In addition, considering that the target objects in medical images may have various shapes and sizes, we designed a four-level feature fusion module. The module includes dilated convolutions with different dilation rates and atrous spatial pyramid pooling, which enables the capture of features at multiple scales within the image, significantly improving segmentation accuracy. To assess the efficacy of our proposed method, we performed experiments using the ISIC 2017 and Kvasir-SEG datasets, comparing it with other leading methods currently available. The experimental outcomes indicate that our proposed architecture excels in segmentation tasks, providing further evidence of its effectiveness and potential in medical image segmentation 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

DT-Net: a hybrid framework of DCNN and transformer for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Xiaoyun Lu,
  • Chunjie Zhou,
  • Yi Lu,
  • Ziyun Zhou

摘要

Automatic segmentation of medical images plays a significant role in improving clinical diagnosis accuracy, enhancing treatment plans, and enabling pathological analysis. Although deep convolutional neural networks (DCNNs) are increasingly used in medical image segmentation, the inherent complexity of medical images can sometimes lead to suboptimal segmentation results. To tackle this issue, we designed a DCNN-Transformer network (DT-Net) to capture information. Specifically, we introduced DCNN to mine local information, while utilizing Transformer to capture the relationships between pixels at a global scale to better understand contextual information in the image for more accurate segmentation. In addition, considering that the target objects in medical images may have various shapes and sizes, we designed a four-level feature fusion module. The module includes dilated convolutions with different dilation rates and atrous spatial pyramid pooling, which enables the capture of features at multiple scales within the image, significantly improving segmentation accuracy. To assess the efficacy of our proposed method, we performed experiments using the ISIC 2017 and Kvasir-SEG datasets, comparing it with other leading methods currently available. The experimental outcomes indicate that our proposed architecture excels in segmentation tasks, providing further evidence of its effectiveness and potential in medical image segmentation applications.