Scholarlink

<p>Deep learning architectures based on Convolutional Neural Networks (CNNs) have demonstrated remarkable success in medical image segmentation, particularly with encoder-decoder frameworks such as U-Net. While U2-Net, with its nested U-Net architecture, exhibits outstanding multi-scale feature extraction capabilities, it faces challenges in precisely segmenting small and low-contrast targets in medical images. To address these limitations, we propose a novel encoder-decoder architecture, DPSAU2-Net(Depthwise Progressive Split-Attention U2-Net). This model incorporates a Residual Depthwise Separable U-block (RSDPU) to enhance local feature texture extraction while integrating a lightweight Split-Attention block (SA block) to improve global feature utilization efficiency and inter-channel feature selection. Extensive experiments on three public datasets–CVC-ClinicDB, 2018 Data Science Bowl, and Lung Image Segmentation–demonstrate the superior performance and generalization of DPSAU2-Net. It outperforms TransUNet by up to 9.1% mIoU and 6.7% F1-score on CVC-ClinicDB, 2.2% and 3.4% on the Data Science Bowl, and achieves 0.935 mIoU and 0.966 F1-score on the Lung dataset. These results demonstrate the effectiveness and potential applicability of DPSAU2-Net in medical image segmentation tasks.</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

Enhanced U2-net with residual depthwise separable and split-attention blocks for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Kai WANG,
  • Jianbin WU,
  • Zhenghao CHEN

摘要

Deep learning architectures based on Convolutional Neural Networks (CNNs) have demonstrated remarkable success in medical image segmentation, particularly with encoder-decoder frameworks such as U-Net. While U2-Net, with its nested U-Net architecture, exhibits outstanding multi-scale feature extraction capabilities, it faces challenges in precisely segmenting small and low-contrast targets in medical images. To address these limitations, we propose a novel encoder-decoder architecture, DPSAU2-Net(Depthwise Progressive Split-Attention U2-Net). This model incorporates a Residual Depthwise Separable U-block (RSDPU) to enhance local feature texture extraction while integrating a lightweight Split-Attention block (SA block) to improve global feature utilization efficiency and inter-channel feature selection. Extensive experiments on three public datasets–CVC-ClinicDB, 2018 Data Science Bowl, and Lung Image Segmentation–demonstrate the superior performance and generalization of DPSAU2-Net. It outperforms TransUNet by up to 9.1% mIoU and 6.7% F1-score on CVC-ClinicDB, 2.2% and 3.4% on the Data Science Bowl, and achieves 0.935 mIoU and 0.966 F1-score on the Lung dataset. These results demonstrate the effectiveness and potential applicability of DPSAU2-Net in medical image segmentation tasks.