Scholarlink

<p>To overcome the limitations of the traditional U-Net architecture in modeling global contextual information and the high computational cost of Transformer models, this paper proposes a novel U-Net variant-LCPA-MSDCE U-Net. The network integrates a Lightweight Channel-Pixel Attention (LCPA) module, a Multi-Scale Dilated Convolution Enhancement (MSDCE) module, and employs a hybrid loss function combining Cross-Entropy, Dice, and Perceptual losses. Specifically, the LCPA module utilizes a parallel architecture, fusing lightweight channel and pixel attention via element-wise multiplication. Compared to the baseline U-Net model, this module increases the parameter count by only 1.04% and FLOPs by 0.46%, with the inference time increasing from 4.64 ms to 5.80 ms, while improving the DSC score by 1.69%, thus achieving a favorable balance between parameters and performance. The MSDCE module uses co-prime dilation rates to expand the receptive field while maintaining compatibility with the U-Net architecture. The hybrid loss strategy, designed to optimize pixel-level accuracy, region overlap, and structural awareness, improves the average DSC score by 1.09% compared to the traditional CE+Dice combination. Experimental results show that the proposed method achieves DSC scores of 84.96%, 91.85%, and 80.40% on the Synapse, ACDC, and MoNuSeg datasets, respectively, achieving a good balance between segmentation accuracy and computational cost while demonstrating strong potential for clinical applications. Meanwhile, its deployment in large-scale and real-time tasks still relies on high-performance computing platforms to ensure efficiency and scalability.</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

Lcpa-msdce UNet: a unet variant integrating lightweight channel-pixel attention and multi-scale dilated convolution enhancement modules for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Yifeng Lin,
  • Mengyu Zhao,
  • Bing Su

摘要

To overcome the limitations of the traditional U-Net architecture in modeling global contextual information and the high computational cost of Transformer models, this paper proposes a novel U-Net variant-LCPA-MSDCE U-Net. The network integrates a Lightweight Channel-Pixel Attention (LCPA) module, a Multi-Scale Dilated Convolution Enhancement (MSDCE) module, and employs a hybrid loss function combining Cross-Entropy, Dice, and Perceptual losses. Specifically, the LCPA module utilizes a parallel architecture, fusing lightweight channel and pixel attention via element-wise multiplication. Compared to the baseline U-Net model, this module increases the parameter count by only 1.04% and FLOPs by 0.46%, with the inference time increasing from 4.64 ms to 5.80 ms, while improving the DSC score by 1.69%, thus achieving a favorable balance between parameters and performance. The MSDCE module uses co-prime dilation rates to expand the receptive field while maintaining compatibility with the U-Net architecture. The hybrid loss strategy, designed to optimize pixel-level accuracy, region overlap, and structural awareness, improves the average DSC score by 1.09% compared to the traditional CE+Dice combination. Experimental results show that the proposed method achieves DSC scores of 84.96%, 91.85%, and 80.40% on the Synapse, ACDC, and MoNuSeg datasets, respectively, achieving a good balance between segmentation accuracy and computational cost while demonstrating strong potential for clinical applications. Meanwhile, its deployment in large-scale and real-time tasks still relies on high-performance computing platforms to ensure efficiency and scalability.