Scholarlink

<p>Skin cancer, particularly malignant melanoma, poses a significant threat to global public health. Precise segmentation of skin lesions is imperative for early diagnosis and subsequent treatment planning. In recent years, deep learning has demonstrated remarkable progress in automated skin lesion diagnosis. However, existing methodologies often suffer from excessive parameter counts and high computational complexity, severely impeding their deployment in resource-constrained scenarios, such as mobile health applications. To mitigate this challenge, this paper proposes EMA-U-Net, a lightweight and efficient model based on the U-Net architecture, which integrates multiple attention mechanisms tailored to the visual characteristics of skin lesions. EMA-U-Net achieves high-performance segmentation while maintaining a minimal parameter count (0.062&#xa0;M) and low computational complexity (0.067&#xa0;GFLOPs). On the ISIC 2017 and ISIC 2018 datasets, the proposed model exhibits superior performance with mIoU scores of 80.11% and 80.73%, respectively. This model provides an effective solution for resource-limited environments, striking an optimal balance between segmentation accuracy and resource consumption. The source code is publicly available at <a href="https://github.com/KWang0217/EMA-UNet">https://github.com/KWang0217/EMA-UNet</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

EMA-U-Net: efficient multi-attention U-Net for skin lesion segmentation

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Lei Ma,
  • Kai Wang,
  • Junfeng Wang,
  • Dangguo Shao,
  • Sanli Yi

摘要

Skin cancer, particularly malignant melanoma, poses a significant threat to global public health. Precise segmentation of skin lesions is imperative for early diagnosis and subsequent treatment planning. In recent years, deep learning has demonstrated remarkable progress in automated skin lesion diagnosis. However, existing methodologies often suffer from excessive parameter counts and high computational complexity, severely impeding their deployment in resource-constrained scenarios, such as mobile health applications. To mitigate this challenge, this paper proposes EMA-U-Net, a lightweight and efficient model based on the U-Net architecture, which integrates multiple attention mechanisms tailored to the visual characteristics of skin lesions. EMA-U-Net achieves high-performance segmentation while maintaining a minimal parameter count (0.062 M) and low computational complexity (0.067 GFLOPs). On the ISIC 2017 and ISIC 2018 datasets, the proposed model exhibits superior performance with mIoU scores of 80.11% and 80.73%, respectively. This model provides an effective solution for resource-limited environments, striking an optimal balance between segmentation accuracy and resource consumption. The source code is publicly available at https://github.com/KWang0217/EMA-UNet.