Scholarlink

<p>Skin scar is a prevalent dermatological concern that impacts both aesthetic appearance and psychological well-being, making precise delineation of scar tissue essential for clinical treatment. To address the challenge of scar image segmentation, this study introduces an innovative deep learning framework integrating CNN and Swin Transformer architectures. The proposed model leverages a multi-scale feature fusion module to combine hierarchical representations from both backbones, while a novel multi-pooling channel-spatial attention mechanism enhances feature refinement during skip connections. Comprehensive experiments demonstrate the model’s superior performance in scar segmentation, achieving metrics of 96.01% Accuracy, 77.43% Precision, 90.17% Recall, 71.38% Jaccard Index, and 83.21% Dice Coefficient, which compare favorably with mainstream methods, and our model performs well in all metrics, highlighting its potential for clinical adoption in scar analysis.</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

A dual encoder network with multiscale feature fusion and multiple pooling channel spatial attention for skin scar image segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月2日

查看全文
  • Weiyuan Yang,
  • Xiaolin Wang,
  • Guangwei Chen,
  • Jianming Wen,
  • Dexing Kong,
  • Jianfeng Zhang,
  • Xinyang Ge,
  • Hao Xu,
  • Jianhua Qin

摘要

Skin scar is a prevalent dermatological concern that impacts both aesthetic appearance and psychological well-being, making precise delineation of scar tissue essential for clinical treatment. To address the challenge of scar image segmentation, this study introduces an innovative deep learning framework integrating CNN and Swin Transformer architectures. The proposed model leverages a multi-scale feature fusion module to combine hierarchical representations from both backbones, while a novel multi-pooling channel-spatial attention mechanism enhances feature refinement during skip connections. Comprehensive experiments demonstrate the model’s superior performance in scar segmentation, achieving metrics of 96.01% Accuracy, 77.43% Precision, 90.17% Recall, 71.38% Jaccard Index, and 83.21% Dice Coefficient, which compare favorably with mainstream methods, and our model performs well in all metrics, highlighting its potential for clinical adoption in scar analysis.