Scholarlink

<p>Pain intensity assessment is crucial for monitoring patient discomfort. While video-based automated pain evaluation has emerged as a promising research area, it faces challenges in effectively capturing facial features and addressing class imbalance. We propose an attention-driven channel–spatial fusion network (ADCSFNet) and a geometric-semantic fusion framework (GSFF) to address these issues. The ADCSFNet comprises a channel subnetwork, utilizing hierarchical residual blocks with exponential channel expansion and hybrid attention modules for feature re-weighting, and a spatial subnetwork, employing a multi-granular processing framework with hierarchical attention mechanisms to aggregate local and global features. The GSFF augments data by integrating facial landmarks and anatomical descriptions into textual prompts, utilizing depth-guided image generation. Evaluation on the UNBC-McMaster shoulder pain expression archive database demonstrated that ADCSFNet outperformed existing methods, achieving a mean absolute error of 0.34, mean squared error of 0.53, and Pearson correlation coefficient of 0.85. This study contributes to advancing automated pain intensity assessment by enhancing feature representation and mitigating class imbalance.</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

Enhancing pain intensity evaluation via an attention-driven channel–spatial fusion network

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Feng Gao,
  • Linbo Qing,
  • Lindong Li,
  • Ge Yang,
  • Risheng Xu,
  • Li Gao

摘要

Pain intensity assessment is crucial for monitoring patient discomfort. While video-based automated pain evaluation has emerged as a promising research area, it faces challenges in effectively capturing facial features and addressing class imbalance. We propose an attention-driven channel–spatial fusion network (ADCSFNet) and a geometric-semantic fusion framework (GSFF) to address these issues. The ADCSFNet comprises a channel subnetwork, utilizing hierarchical residual blocks with exponential channel expansion and hybrid attention modules for feature re-weighting, and a spatial subnetwork, employing a multi-granular processing framework with hierarchical attention mechanisms to aggregate local and global features. The GSFF augments data by integrating facial landmarks and anatomical descriptions into textual prompts, utilizing depth-guided image generation. Evaluation on the UNBC-McMaster shoulder pain expression archive database demonstrated that ADCSFNet outperformed existing methods, achieving a mean absolute error of 0.34, mean squared error of 0.53, and Pearson correlation coefficient of 0.85. This study contributes to advancing automated pain intensity assessment by enhancing feature representation and mitigating class imbalance.