Scholarlink

In this study, we collected over 20,000 colonoscopy images and annotated four classes of objects exhibiting irregular shapes and edges, including polyps, foam, residue, and ileocecal valves. We validated the effectiveness of various attention mechanisms on this dataset by incorparating them into YOLOv5 and YOLOv8. Through experiments and ablation study, we then introduced a new plug-in channel attention module called Linear Channel Attention (LCA). The LCA module reduces complexity to linear by using broadcasted element-wise multiplication when calculating channel attention. Combined with the spatial attention module (SAM), LCA-SAM is well-adapted to colonoscopy scenarios and effectively improves the model’s detecting performance, especially for objects with variable morphologies.

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

Using Linear Channel Attention to Enhance Real-Time Colonoscopy Object Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Qiwen Le,
  • Lanfang Dong,
  • Yingchao Tang,
  • Derun Kong,
  • Aijiu Wu

摘要

In this study, we collected over 20,000 colonoscopy images and annotated four classes of objects exhibiting irregular shapes and edges, including polyps, foam, residue, and ileocecal valves. We validated the effectiveness of various attention mechanisms on this dataset by incorparating them into YOLOv5 and YOLOv8. Through experiments and ablation study, we then introduced a new plug-in channel attention module called Linear Channel Attention (LCA). The LCA module reduces complexity to linear by using broadcasted element-wise multiplication when calculating channel attention. Combined with the spatial attention module (SAM), LCA-SAM is well-adapted to colonoscopy scenarios and effectively improves the model’s detecting performance, especially for objects with variable morphologies.