Scholarlink

<p>Accurate detection of surgical instruments is crucial for precise control in robot-assisted minimally invasive surgeries. Surgical instruments often operate in close proximity to sensitive tissues, necessitating real-time pose information acquisition. In this paper, we present YOLOv10-BFDA, an enhanced detection framework that incorporates bidirectional feature fusion and dual attention mechanisms. The bidirectional feature fusion enables multiscale feature interactions, enhancing the network’s ability to identify targets of varying sizes. The backbone network integrates channel and pixel attention mechanisms to focus on fine-grained critical information. Additionally, an adaptive multi-scale attention module dynamically weights features based on spatial and channel distributions. Experimental results on the M2CAI16-Tool-Locations dataset show that YOLOv10-BFDA achieves 96.5% mAP@0.5 and 62.9% mAP@0.5:0.95, demonstrating improved precision and robustness compared to conventional methods, particularly in multi-object and multi-scale surgical instrument detection tasks. Code can be available at https://github.com/haoyingjys/YOLOv10-BFDA.</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

Enhanced YOLOv10 with bidirectional feature fusion and dual attention for precise surgical instrument detection

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Ying Hao,
  • Wenjie Wang,
  • Xiaohua Wang,
  • Shixiao Guo,
  • Yan Shi

摘要

Accurate detection of surgical instruments is crucial for precise control in robot-assisted minimally invasive surgeries. Surgical instruments often operate in close proximity to sensitive tissues, necessitating real-time pose information acquisition. In this paper, we present YOLOv10-BFDA, an enhanced detection framework that incorporates bidirectional feature fusion and dual attention mechanisms. The bidirectional feature fusion enables multiscale feature interactions, enhancing the network’s ability to identify targets of varying sizes. The backbone network integrates channel and pixel attention mechanisms to focus on fine-grained critical information. Additionally, an adaptive multi-scale attention module dynamically weights features based on spatial and channel distributions. Experimental results on the M2CAI16-Tool-Locations dataset show that YOLOv10-BFDA achieves 96.5% mAP@0.5 and 62.9% mAP@0.5:0.95, demonstrating improved precision and robustness compared to conventional methods, particularly in multi-object and multi-scale surgical instrument detection tasks. Code can be available at https://github.com/haoyingjys/YOLOv10-BFDA.