Computers in Biology and Medicine ( IF 7.7 ) Pub Date : 2020-07-05 , DOI: 10.1016/j.compbiomed.2020.103867 Lingtao Yu 1 , Pengcheng Wang 1 , Yusheng Yan 1 , Yongqiang Xia 1 , Wei Cao 1
Surgical instrument detection is a significant task in computer-aided minimal invasive surgery for providing real-time feedback to physicians, evaluating surgical skills, and developing a training plan for surgeons. In this study, a multi-scale attention single detector is designed for surgical instruments. In the field of object detection, accurate detection of small objects is always a challenging task. We propose an innovative feature fusion technique aimed at small surgical instrument detection. First, the attention map is created from high-level features to act on the low-level features and enrich the semantic information of the low-level features. The original and processed features are then fused by skip connection. Finally, multi-scale feature maps are created to predict fusion features. The experiments on the ATLAS Dione dataset yielded results with a detection time of 0.066 s per frame and a mean average precision of 90.08%. Our proposed feature fusion module can obtain more semantic information for low-level features and significantly enhance the performance of small surgical instrument detection.
中文翻译:
MASSD:用于手术器械的多尺度注意力单发检测器。
手术器械检测是计算机辅助微创手术中的一项重要任务,它可以向医生提供实时反馈,评估手术技能并制定针对外科医生的培训计划。在这项研究中,为手术器械设计了一种多尺度注意力的单一检测器。在物体检测领域,精确检测小物体始终是一项艰巨的任务。我们提出了一种针对小型手术器械检测的创新性特征融合技术。首先,从高级特征创建注意力图,以作用于低级特征并丰富低级特征的语义信息。然后通过跳过连接将原始特征和已处理特征融合在一起。最后,创建多尺度特征图以预测融合特征。在ATLAS Dione数据集上进行的实验得出的结果是,每帧检测时间为0.066 s,平均平均精度为90.08%。我们提出的特征融合模块可以为低级特征获取更多的语义信息,并显着增强小型手术器械检测的性能。