The recognition of surgical tools based on attention mechanism

The application of deep learning technology in healthcare has been widely studied, but there has been limited research on the management of surgical tools. Therefore, we propose a study focusing on the recognition and classification of surgical tools to reduce the risk caused by their loss. Firstly, we design a surgical tool collection system and construct a commonly used surgical tool dataset (STD). Secondly, we investigate two embedding strategies using an attention mechanism in the benchmark network to select suitable tool recognition methods. The first strategy is to embed attention in the base module of the extraction network, known as the embedding strategy. The second strategy is to embed attention in the last step before prediction, called the additional strategy. Our experimental results show that the performance of the additional strategy is superior to that of the embedded strategy in the surgical tool detection task. Using the additional strategy with the improved Yolov5s network based on SimAM attention resulted in an average accuracy of 0.972 while keeping GFLOPs unchanged. The model proposed in this paper outperformed advanced models and showed performance improvement on STD.