ToolNet-X: Surgical Instrument Detection Combined With High-Order Spatial Interaction

Accurately detecting laparoscopic surgical instruments is crucial to ensuring the success of robot-assisted laparoscopic surgery. In this paper, we propose a laparoscopic surgical instrument detection framework called ToolNet-X to address the issues of low accuracy and inadequate real-time performance in current detection methods. The algorithm effectively fuses feature maps of different scales using weighted feature fusion and increasing fusion links. In addition, we introduce gated convolution and recursive operations to enhance the network’s global modeling capabilities. The experimental results show that ToolNet-X achieved a precision of 97.8%, a mAP@0.5 (mean Average Precision) of 96.4%, and a mAP@0.5:0.95 of 59.398% on the M2CAI16-Tool-Locations dataset. Furthermore, we evaluate the performance of the YoloV5 and YoloV7 algorithms in surgical instrument detection and improve the latter using the same approach proposed in this paper, demonstrating significant improvements for both algorithms.