Anchor-Free Convolutional Neural Network Application to Enhance Real-Time Surgical Tool Detection in Computer-Aided Surgery

Computer-aided surgery requires efficient real-time detection of surgical tools. The respective technology should furnish real-time positions of different surgical tools for surgeons or auxiliary robots, improving surgical efficiency and reducing complications. At present, most detection methods of surgical instruments rely on predefined anchor boxes to obtain good detection accuracy. However, they require many complex calculations related to anchor boxes, which may not strike a good balance between detection accuracy and speed. Given the above problem, this study proposes an anchor-free convolutional neural network architecture that avoids complex calculations related to anchor frames and reduces network parameters. It uses the Bridge Network (BriNet18) and the Multiple Cross Stage Path Aggregation Network (MCSPAN) for feature extraction and feature fusion, respectively. In addition, the attention mechanism is incorporated into the output header to enhance the expression ability of the network and improve the network detection performance. This method shows excellent performance on the ATLAS Dione and Cholec80-six datasets, achieving 98.51% mAP@0.5 and 97.60% mAP@0.5 at 56 frames per second, respectively. The experimental results prove that the proposed method is superior to the existing detection methods and truly realizes fast, accurate, and end-to-end detection of surgical tools.