SimPS-Net: Simultaneous Pose and Segmentation Network of Surgical Tools

Localisation of surgical tools during operation is of paramount importance in the context of robotic assisted surgery. 3D pose estimation can be utilised to explore the interaction of tools with registered tissue and improve the motion planning of robotic platforms, thus avoiding potential collisions with external agents. With the problems of traditional tracking systems being cost and the need to redesign surgical tools to accommodate markers, there has been a shift towards image-based, markerless tracking techniques. This study introduces a network capable of detecting and localising tools in 3D using a monocular setup. For training and validation, a novel dataset, 3dStool, was produced, and the network was trained to obtain a mean Dice coefficient of 85.0% for detection, along with a mean position and orientation error of 5.5mm and 3.3° respectively. The presented method is significantly more versatile than various state of the art solutions, as it requires no prior knowledge regarding the 3D structure of the tracked tools. The results were compared to standard pose estimation networks using the same dataset and demonstrated lower errors along most metrics.In addition, the generalisation capabilities of the proposed network were explored by performing inference on a previously unseen pair of scissors.