Anne van Dalen

226 I Chapter 8 Knot-Tying Although in open surgery, knot-tying is part of basic skills and a relative fast procedure, in minimally-invasive surgery, laparoscopic knot-tying can take up to three minutes for a single knot to complete. Mayer et al. 32 described a system to speed up the knot-tying based on RNNs in robotic heart surgery. The surgeon presents a sequence (eg, examples of human-performed knot-tying) to the network and, an RNN with long-term storage learns the task. The preprogrammed controller was able to construct a knot in 33.7 seconds, whereas the use of an RNN provided—after learning from 50 previous runs—a speed improvement of almost 25%, producing a knot in 25.8 sec. 27,33 Automatic Registration and Tracking of the Bone in Orthopedic Surgery In computer-assisted orthopedic surgery, registration of the bone plays a vital role as it describes the position of the patient in regard to the surgical system. This way, the surgical site can be correctly aligned according to the preoperative plan. Therefore, the precision of the registration has influence on all the following steps in the procedure. Liu et al. 29 describe a new way of automatic registration and tracking of the bone, based on depth imaging and deep learning. During surgery, a depth camera repeatedly captures depth images of exposed bone. Using these images, deep neural networks learn to localize, segment, and extract the surface geometry of the target bone. The extracted surface geometry is then compared to a preoperative model of the same bone for registration, making surgical intervention or invasive optical markers superfluous. Ex vivo experiments show a mean translational and rotational error of 2.74 mm and 6.66°, respectively. However, these accuracies are currently lower than conventional intraoperative registration methods based on optical markers. 34,29