A Preliminary Investigation Into the Feasibility of Semi-Autonomous Surgical Path Planning for a Mastoidectomy Using LSTM-Recurrent Neural Networks

Path planning algorithms for robotics can be as simple as having an operator program waypoints into a robot's controller and having the robot perform a simple task such as welding. This works well in an industrial setting but will not work for complicated tasks such as performing surgery. Another approach would be to use a constraint function in programing a robot to perform surgery, but it would be difficult to capture and represent all of the surgeon's information in the mathematical terms required for a cost function. A third approach, and the one utilized in this study, is to train a set of artificial neural networks (ANNs) using recorded surgeons' motions when manipulating a surgical instrument during procedure training using a surgery simulator. This has the advantage of indirectly capturing the surgeon's abilities and intentions without needing to explicitly capture all of the motion information that must be encoded from their trajectory planning and decision-making, and then, say, creating a complex constraint function using that information. In this research effort, virtually captured surgical trajectories from trained surgeons were used to train ANNs, after being preprocessed into three subtasks. Each set of subtask data was used to train a separate ANN. Each of the ANNs was trained using a custom cost function and evaluated using custom metrics. During the training, the positions of fiducial markers, recorded during procedure attempts, were used to orient the recorded path relative to the patient's anatomy. Although the ANN-generated trajectories were not used to perform surgery on a live patient in this study, the fiducial marker position information is intended to be exploited during a real procedure to position, orient, and scale a tool trajectory to suit a patient's specific anatomy. The trained ANNs were subjected to several tests to assess their safety and robustness. We found that even when trained on a small number of datasets, the ANNs converged and could generate output trajectories that were still assessed to be safe even when slight changes in the fiducial marker placement locations were given.