3D Path Planning for Anterior Spinal Surgery Based on CT images and Reinforcement Learning*

Spinal internal fixation is one of the most complex operations, and the planning is the core process of preoperative preparation in robot-assisted surgery. Compared with posterior surgery, the anterior one is gradually accepted by the general public for more accuracy and less aesthetic impact. This paper proposes an anterior surgical path planning method based on the reinforcement learning. Firstly, the multi-task segmentation used for building the planning environment is performed on the chest-enhanced medical images, after which the 3D model of specified organs such as spine and vessels are reconstructed. Then, the surgical path is obtained by the Q-learning algorithm based on the model, while the $\varepsilon$-greedy policy is applied to guarantee the rapid convergence to global optimum. A curved limit scheme is proposed to solve the curse of dimensionality and improve the pathway searching efficiency. The experimental results indicate that the method performs well in the surgical 3D path planning. The path for the end of the robot from the entrance to the lesion site can be planned automatically to avoid the vital organs, even though without the prior knowledge of the environment.