Deep learning-based automatic preoperative planning for deep brain stimulation

WU Weidong; GONG Shun; LI Xinyang; TAO Yingqun

Abstract：Objective To make automated brain nucleus labeling,surgical target localization,and electrode implantation trajectory planning for subthalamic nucleus deep brain stimulation(STN-DBS)based on deep learning.Methods The imaging data of 155 patients receiving bilateral STN-DBS surgery at the Department of Neurosurgery,General Hospital of the Northern Theater Command were analyzed in the following three steps.First,a 3D UX Net convolutional neural network was used to extract deep learning features from MRI images and complete the segmentation of subthalamic nucleus(STN)and red nuclei(RN).Secondly,STN target localization was achieved through algorithms.Finally,no more than 4 electrode implantation trajectories were generated within the specified range area.Clinical feasibility of the target and electrode implantation trajectories were verified through manual review.Results The mean Dice coefficient of RN and STN was 0.90 and 0.84 respectively.The Euclidean distance between the coordinates of the automated STN target,which have been verified to be feasible through manual review,and the coordinates of the STN target in manual surgical planning was 1.2±0.4 mm.25 cases of the automated STN target localization and electrode implantation trajectory in the test set were reviewed manually,and 20 of them(20/25)were feasible.There was no statistically significant difference in automated target localization and electrode implantation trajectory compared to manually planned target and trajectory(P=0.059).Conclusion The 3D UX Net convolutional neural network can be used to segment STN and RN accurately,as the results of the automated target localization and trajectory planning are clinically feasible and can save time,so it can provide reference for clinical doctors in preoperative planning.