Automated Patient-Specific C1-C2 Posterior Cervical Fusion Screw Trajectory Planning using 3D Deep Learning

Posterior cervical fusion at C1-C2 is technically challenging due to the complex anatomy of this region. Optimal screw trajectory planning is crucial, yet currently relies extensively on imaging and manual adjustment. This work investigates an automated planning approach using 3D deep learning and immersive visualization for intuitive verification and editing. Vertebral CTs were converted to 3D point clouds, and ideal screw entry/exit locations were annotated by experienced technicians and approved by neurosurgeons. These ideal locations were replaced with 60-point clusters, with cluster size optimized to balance segmentation and recovery accuracy. Two deep networks were developed based on the PointNet architecture to efficiently learn both local and global 3D vertebral shape features of C1 and C2, respectively. Augmentation techniques generated 11,000 unique training datasets from 50 qualified patient CTs. Cluster center points identified the closest vertebral surface points as suggested entry/exit locations. Performance was quantified by deviation from ideal paths and proximity to vertebral surfaces. Testing on 10 cases not involved in training showed trajectories within 2.7 mm of surgeon-validated ground truth. A preliminary Microsoft HoloLens 2 interface demonstrates the potential of intuitive 3D validation and modification. Overall, deep learning segmentation of vertebral point clouds successfully generated automated, patient-specific screw plans for challenging C1-C2 fixation, eliminating manual measurement errors. The immersive interface further aids preoperative planning. This work demonstrates the potential of deep-learning technologies to enhance precision and safety in posterior cervical fusion.