Robust and Expert-Level Automated Screw Planning in the Entire Spine Based on Deep-Learning [data]

Manual screw planning for spinal stabilization is a time-consuming process,
interrupting the surgical workflow and prolonging operating time. Existing automated
solutions focus on the lumbosacral region due to reliance on vertebral homogeneity. We introduce a deep-learning-based method reformulating screw planning as an image
segmentation task, enabling automated generation of trajectories for all spinal levels
and different screw types. Trained on over 450 real surgical cases, our approach
achieved a 97% clinical acceptability rate on a representative testset, with a mean
absolute deviation from manual planning of 2.6mm, which lies within inter- and
intra-rater variability, demonstrating expert-level performance. Extensive evaluation
on deformity cases and external data confirmed robust generalization across
institutions, scanners, and pathologies. Importantly, this is the first automated
solution for cervical lateral mass screw planning. Our method promises to streamline
surgical workflows, enhance integration with 3D image navigation systems, and improve the safety and efficiency of spine surgery.