Navigated spinal pedicle screw placement has significantly improved surgical outcomes and perioperative complications. Robotic spinal surgery promises to further improve those results. Manual screw planning, while crucial for robotic surgery, remains time-consuming. Deep-learning, and atlas-based methods for automated planning are commercially available for the lumbosacral segments, but not thoracic levels. This study aimed at developing a planning algorithm for the thoracic spine based on a convolutional neural network (CNN).
A CNN was trained on 1143 screws and the corresponding CT scans from patients operated in our center. The test set consisted of 114 thoracic screws. The automatically planned screws were compared to screws independently planned by two surgeons, with regards to intrinsic and geometric parameters. Additionally, the interrater and the intrarater (the same surgeon planning the screws with several weeks interval) variance were evaluated. ANOVA and t-test were used to compare groups, p<0.05 was considered significant.
Automatic planning yielded clinically acceptable screws in all cases. A lateral pedicle breach was observed in 62 screws (mean 1.365±0.67mm), and a medial breach in four (mean 0.54 ±0.08). Planning time per screw was significantly reduced using automated planning (5.84 ± 2.3s vs 123±82.9s; p<0.0001). The distances between automatically and manually planned head and tip points were comparable to the intrarater comparison (p=0.61; p=0.1 respectively) and significantly lower than between two raters (p<0.0001). The deviation in direction was likewise significantly higher in the interrater comparison (p<0.0001).
While the automatically planned screws did not differ in length (p=0.426), the CNN planned markedly thicker screws than the surgeons (p<0.0001).
Table 1: Variations in screw parameters between automatic and manual planning, as well as between different surgeons (interrater) and the same surgeon at different time points (intrarater)
We developed and validated a robust CNN-based algorithm for automated thoracic pedicle screw planning. The screw plans were non-inferior to manually planned screws while markedly reducing the time expenditure. Four planned screws deviated medially, however, <2mm, and can therefore still be considered clinically acceptable. While clinical testing and validation are necessary, our approach promises reliable automated pedicle screw planning.