Development and validation of an automated planning tool for navigated lumbosacral pedicle screws using a convolutional neural network

Abstract

Background contextNavigation and robotic systems have been increasingly applied to spinal instrumentation but dedicated screw planning is a time-consuming prerequisite to tap the full potential of these techniques. PurposeTo develop and validate an automated planning tool for lumbosacral pedicle screw placement using a convolutional neural network (CNN) to facilitate the planning process. Study Design/SettingRetrospective analysis and processing of CT and screw planning data randomly selected from a consecutive registry of CT-navigated instrumentations from a single academic institution. Patient SampleData from 179 cases was processed for CNN training and validation (155 for training, 24 for validation) leveraging a total of 1182 screws (1052 for training, 130 for validation). Outcome MeasuresQuantitative and qualitative (Gertzbein-Robbins classification [GR]) validation via comparison of automatically and manually planned reference screws, inter-rater and intra-rater variability. MethodsAnnotated data from CT-navigated instrumentation was used to train a CNN operating in a vertebra instance-based approach employing a state-of-the-art U-Net framework. Internal five-fold cross-validation and external validation on an independent cohort not previously involved in training was performed. Quantitative validation of automatically planned screws was performed in comparison to corresponding manually planned screws by calculating the minimal absolute difference (MAD) of screw head and tip points, length and diameter, screw direction and Dice coefficient. Results were evaluated in relation to inter-rater and intra-rater variability of manual screw planning. ResultsAutomated screw planning was successful in all targeted 130 screws. Compared with manually planned screws as a reference, mean MAD of automatically planned screws was 4.61±2.27 mm for screw head, 3.96±2.19 mm for tip points and 5.51±3.64° for screw direction. These differences were either statistically comparable or significantly smaller when compared with interrater variability of manual screw planning (p>.99 for head point and direction, p=.004 for tip point, respectively). Mean Dice coefficient of 0.61±0.16 indicated significantly greater agreement of automatic screws with the manual reference compared with interrater agreement (Dice 0.56±0.18, p<.001). Automatically planned screws were marginally shorter (MAD 3.4±3.2 mm) and thinner (MAD mean 0.3±0.6 mm) compared with the manual reference, but with statistical significance (p<.0001, respectively). Automatically planned screws were GR grade A in 96.2% in qualitative validation. Planning time was significantly shorter with the automatic approach (0:41 min vs. 6:41 min, p<.0001). ConclusionsWe derived and validated a fully automated planning tool for lumbosacral pedicle screws using a CNN. Our validation showed noninferiority to manual screw planning and provided sufficient accuracy to facilitate and expedite the screw planning process. These results offer a high potential to improve workflows in spine surgery when integrated into navigation or robotic assistance systems.