Article
Automated screw planning of navigated lumbosacral pedicle screws using artificial intelligence
Vollautomatische Planung navigierter lumbosakraler Pedikelschrauben mittels künstlicher Intelligenz
Search Medline for
Authors
Published: | June 4, 2021 |
---|
Outline
Text
Objective: Use of navigation for spinal instrumentation has gained traction in recent years but preoperative trajectory-planning is a time-consuming task. We sought to develop and validate an automated planning tool for lumbosacral pedicle screws using a convolutional neural network (CNN).
Methods: We used planning data from random 155 CT-navigated instrumentations and extracted screw parameters from 1052 pre-planned pedicle screws covering L2-S1 levels, which served as training data for a CNN. A vertebra instance-based approach employing a state-of-the-art U-Net framework was developed and trained followed by internal 5-fold cross-validation. The retrieved net was evaluated on an external test-set of 30 cases not involved in training. Automatic screw parameters were compared to corresponding pre-planned screws in the test-set by mean absolute difference (MAD) of screw head and tip points, length and diameter, respectively. Clinical acceptability of algorithm-generated screws was evaluated by experts using the Gertzbein-Robbins (GR) classification.
Results: Automated planning was feasible for all targeted 198 screws. Compared to pre-planned screws, MAD was 4.3±2.1mm for screw head, 4.2±2.4mm for tip points, 4.6±3.1mm for length and 0.4±0.3mm for diameter. In ANOVA followed by Dunn’s multiple comparison, MAD for head and tip points was significantly greater at L5 and S1 compared to other segments (p<0.001), reflecting increasing degrees of freedom in caudal screw placement. No difference between segments was found for screw length and diameter. Upon expert rating, screws were predominantly classified grade A (189, 95%) with only 9 grade B screws (5%) according to GR indicating that screws showed either no, or only minor (<2mm) cortical breach. All planned screws were classified clinically acceptable. Algorithm generated screws were not inferior to their manually planned companions.
Conclusion: We derived a fully automated planning tool for lumbosacral pedicle screws using CNN. Validation showed sufficient accuracy to facilitate screw planning with high potential to increase time-efficiency in navigated spinal instrumentation when integrated into commercial navigation systems.