Article
Automated planning of lumbar pedicle screws – comparison of a self-derived deep learning-based to a commercial atlas-based approach
Automatische Planung von lumbalen Pedikelschrauben: Vergleich eines Deep Learning-basierten mit einem kommerziellen Atlas-basierten Ansatz
Search Medline for
Authors
-
Moritz Scherer
- Universitätsklinikum Heidelberg, Neurochirurgische Klinik, Heidelberg, Deutschland
-
Lisa Kausch
- DKFZ Heidelberg, Medical Image Computing, Heidelberg, Deutschland
-
Basem Ishak
- Universitätsklinikum Heidelberg, Neurochirurgische Klinik, Heidelberg, Deutschland
-
Karl Kiening
- Universitätsklinikum Heidelberg, Neurochirurgische Klinik, Heidelberg, Deutschland
-
Jan-Oliver Neumann
- Universitätsklinikum Heidelberg, Neurochirurgische Klinik, Heidelberg, Deutschland
-
Klaus Maier-Hein
- DKFZ Heidelberg, Medical Image Computing, Heidelberg, Deutschland
-
Andreas W. Unterberg
- Universitätsklinikum Heidelberg, Neurochirurgische Klinik, Heidelberg, Deutschland
| Published: | May 25, 2022 |
|---|
Outline
Text
Objective: The application of Artificial Intelligence is a hot topic in neurosurgery. We recently described a deep learning-based (DL) approach to automated planning of pedicle screws which facilitates navigated or robotic pedicle screw placement. The aim of this study was to compare accuracy of a self-derived DL approach to a commercially available atlas-based (AB) system for pedicle screw planning.
Methods: From a consecutive registry of CT-navigated instrumentations, we randomly selected 24 cases with 102 screws placed in L2-L5 for analysis. Using the Brainlab® iPlan Spine App, screw planning was manually performed by two independent raters which defined the ground truth (GT) for screw positions and dimensions. Additionally, screws were planned using the DL approach and the Brainlab® iPlan Spine automatic planning tool, which is an atlas-based planning algorithm for pedicle screws. Using a Python script, both automatic planning results were compared to the GT. Screws were compared by mean absolute differences (MAD) of respective head and tip points (in mm) and their angular deviation (in degree) in 3D space. Results were evaluated in comparison to interrater variability of manual screw planning.
Results: Automatic planning was successful in all 102 screws with the DL approach and in 82/102 (80%) with the AB approach. Compared to the GT, MADs in DL planning for angular deviation, head and tip points were 5.5±3.2°, 4.4±2.6mm and 3.8±1.9mm, respectively. For AB planning, corresponding MADs were 8.4±6.8°, 10.9±5.7mm and 11.8±10.7mm, respectively. Interrater variance for manual screw planning was 5.3±3.2°, 4.6±1.8mm and 3.9±1.8mm, respectively. Evaluated MADs for DL planning were statistically comparable to interrater variance of manual screw planning (p=0.62). AB planning exhibited significantly greater differences compared to either DL (p<0.001) and interrater variability (p<0.001), respectively.
Conclusion: DL based screw planning was confirmed with convincing accuracy compared to the GT in all targeted screws. Moreover, DL screws were indifferent to interrater variance of manual screw planning. In contrast, AB screw planning was feasible in a majority of targeted cases but showed inferior results that would need post-processing prior to screw implantation. DL based applications appear as a promising approach to automated screw planning given the frequent anatomic variations of the spine severely limiting the accuracy of AB systems.
