Automated multi-objective pedicle screw planning

Introduction

Spinal fusion is a well-established surgical intervention for the treatment of spinal pathologies, including degenerative disc disease, spinal instability, or deformity. The procedure immobilizes spinal segments using pedicle screws interconnected by metal rods (1-3). The advent of surgical guidance systems for pedicle screw placement has made software-based screw planning a necessary step in spinal fusion procedures (4). To date, pedicle screws are planned manually, which is time-consuming and cognitively demanding (5). Assistive systems could facilitate the planning of pedicle screws, saving valuable time and resources. The planned screws must pass through the pedicle without threatening sensitive neighboring structures and be sized to yield sufficient stability without compromising pedicle integrity (6-8). Additionally, since the metal rod interconnecting the screws must be bent to conform to the screw head positions, the alignment of the screw heads determines the ease of this subsequent work step. Further, the trabecular structure of vertebrae is locally heterogeneous, and different screw trajectories may hence yield significantly different screw stability. While this heterogeneity is difficult to exploit manually, computational methods can identify trajectories with higher stability. In a previous study, we investigated the mechanical benefit of automated bone density-optimized screw planning, which selects trajectories through regions of higher computed tomography (CT)-derived bone mineral density, resulting in 40% higher pull-out forces compared to traditional standard trajectories. Computer-assisted systems, hence, offer the opportunity to optimize screw placement for safety, alignment, and mechanical stability.

Here, we propose and evaluate a fully automated screw planning pipeline. Both manual and automated screw plans were generated from clinical CT scans. The automated plans were evaluated for clinical safety and compared against manual plans for bone density coverage and screw head alignment. We present this article in accordance with the STROBE reporting checklist (available at https://jss.amegroups.com/article/view/10.21037/jss-2025-aw-188/rc).

Methods

Study design

Pedicle screw plans were generated on clinical CT scans once through the automated planning pipeline and once manually by an experienced surgeon, as a reference. Clinical safety of the automated pedicle screw planning was evaluated using the Gertzbein-Robbins classification system, with additional reassessment of perforation occurrence, location, and extent. These ratings were performed by two separate readers, where inter-reader agreement was quantified with Fleiss’ kappa (9) and associated 95% confidence interval (CI). Kappa values were classified as poor (<0), slight (0–0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), and almost perfect (0.81–1.00) (10). A final, unified Gertzbein-Robbins classification was established by consensus in a meeting of the two readers. Total Hounsfield unit (HU) coverage as a proxy for screw stability and screw-head misalignment to gauge subsequent ease of rod contouring of the manual and automated plans were compared for each screw plan. Volume-standardized (average) HU coverage was additionally compared. Appropriate screw types were selected from the M.U.S.T. implant catalogue (M.U.S.T., Medacta International SA, Verona, Switzerland), and the chosen screw dimensions were recorded and tabulated. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The Cantonal Ethics Committee of Zurich confirmed that ethical approval was not required (No. Req-2025-00600).

Dataset

We extracted 20 CT scans from the hospital imaging database of spines without any implants, where at least five intact thoracolumbar vertebrae up to T9 were present, and where no contrast agent was used. Written informed consent was obtained from all patients for the use of their data for research purposes. The scans were anonymized before being transferred to the research environment. Median patient age was 63 [interquartile range (IQR), 54.3–73.5] years, and 72.73% were male. CT acquisition parameters varied across scans, with a median in-plane resolution of 0.4 (IQR, 0.3–0.7) mm, and a median slice thickness of 0.7 (IQR, 0.7–1) mm. Tube potential was 120 kVp for all scans, and tube current was 149 (IQR, 94–204) mA. Spinal fusion pedicle screw planning was conducted for all intact vertebrae visible in the scan.

Automated pedicle screw planning

The proposed pipeline for automated pedicle screw planning provides an end-to-end solution from CT scan to screw plan (Figure 1). The pipeline starts with the segmentation of the CT scan, wherein vertebral structures are identified and labeled image masks generated by a validated pipeline of neural networks (11) implemented in a commercial software (SpineR, v1.1, Bonescreen GmbH, Munich, Germany). Extracted image masks are further processed through level-specific template mesh fitting (12). For each vertebral level, a pre-defined template surface mesh with annotated anatomical regions is registered to the segmentation and mapped using nearest-neighbor correspondence. Pedicle screw placement is typically performed centrally through the pedicle, avoiding cortical bone perforation and damage to surrounding sensitive tissues. To operationalize these clinical constraints, the anatomical landmarks are used to define a pedicle coordinate frame, to localize the pedicle’s minimal cross-section. At this location, the pedicle is intersected by a plane orthogonal to the estimated pedicle axis, and a two-dimensional point grid with 0.5 mm spacing is generated within the pedicle cross-section to represent candidate intrapedicular target points. Screw insertion points are available from prior region mapping on the posterior vertebral surface within a constrained entry region around the standard anatomical entry zone. The effective spatial resolution of this sampling is governed by the underlying segmentation and mesh discretization and was approximately 1 mm in the present dataset. The permissible zone is then constructed as the set of all valid combinations of insertion points and intra-pedicular grid points. Each pair defines a candidate trajectory axis spanning from the insertion point to a target point within the pedicle isthmus, resulting in a discrete, bounded set of feasible trajectories per pedicle. To reduce computational complexity while preserving high-quality candidates, trajectories are evaluated and grouped into bins per insertion point, where the best-scoring trajectory within each bin is retained. A multi-hierarchical grid search is then performed over all pedicle grids to identify the best set of trajectories across the entire construct deterministically. The core objective of the function balances mechanical stability, clinical safety, and rod bending, combining a construct-level rod-bending loss with per-level safety and stability terms aggregated across the instrumented vertebrae. All terms are defined as scalar, unitless quantities after normalization. Individual loss weights were empirically determined, such that each objective’s variance contribution to the outcome yielded equal attribution of importance to the optimization target. Clinical safety quantifies screw trajectory divergence from standard clinical practice, ensuring alignment parallel to the superior endplate and centrally through the pedicle. Mechanical stability computes the integrated HU along a cylinder representing the screw. The rod bending term penalizes relative caudal-cranial screw head misalignments, encouraging screw positions that require less rod manipulation. The automated pedicle screw planning pipeline was executed on a workstation equipped with 128 GB RAM, an Intel Core™ i9-7980XE CPU @ 2.60 GHz, and an NVIDIA GeForce GTX 1080. Including all pipeline steps, median per-case runtime was 5.4 (IQR, 4.6–6.9) min.

Figure 1 Schematic overview illustrating the main components of the automated pedicle screw planning pipeline, comprising segmentation, template mesh fitting and permissible zone delineation, optimization function, and generated surgical plan.

Manual pedicle screw planning

Manual surgical planning was conducted with a custom-developed module in 3D Slicer (13), an interactive platform for visualization and manipulation of medical imaging data. Screw planning was conducted in accordance with the clinical standard of care by a board-certified spine surgeon with 2 years of post-residency experience. The screws were placed centrally through the pedicle, aligned parallel to the superior endplate, and oriented along the longitudinal axis of the pedicle. Screw length was selected to maintain a safety margin of at least 4 mm from the cortical wall. Screw diameter was selected to the most appropriate fit, maintaining a minimum spacing of 1 mm to cortical walls (Figure 2). Manual planning was performed based solely on the pre-operative anatomy and did not explicitly account for anticipated intraoperative deformity correction.

Figure 2 Manually planned screw following standard clinical practice—centrally through the pedicle and parallel to the superior endplate.

Statistical analysis

Sample size estimation indicated 20 cases (i.e., 40 rods) to be sufficient to yield a power of 87% to detect a difference, assuming a paired effect size of d=0.5. The Wilcoxon signed-rank test was applied to assess absolute differences between automated and manual plan outcomes. HU coverage was compared on a per-rod basis through mean aggregation. Similarly, screw head misalignment was compared on a per-rod basis by extracting median and maximum angles between subsequent screws. Summary statistics are provided as median and IQR and absolute and relative counts, as applicable. The analysis was conducted with MATLAB [The MathWorks Inc. (2024b), Natick, MA, USA]. P values below 0.05 were considered statistically significant.

Results

A total of 238 pedicle screws from 20 clinical CTs of 4–8 vertebrae-level fusions were analyzed. The Gertzbein-Robbins classification yielded 87.8% (209 of 238) of pedicle screws from automated planning as grade A (no cortical breach), with the remaining 12.2% of screws classified as grade B (minor breaches <2 mm). Among the grade B screws, medial breaches were most common, accounting for 48.5% of all cortical violations, followed by lateral breaches at 39.4%, and superior breaches at 12.1%. No inferior breaches were observed. Grade B screws exhibited a median breach distance of 0.8 (IQR, 0.65–1.2) mm. Inter-reader reliability of the Gertzbein-Robbins classification yielded a kappa of 0.81 (95% CI: 0.78, 0.84), which is classified as almost perfect. Median HU coverage was improved by 74.9% (IQR, 41.9–136.8%) (P<0.001) and average HU coverage by 17.2% (IQR, 1.7–55%) (P<0.001). Automated planning resulted in improved screw head alignment both in terms of maximum 58.6% (IQR, 5.4–77.5%] (P<0.001) and in median 67.9% (IQR, 33.4–81.8%) (P<0.001) per-rod misalignment, respectively (Figure 3).

Figure 3 Screw-head misalignment and HU coverage resulting from automated and manual screw planning. CT, computed tomography; HU, Hounsfield unit.

Figure 4 provides an exemplary visualization of the screw-head alignment of a manually and an automatically generated screw plan.

Figure 4 Posterior view of a manually (A) and an automatically (B) generated screw plan showing improved screw head alignment through automated planning.

Table 1 reports computed cohort data from automated and manual planning across HU coverage, screw-head misalignment, and screw dimensions.

Discussion

This study evaluated an automated pedicle screw planning framework for multi-level fusion, optimizing for clinical safety, biomechanical stability, and screw head alignment to overcome manual planning limitations. The automated pipeline demonstrated significant improvements over conventional manual planning.

The automated system demonstrated relatively high placement accuracy regarding anatomical boundaries, with 87.8% of screws classified as Gertzbein-Robbins grade A and the remaining 12.2% as grade B, with all cortical breaches being minor, measuring less than 2 mm. These results are comparable to the reported accuracy of other automated planning systems, which report grade A rates ranging from 80% to 95% (14,15). In contrast to registration- or atlas-based approaches currently explored in clinical research and early-stage systems (16-18), our method integrates safety, biomechanical stability, and rod conformity within a single optimization framework. Further improvements in the Gertzbein-Robbins classification could likely be achieved by adjusting the weights to place a higher emphasis on the clinical safety loss term. Qualitative review of the segmentations revealed that a significant portion of the perforations likely occurred due to minor segmentation discrepancies. The fact that most breaches were either in the medial or the lateral direction is most likely a result of the cranio-caudally oriented oval shape of the pedicle cross-section.

The automated approach yielded superior HU coverage, implying higher bone purchase. While the automated plans utilized larger screw dimensions on average, the observed improvements in bone purchase were not merely a product of increased volume. After normalizing for screw size, automated planned trajectories achieved a 17.2% higher average HU density compared to manual plans. In a recent study on bone density-optimized pedicle screw planning, our algorithm demonstrated its effectiveness in enhancing mechanical stability, showing a significant improvement in screw pull-out resistance (19).

Further, a substantial reduction in the angular divergence between consecutive screw heads was observed, which would likely reduce the amount of rod manipulation required during the surgery. In the long term, potential risk, such as rod fracture might be lowered, since the reduced extent of deformation inherently preserves the rod’s structural integrity and mitigates fatigue (20,21).

This study is subject to limitations that warrant consideration. Manual planning was performed by a single surgeon following standard clinical practice, which may introduce rater-specific bias. In particular, manual screw head alignment may reflect individual planning preferences or implicit anticipation of intraoperative correction strategies. While this reflects real-world planning behavior, it limits the ability to attribute differences in screw head alignment exclusively to planning quality. The current automated planning system focuses on optimizing individual screw placement and construct metrics based on a pre-operative anatomical configuration and does not explicitly incorporate deformity correction objectives or global spinal alignment targets. In cases involving mobile segments or planned intraoperative realignment, however, estimated vertebral pose changes could be applied to the anatomical surface models and underlying CT data, enabling re-evaluation or re-optimization of screw trajectories under corrected alignment conditions. This preserves the applicability of the planning framework even when spinal geometry is altered during surgery.

While automated planning promises a reduction in planning time, the study does not provide data on the potential time gains of the proposed system. Previous studies report a 26% time saving (18) of similar systems compared to traditional manual planning.

This study evaluates the planning stage of pedicle screw instrumentation in isolation by comparing virtual screw plans, thereby quantifying the effect of automated optimization on screw trajectory selection and construct configuration independent of execution-related confounders. A formal sensitivity analysis to segmentation uncertainty and execution perturbations was not performed, as this would require explicit modeling of system- and workflow-specific error characteristics beyond the scope of the present planning-stage evaluation. The true clinical impact of automated planning ultimately depends on how accurately virtual plans can be translated into screw placement. While the present work focuses on virtual planning to assess achievable improvements in safety, stability, and alignment, end-to-end validation comparing planning-based and non-planning-based placement under surgical or simulated conditions remains an important direction for future research. Accordingly, HU coverage was used as a surrogate planning metric for mechanical stability rather than a more direct outcome measure.

Translating virtual plans to real-world screw placement inherently involves some degree of inaccuracy. The inherent divergence between pre-operative plans and actual screw placement, typically exhibiting systematic errors in the order of 1–3 mm (22,23), means that the extent to which these optimized plans can be reliably executed in surgery remains to be fully established. In a worst-case scenario, a 3 mm deviation could potentially shift screws planned with the automated pipeline from a Gertzbein-Robbins grade A to a grade C classification, or a grade B.

Additionally, the current anatomical scope of the algorithm is presently validated for planning in the thoracolumbar spine from L5 up to T9. Extension towards higher thoracic levels is technically feasible and mainly a matter of further development and validation.

The dataset used for this validation study represents a convenience sample selected by imaging-based criteria rather than surgical indication, which may introduce selection bias and limit generalizability across diagnoses and clinical populations.

The Gertzbein-Robbins classification assesses pedicle wall breach but does not capture other clinically relevant malpositions, such as superior facet joint violation or proximity to the neural foramen. In the present study, facet joint violation was not explicitly assessed and is not directly constrained by the optimization. Accordingly, safety-related conclusions in this study primarily relate to pedicle containment. Notably, the planning pipeline already identifies vertebral anatomical regions through level-specific template mesh fitting and region labeling, offering an existing framework for integrating facet joint protection into the optimization logic. This existing anatomical annotation could be leveraged to add an explicit facet-related constraint with minimal additional implementation effort.

Ultimately, comparative clinical safety relevant for clinical translation should be evaluated in an end-to-end setting, where both automated and manual plans are executed, and placement accuracy is assessed under operative conditions.

Conclusions

The proposed automated pedicle screw planning approach demonstrated clear improvements over manual planning in terms of safety, stability, and screw head alignment. This data-driven optimization framework represents a step towards intelligent planning systems capable of identifying construct configurations that are both surgically feasible and biomechanically advantageous. Further clinical validation is needed to confirm its applicability in practice.

Acknowledgments

We greatly appreciate Osamu Hirose’s support in algorithm development. We thank Bonescreen GmbH for kindly providing a SpineR license for the study.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://jss.amegroups.com/article/view/10.21037/jss-2025-aw-188/rc

Data Sharing Statement: Available at https://jss.amegroups.com/article/view/10.21037/jss-2025-aw-188/dss

Peer Review File: Available at https://jss.amegroups.com/article/view/10.21037/jss-2025-aw-188/prf

Funding: This work was supported by the University of Zurich (No. MEDEF24-040).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jss.amegroups.com/article/view/10.21037/jss-2025-aw-188/coif). T.G. is the recipient of University of Zurich Fellowship (No. MEDEF24-040). M.F. has received consulting fees from Arthrex and Medacta; holds an issued patent (WIPO Reference No. WO2022128956A1); serves as a Board Member for Swiss Orthopaedics and the Stiftung zur Förderungder Ausbildung in medizinischer Radiologie (IDKD), the President of the Board for Moving Spine AG, and a Board Member for Artonis; holds stock or stock options in Zurimed and Artonis; and has received equipment/material support from Moving Spine AG. J.W. is a co-inventor of an issued patent (No. WO2022128956A1). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The Cantonal Ethics Committee of Zurich confirmed that ethical approval was not required (No. Req-2025-00600). Written informed consent was obtained from all patients for the use of their data for research purposes.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Loenen ACY, Noriega DC, Ruiz Wills C, et al. Misaligned spinal rods can induce high internal forces consistent with those observed to cause screw pullout and disc degeneration. Spine J 2021;21:528-37. [Crossref] [PubMed]
2. Sumiya S, Fukushima K, Kurosa Y, et al. Comparative analysis of clinical factors associated with pedicle screw pull-out during or immediately after surgery between intraoperative cone-beam computed tomography and postoperative computed tomography. BMC Musculoskelet Disord 2021;22:55. [Crossref] [PubMed]
3. Ohba T, Ebata S, Oda K, et al. Utility of a Computer-assisted Rod Bending System to Avoid Pull-out and Loosening of Percutaneous Pedicle Screws. Clin Spine Surg 2021;34:E166-71. [Crossref] [PubMed]
4. Benito R, Bertelsen Á, de Ramos V, et al. Fast and versatile platform for pedicle screw insertion planning. Int J Comput Assist Radiol Surg 2023;18:1151-7. [Crossref] [PubMed]
5. Caprara S, Fasser MR, Spirig JM, et al. Bone density optimized pedicle screw instrumentation improves screw pull-out force in lumbar vertebrae. Comput Methods Biomech Biomed Engin 2022;25:464-74. [Crossref] [PubMed]
6. Shin BJ, James AR, Njoku IU, et al. Pedicle screw navigation: a systematic review and meta-analysis of perforation risk for computer-navigated versus freehand insertion. J Neurosurg Spine 2012;17:113-22. [Crossref] [PubMed]
7. Gelalis ID, Paschos NK, Pakos EE, et al. Accuracy of pedicle screw placement: a systematic review of prospective in vivo studies comparing free hand, fluoroscopy guidance and navigation techniques. Eur Spine J 2012;21:247-55. [Crossref] [PubMed]
8. Tian NF, Huang QS, Zhou P, et al. Pedicle screw insertion accuracy with different assisted methods: a systematic review and meta-analysis of comparative studies. Eur Spine J 2011;20:846-59. [Crossref] [PubMed]
9. Fleiss JL. Measuring nominal scale agreement among many raters. Psychological Bulletin 1971;76:378-82.
10. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics 1977;33:159-74.
11. Sekuboyina A, Husseini ME, Bayat A, et al. VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images. Med Image Anal 2021;73:102166. [Crossref] [PubMed]
12. Hirose O. Geodesic-Based Bayesian Coherent Point Drift. IEEE Trans Pattern Anal Mach Intell 2023;45:5816-32. [Crossref] [PubMed]
13. Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 2012;30:1323-41. [Crossref] [PubMed]
14. Zhao J, Zhang Y, Zhan S, et al. Pedicle screw path planning for multi-level vertebral fixation. Med Phys 2024;51:1547-60. [Crossref] [PubMed]
15. Kausch L, Scherer M, Thomas S, et al. Automatic image-based pedicle screw planning. in Medical Imaging 2021: Image-Guided Procedures, Robotic Interventions, and Modeling 2021;11598:406-14. Available online: https://doi.org/10.1117/12.2582571
16. Maik V, Purayath A, Lakshmanan M, et al. GUI-based Pedicle Screw Planning on Fluoroscopic Images Utilizing Vertebral Segmentation. In: 2024 IEEE International Symposium on Medical Measurements and Applications (MeMeA). IEEE; 2024:1-6.
17. Goerres J, Uneri A, De Silva T, et al. Spinal pedicle screw planning using deformable atlas registration. Phys Med Biol 2017;62:2871-91. [Crossref] [PubMed]
18. Bertram U, Köveshazi I, Michaelis M, et al. Man versus machine: Automatic pedicle screw planning using registration-based techniques compared with manual screw planning for thoracolumbar fusion surgeries. Int J Med Robot 2023; Epub ahead of print. [Crossref]
19. Götschi T, Maranta G, Bernet M, et al. Bone density-optimized pedicle screw planning enhances mechanical stability. Eur Spine J 2025; Epub ahead of print. [Crossref]
20. Lindsey C, Deviren V, Xu Z, et al. The effects of rod contouring on spinal construct fatigue strength. Spine (Phila Pa 1976) 2006;31:1680-7. [Crossref] [PubMed]
21. Demura S, Murakami H, Hayashi H, et al. Influence of Rod Contouring on Rod Strength and Stiffness in Spine Surgery. Orthopedics 2015;38:e520-3. [Crossref] [PubMed]
22. Choi JY, Park SM, Kim HJ, et al. Recent Updates on Minimally Invasive Spine Surgery: Techniques, Technologies, and Indications. Asian Spine J 2022;16:1013-21. [Crossref] [PubMed]
23. Russo F, Nardi N, Papalia GF, et al. Deviation from preoperative planning and pedicle screw accuracy in navigated and robotic spinal fusion: a systematic review. Eur Spine J 2025;34:1890-9. [Crossref] [PubMed]