Precision in Cervical Pedicle Screw Placement

: Predictive Insights from Machine Learning Models

Abstract

Accurate pedicle screw placement is essential for stable fixation in spinal surgery, yet remains one of the most technically challenging tasks. The cervical and thoracic spine are particularly di!cult due to small pedicle dimensions, anatomical variability, and proximity to critical neurovascular structures. Even with image-guided navigation and robotic assistance, malposition rates remain significant. This thesis explores whether machine learning methods can support surgical decision-making by predicting pedicle screw placement quality from multimodal data, presenting a novel approach that integrates clinical parameters with features extracted from intraoperative CT segmentations. A dataset was compiled combining demographic and clinical information with automatically segmented vertebral structures. Segmentation was performed using the TotalSegmentator tool to obtain reproducible anatomical features, though the quality and granularity of intraoperative CT segmentations posed limitations. Eight classifiers were trained and evaluated, ranging from simple baselines to advanced ensemble and deep learning–inspired methods. The primary predictive task was binary classification between good (Gertzbein grades 0–1) and breach (grades 2–3), with multiclass classification across all grades included for completeness. Overall predictive performance was modest, reflecting both task complexity and the severe class imbalance. Reported metrics, such as a maximum ROC AUC of 0.61, F1-score of 0.21, and accuracy of 0.90, each from di"erent models and configurations, highlight the current challenges rather than clinical readiness. The main outcome of this work lies instead in the interpretability analyses. Feature importance and SHAP evaluations consistently surfaced variables that domain experts considered clinically plausible and relevant, suggesting that the models capture meaningful anatomical and procedural patterns even where predictive accuracy is limited. This thesis contributes to the emerging field of machine learning in spine surgery by demonstrating the feasibility and constraints of combining intraoperative imaging with patient-specific data for screw placement prediction. It underscores the potential of interpretability analyses to generate new insights, while also pointing to the need for improved segmentation methods, larger and more balanced datasets, and refined feature engineering before such models can be considered for clinical translation.

Date of Award	2025
Original language	English
Supervisor	Alessio Montuoro (Supervisor)