Preoperative evaluation of C2 pedicle screw placement using a deep learning model: Development and validation study

Abstract

Background Current preoperative assessment methods for C2 pedicle screw placement face challenges including low consistency, operational complexity, and high skill demands. Objective This study aimed to develop and validate a deep learning model for rapid and accurate assessment of C2 pedicle screw placement feasibility. Materials and methods We developed C2-Net, an automated deep learning pipeline incorporating an image segmentation module for delineating C2 pedicles in CT images and a screw placement probability assessment module. The model's performance was evaluated using 3D-printed manually placed screws as ground truth and compared with surgeons of different experience levels. Results On the test set, C2-Net achieved an accuracy of 89.4%, sensitivity of 90.0%, and specificity of 89.0%. The model demonstrated performance comparable to senior surgeons and numerically superior to junior surgeons, with higher consistency in diagnostic metrics. Attention maps generated by the model provided visual interpretation of the decision-making process. The predicted probabilities demonstrated capability in differentiating structural variations of C2 pedicles. Conclusion C2-Net shows high accuracy and efficiency in assessing C2 pedicle screw placement, outperforming junior surgeons. With its ability to provide rapid, consistent evaluations and visual interpretations, C2-Net demonstrates potential as a valuable assistive tool for clinical decision-making in spinal surgery. Trial Registration: ChiCTR2500101655