Deep Learning-Based Spine Vertebrae and Pedicle Screw Segmentation from Postoperative CT Images: a Feasibility Study Using Limited Data

Objective verification of pedicle screw placement on postoperative imaging is essential for patient safety and for evaluating the biomechanical integrity of spinal constructs. Inaccurate screw positioning may compromise stability and increase the risk of neurological injury. Deep learning provides opportunities to automate segmentation of implants and surrounding anatomy, enabling objective and reproducible assessment of implant-anatomy relationships. This study evaluated the feasibility of training readily available convolutional neural networks (CNNs) to segment vertebrae, the spinal canal, and pedicle screws in postoperative thoracolumbar CT images using a limited dataset and 2D slice-based segmentation. One hundred manually annotated axial postoperative CT slices from 20 patients who had undergone posterior fixation were included and divided into training, validation, and test sets across two data splits. Four CNN architectures: YOLOv8, DeepLabv3+, U-net, and Attention U-net, were trained for multi-class segmentation. Model performance was assessed using class-wise Dice Similarity Coefficients (DSC). Across both splits, segmentation accuracy varied by model and anatomical class, mean DSC 0.62-0.90, reflecting variations across classes and models. DeepLabv3+achieved the highest overall performance, mean DSC 0.83 and 0.78. YOLOv8 demonstrated stable results across splits, particularly for screw segmentation. U-net and Attention U-net performed comparably for spinal canal and screw-head segmentation in the first split but showed reduced performance in the second split. Qualitative assessment confirmed that all models were able to identify key anatomical structures and implants, although metal artifacts affected performance. As a feasibility demonstration, the study showed that DeepLabv3+and YOLOv8 provided the most consistent results, indicating that robust automated screw-canal assessment is achievable even with limited clinical data.