Development of an automated CT-Based vertebral segmentation pipeline for digital 3D surgical planning

The main objective of this project was to develop a pipeline for the automatic segmentation of vertebrae in Computed Tomography (CT) scans to accelerate and support 3D spine surgical planning. To achieve this goal, several specific objectives were defined, ranging from adapting an existing generic segmentation model to a task-specific dataset to applying post-processing techniques to improve contour accuracy and validating the results with clinical experts. The implemented method is based on the multi-organ nnU-Net segmentation model, finetuned with a CT dataset focused on vertebrae. A comprehensive quantitative evaluation was conducted at both vertebral region and individual vertebra levels, combined with visual assessment to validate the anatomical consistency of the results. Additionally, postprocessing techniques, including Chan–Vese active contours and Conditional Random Fields (CRFs), were applied to refine boundaries and reduce false positives. Results indicate that, although the original general model achieves superior global performance metrics (Dice), the fine-tuned model demonstrates improved local boundary definition and spatial resolution, despite exhibiting label consistency challenges in the thoracic region. Post-processing further increased contour precision by reducing false positives, while slightly decreasing recall, highlighting a trade-off between boundary accuracy and anatomical integrity. In conclusion, reliable vertebral segmentation for surgical planning cannot rely solely on deep learning models. A carefully designed pipeline that balances model selection, label consistency, spatial resolution, and post-processing strategies, guided by quantitative metrics and clinical expertise, is required. This work provides a structured and clinically informed framework that can serve as a foundation for future improvements toward fully automated and reliable 3D spine segmentation.