Computer Vision Applications in Spinal Orthopaedics: A Scoping Review of Imaging-Based Algorithms for Diagnosis, Measurement, and Surgical Planning

TL;DR

This paper reviews how computer vision is used in spinal imaging for diagnosis and surgery, highlighting strong technical results but limited real-world application.

Contribution

The paper provides a comprehensive scoping review of computer vision applications in spinal orthopaedics, identifying trends and gaps in clinical translation.

Findings

Deep learning methods like U-Net and ResNet are widely used for spinal image segmentation and labeling.

Fracture detection algorithms on CT and MRI achieve high AUCs (0.91-0.95), but external validation is rare.

Morphometric measurement algorithms show strong agreement with human analysis (ICC 0.93-0.98).

Abstract

Computer vision has advanced in spinal imaging, enabling automated interpretation of radiographs, CT, and MRI for diagnosis, surgical planning, and postoperative assessment. The spine’s complex anatomy and high imaging volume make it a key area for algorithmic assistance. This scoping review maps current applications of computer vision in spinal orthopaedics and describes the clinical tasks, imaging modalities, and computational methods used in published studies.

A systematic search of Ovid MEDLINE and Embase was performed from January 1995 to October 2025. Studies were included if they applied automated or semi-automated computer vision techniques to spinal imaging for diagnostic, morphometric, or surgical planning. Two reviewers screened and recorded data in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) guidelines.

Twenty studies met the inclusion criteria. CT (45%) and MRI (35%) were the dominant imaging modalities, followed by radiographs (15%) and ultrasound (5%). Deep learning methods were employed in 90% of studies, mainly convolutional architectures such as U-Net, ResNet, and YOLOv5. Segmentation and vertebral labeling were the most common tasks (60%), achieving Dice coefficients of 0.86-0.97 and accuracies of 90-98%. Fracture detection networks (25%) reached AUCs of 0.91-0.95, while morphometric measurement algorithms (15%) produced intraclass correlations of 0.93-0.98 compared with human analysis. Despite strong technical performance, only 20% of studies included external validation, and none conducted prospective testing.

Computer vision has demonstrated strong performance in spinal image segmentation and fracture detection, particularly on CT and MRI. Nevertheless, clinical translation remains limited. Future research should prioritize multi-center datasets, real-world validation, and integration into surgical clinical practice to support preoperative and intraoperative care.

Figures