Artificial Intelligence in CT Imaging for Thoracolumbar Fracture Detection and Classification: A Narrative Review
Hassan Soufan, Joya M. Mrad, Rony J. Imad El Hachem, Jean-Paul Rizk, Wajih Nasr, Imad NahlePrécis:
Artificial intelligence applications in CT-based thoracolumbar fracture analysis show promise in detection and classification, including AO Spine and TLICS frameworks. Despite promising results, further validation, multimodal integration, and clinical workflow studies are required before routine implementation.
Study Design:
A narrative review of published literature was conducted.
Objective:
Thoracolumbar fractures represent a major source of spinal morbidity. Accurate detection and classification on computed tomography (CT) directly informs operative
Summary of Background Data:
Thoracolumbar fractures require accurate CT-based detection and classification to guide management, yet increasing imaging volume and classification complexity challenge clinical workflows.
Methods:
Studies reporting AI model performance for thoracolumbar CT fracture detection, classification by the AO Spine system or the Thoracolumbar Injury Classification and Severity Score (TLICS), fracture etiology differentiation, and chronicity assessment were included. Architecture reviewed included convolutional neural networks (CNNs), U-Net segmentation pipelines, Vision Transformers (ViT), and radiomics-based machine learning frameworks.
Results:
Detection models achieved vertebra-level accuracy of up to 97.93% and area under the receiver operating characteristic curve (AUC) values of 0.88 to 0.96 across validation cohorts. For fracture classification, the AO Spine classification reached a balanced accuracy of 79.56% to 93.68% depending on the architecture, while TLICS component automation achieved AUC values of 0.85 to 0.93. Models assessing fracture chronicity attained an AUC of 0.96 on internal and an AUC of 0.89 on external validation. Generative adversarial network (GAN)-based augmentation improved classification accuracy by up to 16%.
Conclusion:
AI-based models show promising results for selected thoracolumbar fracture detection and classification tasks on CT. Nonetheless, current evidence is limited by retrospective study designs, heterogeneous data sets, and variable reference standards. Prospective multicenter validation, multimodal integration, and regulatory clarity remain prerequisites for routine clinical adoption.
Levels of Evidence:
Level V.