Hierarchical classification of acute and chronic osteoporotic vertebral compression fractures of the lumbar spine using X-ray images.

Abstract

Objective: To develop and validate a hierarchical deep learning model for differentiating acute and chronic lumbar osteoporotic vertebral compression fractures (OVCFs) using X-ray images.

Materials and methods: We retrospectively reviewed approximately 2600 lateral lumbar radiographs obtained from patients clinically suspected of having OVCFs between 2007 and 2022. After excluding poor-quality images and surgically instrumented vertebrae, 1299 radiographs (6495 vertebral patches, L1-L5) were included. Labeling was performed by neurosurgeons and radiologists using X-ray images, with CT and/or MRI findings serving as the reference standard. A two-step hierarchical classification was implemented: first classifying vertebrae into Normal-Chronic, Acute, and Indeterminate (cement-augmented vertebrae without instrumentation) groups, followed by subdivision of the Normal-Chronic group into Normal and Chronic categories.

Results: A total of 1299 radiographs were evaluated. The hierarchical model achieved an accuracy of 91% in the initial three-class step. For the detection of acute fractures in the final classification step, the model demonstrated a sensitivity of 91.0% (95% CI 84.8-95.0%), a specificity of 82.1% (95% CI 79.8-84.5%), and a high negative predictive value (NPV) of 98.8% (95% CI 97.9-99.3%). The Normal-aligned hierarchical approach outperformed the Acute-aligned and end-to-end models, particularly for acute and chronic cases.

Conclusion: The proposed hierarchical approach enhances the diagnostic utility of standard X-ray images by enabling more accurate classification of lumbar fracture types. This study is limited by its single-institution retrospective design. This model may reduce reliance on advanced imaging and support faster and more informed clinical decision-making.