A deep-learning model for identifying fresh vertebral compression fractures on digital radiography

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Vertebral compression fractures (VCFs) are common, and differentiating fresh from old VCFs is crucial for treatment decisions.
• Digital radiography (DR) is widely available, but its ability to distinguish fresh VCFs is limited.
• Magnetic resonance imaging (MRI) is the gold standard for VCF assessment but is less accessible than DR.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for identifying fresh VCFs using DR images.
• To compare the DL model's performance against MRI as the reference standard.
• To analyze the model's performance across different radiographic views and VCF characteristics.

Main Methods:

• A retrospective cohort of patients with lumbar VCFs who underwent both DR and MRI was analyzed.
• VCFs were classified as fresh or old based on MRI findings.
• A DL-based prediction model was trained on DR data and its diagnostic performance was assessed using metrics like AUC, accuracy, sensitivity, and specificity.
• The DeLong test was used to compare ROC curves of different models.

Main Results:

• The ensemble DL model achieved an AUC of 0.80, 74% accuracy, 80% sensitivity, and 68% specificity in identifying fresh VCFs from DR.
• Lateral DR views (AUC, 0.83) showed superior performance compared to anteroposterior views (AUC, 0.77).
• The model performed best for grade 3 VCFs (AUC, 0.89) and crush-type VCFs (AUC, 0.87).

Conclusions:

• The developed DL model demonstrates adequate performance for identifying fresh VCFs from DR images.
• The model shows potential as a non-invasive tool to aid radiologists in VCF assessment.
• Further validation and clinical integration of this DL model are warranted.