Enhancing Diagnostic Accuracy of Fresh Vertebral Compression Fractures With Deep Learning Models.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Spine Surgery Diagnostics

Background:

• Diagnosing fresh thoracolumbar vertebral compression fractures often relies on MRI, which is resource-intensive and costly.
• Limited MRI availability can delay diagnosis and treatment for patients with vertebral compression fractures.
• Deep learning models utilizing X-ray images offer a potential alternative for efficient fracture diagnosis.

Purpose of the Study:

• To develop and validate deep learning models for accurate diagnosis of fresh thoracolumbar vertebral compression fractures using X-ray images.
• To assess the diagnostic performance of these models compared to traditional methods and expert evaluation.

Main Methods:

• A retrospective study utilizing 3025 lateral X-ray images from a shared database.
• Development of deep learning models using EfficientNet, MobileNet, and MnasNet architectures.
• Evaluation of model performance using metrics like accuracy, AUC, sensitivity, and specificity, with comparison to spine surgeon assessments.

Main Results:

• Deep learning models achieved accuracies ranging from 82.2% to 83.0% in the validation set and 78.1% to 80.7% in the control set.
• Area Under the Curve (AUC) values for the models were between 0.852 and 0.865.
• The developed deep learning models demonstrated superior diagnostic performance compared to both senior and junior spine surgeons in the control set.

Conclusions:

• Deep learning models effectively diagnose fresh vertebral compression fractures from X-ray images.
• These AI-powered tools show promise as a cost-effective and accessible diagnostic aid, potentially reducing reliance on MRI.
• The study highlights the potential of artificial intelligence to enhance diagnostic accuracy and efficiency in spinal imaging.