Assessing deep learning model performance in osteoporosis screening with lumbar spine radiographs

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Osteoporosis diagnosis and fracture risk assessment commonly rely on dual energy X-ray absorptiometry (DXA).
• Limited DXA resources necessitate alternative screening methods.
• Lumbar spine radiographs offer a potentially accessible imaging source.

Purpose of the Study:

• To develop and evaluate deep learning models for osteoporosis screening.
• To assess the accuracy of these models using lumbar spine radiographs.
• To explore AI as a complementary tool for osteoporosis detection.

Main Methods:

• Deep learning models (ResNet-18, DarkNet-19) were trained on anteroposterior (AP) and lateral lumbar spine radiographs.
• Patients were classified as non-osteoporosis (T-score > -2.5) or osteoporosis (T-score ≤ -2.5) based on DXA.
• Model performance was evaluated using area under the curve (AUC), sensitivity, and specificity on a separate test dataset.

Main Results:

• For AP images, ResNet-18 achieved an AUC of 0.79 (sensitivity 79.7%, specificity 66.5%).
• For lateral images, DarkNet-19 achieved the highest AUC of 0.82 (sensitivity 87.5%, specificity 79.4%).
• The models demonstrated varying but significant accuracy in detecting osteoporosis from radiographs.

Conclusions:

• Deep learning models show potential for effective osteoporosis screening from lumbar spine radiographs.
• These AI-driven tools could serve as accessible alternatives or complements to DXA.
• This approach may facilitate broader osteoporosis risk assessment and treatment decisions.