Deep learning on lumbar radiographs versus DEXA for osteoporosis screening

Deep learning analysis of lumbar radiographs distinguished normal from osteopenia with AUC 0.90 and normal from osteoporosis with 90.08% accuracy, compared to DEXA as reference, in 1039 patients.

• For multiclass classification (normal, osteopenia, osteoporosis), DenseNet169 achieved a test AUC of 0.81.
• For binary tasks, Xception differentiated normal from osteopenia (AUC 0.90) and InceptionResNetV2 separated normal from osteoporosis (accuracy 90.08%).
• Retrospective study of 1039 patients who had both DEXA and lumbar radiographs within three months.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.