Deep learning–based BMD estimation from knee radiographs with conformal uncertainty quantification
Wai Lok Yeung, Long HuiBackground
Population-wide osteoporosis screening is hindered by limited access to Dual-energy X-ray Absorptiometry (DXA). Opportunistic screening using plain radiographs is a promising alternative, but safe clinical adoption requires models that can reliably quantify their uncertainty. We evaluated a deep learning model that estimates femoral-neck bone mineral density (BMD) from knee radiographs and employs conformal prediction to provide calibrated, patient-specific uncertainty intervals.
Methods
We utilized bilateral knee radiographs with paired femoral-neck DXA scans (within 180 days) from the Osteoarthritis Initiative. The dataset was split at the participant level into training (70%), validation (10%), test (10%), and calibration (10%) sets. An EfficientNet-V2-M model was trained to predict BMD. We implemented Split Conformal Prediction with two test-time augmentation (TTA) strategies: (1) averaging augmented predictions before conformalization (Standard TTA), and (2) treating each augmented view as a separate sample (Multisample TTA). Performance was evaluated using Pearson correlation (
Results
The model achieved a Pearson correlation of
Conclusion
Deep learning can extract BMD signals from knee radiographs, and conformal prediction provides a rigorous “safety wrapper” by quantifying uncertainty. This approach offers a foundational step towards trustworthy opportunistic osteoporosis screening in orthopedic settings.