A multi-stage framework used pretrained CNNs to extract knee-radiograph features and machine-learning classifiers to distinguish normal, osteopenia and osteoporosis. External testing exposed a gap between AUC and threshold performance, especially probability collapse for osteopenia, requiring recalibration and threshold adaptation.
Key findings
- EfficientNetB0 and DenseNet121 generally outperformed ResNet18; DenseNet121-neural network had the highest overall classification, while EfficientNetB0-efficient linear was stable. External data revealed class-prior mismatch and osteopenia probability collapse; recalibration and prior boosting improved sensitivity, balanced accuracy and macro F1.
Why this matters globally
Routine knee radiographs could flag otherwise unscreened people for definitive osteoporosis assessment. The appropriate role is opportunistic screening and referral, not replacement of clinical diagnosis.
Thai researcher contribution
University of Phayao medicine, health-data and science researchers collaborated with Chiang Mai University's optimization and computational-intelligence center on both modeling and domain-shift analysis.
Limitations to consider
This remains a proof of concept despite public external validation. The abstract does not detail reference standards, demographic and scanner diversity or clinical decision impact, and thresholds adapted to one external domain may not transfer to the next.
Verify the original sources
Journal of Clinical MedicineRead the original article↗DOI: 10.3390/jcm15135222