Opportunistic Osteoporosis Screening from Routine Knee Radiographs Using a Multi-Stage CNN Framework with External...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Public Health

Background:

• Osteoporosis poses a significant public health risk, leading to fractures and reduced quality of life.
• Early detection through opportunistic screening of knee X-ray images using AI is a promising approach.

Purpose of the Study:

• To develop and evaluate a multi-stage deep learning and machine learning framework for osteoporosis classification.
• To emphasize external validation, calibration drift, and cross-domain generalization performance.

Main Methods:

• Extracted deep features from knee X-rays using pretrained CNNs (ResNet18, EfficientNetB0, DenseNet121).
• Classified features using ML models (Neural Network, Efficient Linear, SVM, Naive Bayes).
• Investigated data augmentation and evaluated performance using accuracy, F1-score, AUC, and reliability calibration.

Main Results:

• EfficientNetB0 and DenseNet121 outperformed ResNet18.
• External validation revealed calibration drift and class-prior mismatch.
• Post-hoc recalibration and class-prior boosting improved performance on external datasets, especially for the Osteopenia class.

Conclusions:

• The AI framework shows feasibility for osteoporosis classification from knee X-rays.
• Adaptive recalibration and threshold optimization are crucial for maintaining performance across different domains.
• Further validation on diverse clinical cohorts is needed for generalizability and clinical utility.