Entropy-Guided Adaptive Confidence-Weighted Complementary Deep Ensemble for osteopenia and osteoporosis classification from knee radiographs

Abstract Automated osteoporosis assessment from knee radiographs remains a challenging computer-aided diagnosis task due to subtle structural differences among normal, osteopenic, and osteoporotic bone patterns, together with the limited availability of transparent and reproducible evaluation frameworks based on publicly accessible datasets. Existing studies predominantly rely on standalone deep learning models or conventional ensemble methods that assume fixed classifier reliability and rarely incorporate predictive uncertainty into decision-level fusion. To address this limitation, this study proposes an Entropy-Guided Adaptive Confidence-Weighted Complementary Deep Ensemble (E-ACW Ensemble) for uncertainty-aware osteoporosis-related knee radiograph classification in both binary and multi-class settings. The proposed framework integrates standardized radiographic preprocessing, fold-specific data augmentation, complementary deep feature learning, and an entropy-guided adaptive fusion strategy. Predictive uncertainty is estimated using Shannon entropy to enable sample-specific confidence weighting during decision-level fusion. The framework was evaluated on publicly available knee radiograph datasets using a rigorous five-fold stratified cross-validation protocol. Comprehensive experimental evaluation, including comparative benchmarking, ablation studies, fold-wise stability analysis, statistical significance testing, and computational efficiency assessment, was conducted to validate predictive performance, robustness, and reproducibility. Among the investigated configurations, the proposed EfficientNetB0 + MobileNetV2 + E-ACW Ensemble achieved the best overall performance, reaching 99.2% accuracy and 99.1% F1-score for binary classification and 98.7% accuracy and 98.5% F1-score for multi-class classification while requiring substantially lower computational cost than the corresponding high-capacity ensemble. These findings demonstrate that entropy-guided adaptive confidence weighting provides an effective and computationally efficient decision-level fusion strategy for osteoporosis-related knee radiograph classification. The proposed E-ACW Ensemble establishes a reproducible framework that can be readily extended to future computer-aided musculoskeletal image analysis and uncertainty-aware medical imaging applications.