Аңдатпа
Osteoporosis is a systemic skeletal disorder affecting millions worldwide, characterized by decreased bone mineral density and increased fracture risk. Early and accurate detection from knee radiographs remains clinically challenging due to the subtlety of trabecular changes and limited availability of annotated datasets. This study presents a comparative analysis of three deep learning architectures — a custom Convolutional Neural Network (CNN), ResNet50 with partial fine-tuning, and EfficientNet-B3 with full fine-tuning — for binary classification of knee X-ray images into Normal and Osteoporosis categories. Methods. The Osteoporosis Knee X-ray Dataset (372 images) was used as the benchmark. All models were trained with Contrast-Limited Adaptive Histogram Equalization (CLAHE)-based contrast enhancement, geometric and occlusion augmentations, and a class-weighted cross-entropy loss function to address class imbalance. Early stopping and L2 regularization were applied to improve generalization. Results. Experimental results demonstrate that EfficientNet-B3 achieved the highest overall performance with an accuracy of 80.00%, recall of 89.74%, and F1-score of 82.35%, making it the most suitable model for clinical screening where sensitivity is prioritized. Custom CNN attained the highest Receiver Operating Characteristic – Area Under the Curve of 87.75%, indicating strong discriminative capacity despite its compact architecture. Conclusions. ResNet50 with partial fine-tuning showed competitive precision (76.32%) and Area Under the Curve (86.89%). These findings provide insights into the trade-offs between model complexity, transfer learning strategy, and diagnostic performance for small-scale medical imaging tasks.