Comparative Analysis of Deep Learning Architectures for Osteoporosis Detection in Knee Radiographs: From Custom Cnn to Transfer Learning
Year: 2026
Type: article
Abstract: Osteoporosis is a systemic skeletal disease associated with reduced bone mineral density and increased fracture risk. Although dual-energy X-ray absorptiometry remains the diagnostic gold standard, conventional knee radiographs are more widely available and may support preliminary screening in resource-limited clinical settings. The aim of this study was to compare deep learning-based approaches for classifying knee radiographs into Normal and Osteoporosis categories. A publicly available dataset of 372 knee X-ray images was used; image preprocessing included Contrast Limited Adaptive Histogram Equalization (CLAHE), and the models were trained and evaluated under the same experimental conditions. The best screening-oriented performance was achieved by EfficientNet-B3, which reached 80.00% accuracy, 89.74% recall, and an F1-score of 82.35%. The custom convolutional neural network (CNN) demonstrated the highest receiver operating characteristic area under the curve (ROC-AUC), reaching 87.75%. The results suggest that both transfer learning and compact task-specific models may be useful for preliminary osteoporosis detection from knee radiographs. Additional analyses included five-fold cross-validation, confusion matrix-based error analysis, ablation experiments, and Grad-CAM visualization. However, the study is exploratory and limited by the small dataset size and the absence of external validation. Further evaluation on larger independent clinical cohorts is required before clinical application.
Source: Вестник КазАТК
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225088427
FWCI:
Cites:
Cited by:
Related to: