Comparative analysis of deep learning models for early detection of osteoarthritis using knee radiographs: A retrospective study

Background

Knee osteoarthritis is a prevalent degenerative joint disease leading to pain and disability worldwide. Early detection is critical to initiating treatment strategies that can delay disease progression. While deep learning models have shown promise in automating Osteoarthritis detection from radiographs, comparative studies assessing their efficacy for early-stage detection remain limited. The aim of this study was to evaluate and compare the performance of three deep learning architectures for early detection of knee osteoarthritis using radiographic imaging.

Materials and methods

A retrospective study analysing 1200 knee radiographs (1000 training, 200 validation) collected from 2022 to 2024. Three deep learning models (custom CNN, ResNet-50, and VGG-16) were implemented and trained using PyTorch. Performance was evaluated using accuracy, sensitivity, specificity, and AUC-ROC metrics. Ground truth was established through independent assessment by three experienced orthopaedic surgeons using the Kellgren-Lawrence grading system.

Results

ResNet-50 demonstrated superior performance with accuracy 0.912 ± 0.018, sensitivity 0.908 ± 0.021, specificity 0.916 ± 0.017, and AUC 0.934 ± 0.013. VGG-16 followed with accuracy 0.887 ± 0.020, while the custom CNN achieved 0.853 ± 0.025. Statistical analysis confirmed significant differences between models (p < 0.01). Inter-observer agreement (kappa = 0.83 ± 0.02) indicated strong concordance between AI predictions and expert assessments. Model performance remained consistent across demographic subgroups, with only minimal variations based on age and BMI.

Conclusion

ResNet-50 architecture demonstrated optimal performance for early osteoarthritis detection, combining high accuracy with clinically viable processing speeds. The model's consistency across demographic subgroups and strong inter-observer agreement suggests potential for reliable clinical implementation in automated screening workflows.