Related Experiment Video
Updated: Jul 1, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Automated Kellgren-Lawrence grading of knee osteoarthritis using a multi-scale attention-based deep learning
Henghui Zhang1, Chui Kong2, Hanwen Chang1
1Shanghai Key Laboratory of Orthopaedic Implants, Department of Orthopaedic Surgery, Shanghai Jiao Tong University School of Medicine, School of Biomedical Engineering, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University, Shanghai, China.
Background:
Accurate radiographic grading of knee osteoarthritis (KOA) using the Kellgren-Lawrence (KL) system is essential for clinical decision-making but suffers from subjective interpretation and inter-observer variability. Deep learning approaches have shown promise, yet existing models often rely on single-scale features, lack effective attention mechanisms, and generalize poorly to external datasets.
Methods:
We developed a novel multi-scale attention-guided deep learning framework that synergistically integrates three key components: (1) a Feature Pyramid Network (FPN) for multi-scale feature fusion across different spatial resolutions; (2) a dual-attention mechanism combining a Squeeze-and-Excitation (SE) module (channel-wise recalibration) and a Self-Attention module (long-range spatial dependency modeling); and (3) knowledge distillation to enhance generalization. A total of 5,000 radiographs from the OSAIL KL dataset were used for model development (3,500 training, 1,500 internal validation), with external validation on an independent dataset of 2,000 radiographs. Performance was compared against ResNet50, ResNet18, MobileNet V2, and DenseNet using 5-fold cross-validation.
Results:
On internal validation, our model achieved superior performance across all metrics (F1: 0.726, precision: 0.740, MCC: 0.620, accuracy: 0.726), outperforming all baseline architectures. On external validation, the model maintained robust generalization (F1: 0.656, precision: 0.683, MCC: 0.564, accuracy: 0.685), with only a 4.04% drop in accuracy-demonstrating clinical reliability. Misclassifications were predominantly confined to adjacent KL grades, with no extreme errors (e.g., Grade 0 vs. Grade 4). Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps revealed that the model's attention aligns with clinically relevant regions (joint space, osteophyte sites), enhancing interpretability.
Conclusion:
The proposed multi-scale attention-guided framework enables reliable, automated KL grading of KOA from radiographs with strong generalization across independent datasets. By integrating FPN, dual-attention mechanisms, and knowledge distillation, this approach addresses key limitations of prior models and may support standardized, objective, and interpretable clinical assessment of KOA severity.