An Attention-Enhanced Residual Network for Osteoarthritis Grading from Medical Images

Knee osteoarthritis (KOA) is a common chronic degenerative joint disease, and its severity is typically assessed using the Kellgren-Lawrence (KL) grading system on radiographs. However, manual grading is often subjective, time-consuming, and particularly challenging when adjacent grades differ only in subtle imaging features. To address these limitations, this study proposes an automatic KOA grading method based on an attention-enhanced residual network, termed ECA-ResNet34. The model adopts ResNet34 as the backbone and introduces Efficient Channel Attention (ECA) modules into the high-level semantic stages (Stage 3 and Stage 4) to enhance lesion-related feature representation while maintaining low computational overhead. This design enables the network to better capture fine-grained pathological cues, such as joint space narrowing, osteophyte formation, and subchondral sclerosis. Experiments were conducted on a KL classification dataset constructed from the Osteoarthritis Initiative (OAI). Comparative experiments with mainstream convolutional neural networks, attention mechanism comparisons, and ablation studies were performed to evaluate the proposed model. The results show that ECA-ResNet34 achieved an accuracy of 94.60%, precision of 94.31%, recall of 94.18%, and F1-score of 94.24% on the test set, outperforming the baseline ResNet34 by 0.76 percentage points in accuracy. In addition, ECA performed better overall than SE and CBAM under the same backbone, and the Stage 3 + Stage 4 insertion strategy yielded the best results. These findings demonstrate that lightweight channel attention can effectively improve fine-grained KOA grading and provide a practical solution for intelligent computer-aided diagnosis of knee osteoarthritis.