A Knee OA Diagnosis Method via the Attention Mechanism and MSGD-Based Automatic Learning Rate Decay Strategy

Knee osteoarthritis (OA) stands as a highly prevalent degenerative joint disorder, imposing significant physical and psychological burdens on affected patients. Typically, clinicians rely on X-ray images for OA diagnosis, which is susceptible to subjective bias and fatigue-related errors. Thus, computer-aided diagnostic systems for knee OA utilizing X-rays have been developed. Current approaches primarily adopt a two-stage framework involving knee joint localization and OA severity prediction, but face several challenges: (1) Existing methods fail to adequately capture rich contextual semantic information from knee joint regions, compromising diagnostic performance. (2) Conventional learning rate decay strategies in neural network training often rely on manual configuration, potentially leading to excessively small learning rates that slow convergence, while adaptive methods like Adam may exhibit suboptimal generalization. To overcome above limitations, a novel knee OA diagnosis method incorporating the attention mechanism and an automated learning rate decay strategy via MSGD is presented. Our framework first preprocesses input data and localizes knee joint regions using KneeDetNet. Detected regions are then fed into the ResNet34 backbone network, where we integrate a coordinate attention module to enhance positional and contextual semantic information extraction. Additionally, we extract five local feature maps, which are concatenated with global features before the fully connected layer for comprehensive feature representation. The model employs a weighted joint loss function combining Center loss and Softmax loss, optimized through our MSGD-based automated learning rate decay strategy. Extensive evaluations on the public OAI dataset demonstrate that our approach significantly improves knee OA classification accuracy, showing promising potential for clinical decision support.