Enhanced Knee Osteoarthritis Diagnosis Using Transfer Learning: Optimizing CNN Architectures and Activation Functions for KL Grading

Millions of people worldwide suffer from knee osteoarthritis (KOA), a chronic degenerative joint disease, significantly impacting physical and psychological health. Timely and precise diagnosis is essential for enhancing patient care and slowing disease progression. The Kellgren-Lawrence (KL) grading system, frequently employed for the evaluation of KOA severity by X-ray imaging, is prone to human bias as well as inconsistency. Deep learning, particularly transfer learning, offers a promising solution for automating KOA diagnosis and KL grading, ensuring more reliable results. This study evaluates the four pretrained Convolutional Neural Network (CNN) models, namely NASNetLarge, ResNet101V2, InceptionReNetV2, and InceptionV3 for KOA classification. The base layers of each model are frozen, and customized layers incorporating Swish, ReLu, and ELU activation functions in various combinations are added. A final sigmoid activation layer is used for multi-class KL grading (Grade 0-4). The results highlight that activation function selection significantly influences model accuracy and training efficiency. These findings emphasize the importance of optimizing activation functions and model architecture for deep learning applications in medical imaging. Future research will explore advanced hybrid models and larger datasets to enhance diagnostic precision and facilitate clinical implementation, contributing to the development of automated KOA diagnostic tools.