Fine-Tuning Pretrained CNN Models for Knee Osteoarthritis Diagnosis: Impact of Layer Freezing and Activation Function Selection

Knee osteoarthritis, often known as KOA, is a degenerative joint condition that lasts for a long time and affects millions of people around the world and has a big effect on their mental and physical health. A correct and early diagnosis is very important for better patient care and slowing the progression of illness. People can be biased and make mistakes when using the Kellgren-Lawrence (KL) classification system to assess the severity of KOA using X-ray images. It looks like deep learning, especially transfer learning, could be a good way to automate KOA diagnosis and KL grades, which would lead to more accurate results. This study looks at four Convolutional Neural Network (CNN) models that have already been trained: NASNetLarge, ResNet101V2, InceptionResNetV2, and InceptionV3. The models are used to classify KOAs. The base layers of each model are frozen, and then personalized layers are added that use the Swish, ReLu, and ELU activation functions in different ways. A last softmax activation layer is used for KL grading with more than one class (Grade 0-4). The results show that selecting the proper activation function greatly impacts model performance and training time. This shows how important it is to make activation functions and model building better before deep learning can be used in medical imaging. Using more datasets and complex hybrid models, future research will improve diagnosis accuracy and clinical usage. This helps create automatic KOA diagnosis tools.